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
1148 changed files with 184135 additions and 56064 deletions
@@ -11,6 +11,7 @@ ENV/
 .uv/
 *.egg-info/
 dist/
+!aether-hub/dist/aether-hub
 build/
 *.egg

@@ -21,11 +21,9 @@ JWT_SECRET_KEY=change-this-to-a-secure-random-string
 # 注意：更换此密钥后需要在管理面板重新配置所有 Provider API Key
 ENCRYPTION_KEY=change-this-to-another-secure-random-string

-# 代理节点 HMAC 密钥（用于 aether-proxy 认证）
-# 可选：不设置时会从 ENCRYPTION_KEY 自动派生
-# 显式设置时，aether-proxy.toml 的 hmac_key 配置相同值即可
-# 可通过 python generate_keys.py 生成
-# PROXY_HMAC_KEY=change-this-to-a-proxy-hmac-key
+# 支付回调共享密钥（公开 /api/payment/callback/* 入口必须携带 x-payment-callback-token）
+# 建议使用 32+ 位随机字符串
+PAYMENT_CALLBACK_SECRET=change-this-to-a-secure-callback-secret

 # 管理员账号（仅首次初始化时使用, 创建完成后可在系统内修改密码）
 ADMIN_EMAIL=[email protected]
@@ -38,6 +36,53 @@ ADMIN_PASSWORD=admin123456
 # 应用端口（默认 8084）
 # APP_PORT=8084

+# 生产部署镜像（deploy.sh 会读取）
+# APP_IMAGE=ghcr.io/fawney19/aether:latest
+
+# Gunicorn Worker 数量（默认 2）
+# Tunnel 请求统一经 Hub 转发，可安全使用多 worker。
+# 非 Docker 运行时若使用 ProxyNode tunnel，请确保 aether-hub 可达（默认 ws://127.0.0.1:8085）。
+# GUNICORN_WORKERS=2
+
+# Gunicorn Max Requests（默认 4000）
+# Worker 处理指定数量请求后自动重启，防止内存泄漏
+# 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

@@ -49,6 +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
@@ -0,0 +1,91 @@
+name: Build aether-hub
+
+on:
+  push:
+    tags: ['hub-v*']
+  workflow_dispatch:
+
+permissions:
+  contents: write
+
+jobs:
+  build:
+    name: ${{ matrix.name }}
+    runs-on: ubuntu-latest
+    strategy:
+      fail-fast: false
+      matrix:
+        include:
+          - name: linux-amd64
+            target: x86_64-unknown-linux-gnu
+            use_cross: true
+          - name: linux-arm64
+            target: aarch64-unknown-linux-gnu
+            use_cross: true
+
+    steps:
+      - uses: actions/checkout@v5
+
+      - name: Install Rust toolchain
+        uses: dtolnay/rust-toolchain@stable
+        with:
+          targets: ${{ matrix.target }}
+
+      - name: Rust cache
+        uses: Swatinem/rust-cache@v2
+        with:
+          workspaces: aether-hub -> target
+          key: ${{ matrix.target }}
+
+      - name: Install cross
+        if: matrix.use_cross
+        uses: taiki-e/install-action@cross
+
+      - name: Build
+        working-directory: aether-hub
+        shell: bash
+        run: |
+          if [ "${{ matrix.use_cross }}" = "true" ]; then
+            cross build --release --target ${{ matrix.target }}
+          else
+            cargo build --release --target ${{ matrix.target }}
+          fi
+
+      - name: Package
+        shell: bash
+        run: |
+          cd aether-hub/target/${{ matrix.target }}/release
+          chmod +x aether-hub
+          tar czf ../../../../aether-hub-${{ matrix.name }}.tar.gz aether-hub
+
+      - name: Upload artifact
+        uses: actions/upload-artifact@v5
+        with:
+          name: aether-hub-${{ matrix.name }}
+          path: aether-hub-*.tar.gz
+          if-no-files-found: error
+
+  release:
+    needs: build
+    runs-on: ubuntu-latest
+    if: startsWith(github.ref, 'refs/tags/')
+    steps:
+      - name: Download all artifacts
+        uses: actions/download-artifact@v5
+        with:
+          merge-multiple: true
+          path: artifacts
+
+      - name: Generate checksums
+        working-directory: artifacts
+        run: sha256sum aether-hub-* > SHA256SUMS.txt
+
+      - name: Create GitHub Release
+        uses: softprops/action-gh-release@v2
+        with:
+          name: "${{ github.ref_name }}"
+          generate_release_notes: true
+          files: |
+            artifacts/aether-hub-*
+            artifacts/SHA256SUMS.txt
+          fail_on_unmatched_files: true
@@ -1,4 +1,4 @@
-name: Build aether-proxy Binaries
+name: Build aether-proxy

 on:
  push:
@@ -7,6 +7,12 @@ on:

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

 jobs:
  build:
@@ -38,7 +44,7 @@ jobs:
            use_cross: false

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

      - name: Install Rust toolchain
        uses: dtolnay/rust-toolchain@stable
@@ -81,7 +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: |
@@ -95,7 +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
@@ -107,44 +113,115 @@ jobs:
      - name: Create GitHub Release
        uses: softprops/action-gh-release@v2
        with:
-          name: "aether-proxy ${{ github.ref_name }}"
-          body: |
-            ## aether-proxy ${{ github.ref_name }}
-
-            ### 下载
-
-            | 平台 | 文件 |
-            |------|------|
-            | Linux x86_64 | `aether-proxy-linux-amd64.tar.gz` |
-            | Linux ARM64 | `aether-proxy-linux-arm64.tar.gz` |
-            | macOS x86_64 (Intel) | `aether-proxy-macos-amd64.tar.gz` |
-            | macOS ARM64 (Apple Silicon) | `aether-proxy-macos-arm64.tar.gz` |
-            | Windows x86_64 | `aether-proxy-windows-amd64.zip` |
-
-            ### 使用
-
-            ```bash
-            # Linux/macOS: 解压
-            tar xzf aether-proxy-<platform>.tar.gz
-
-            # Windows: 解压 zip 文件
-
-            # 配置环境变量 (或写入 .env 文件)
-            export AETHER_PROXY_AETHER_URL=https://your-aether.example.com
-            export AETHER_PROXY_MANAGEMENT_TOKEN=ae_xxx
-            export AETHER_PROXY_HMAC_KEY=your-hmac-key
-
-            # 运行
-            ./aether-proxy
-            ```
-
-            ### 校验
-
-            下载 `SHA256SUMS.txt` 后可验证文件完整性:
-            ```bash
-            sha256sum -c SHA256SUMS.txt
-            ```
+          name: "${{ github.ref_name }}"
+          generate_release_notes: true
          files: |
            artifacts/aether-proxy-*
            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@v5
+
+      - name: Download Linux artifacts
+        uses: actions/download-artifact@v5
+        with:
+          pattern: aether-proxy-linux-*
+          merge-multiple: true
+          path: artifacts
+
+      - name: Prepare binaries
+        run: |
+          mkdir -p aether-proxy/build/linux-amd64 aether-proxy/build/linux-arm64
+          tar xzf artifacts/aether-proxy-linux-amd64.tar.gz -C aether-proxy/build/linux-amd64
+          tar xzf artifacts/aether-proxy-linux-arm64.tar.gz -C aether-proxy/build/linux-arm64
+
+      - 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=proxy-v(.*),group=1
+            type=match,pattern=proxy-v(\d+\.\d+),group=1
+            type=sha,prefix=
+          flavor: |
+            latest=auto
+
+      - name: Build and push
+        uses: docker/build-push-action@v6
+        with:
+          context: ./aether-proxy
+          push: true
+          tags: ${{ steps.meta.outputs.tags }}
+          labels: ${{ steps.meta.outputs.labels }}
+          platforms: linux/amd64,linux/arm64
+
+  update-readme:
+    needs: release
+    runs-on: ubuntu-latest
+    if: startsWith(github.ref, 'refs/tags/')
+    steps:
+      - uses: actions/checkout@v5
+        with:
+          ref: master
+
+      - name: Update README download links
+        env:
+          TAG: ${{ github.ref_name }}
+        run: |
+          VERSION="${TAG#proxy-v}"
+          BASE="https://github.com/fawney19/Aether/releases/download/${TAG}"
+          cd aether-proxy
+
+          TABLE="| Platform | Download |\n|----------|----------|\n"
+          TABLE+="| Linux x86_64 | [aether-proxy-linux-amd64.tar.gz](${BASE}/aether-proxy-linux-amd64.tar.gz) |\n"
+          TABLE+="| Linux ARM64 | [aether-proxy-linux-arm64.tar.gz](${BASE}/aether-proxy-linux-arm64.tar.gz) |\n"
+          TABLE+="| macOS x86_64 | [aether-proxy-macos-amd64.tar.gz](${BASE}/aether-proxy-macos-amd64.tar.gz) |\n"
+          TABLE+="| macOS ARM64 | [aether-proxy-macos-arm64.tar.gz](${BASE}/aether-proxy-macos-arm64.tar.gz) |\n"
+          TABLE+="| Windows x86_64 | [aether-proxy-windows-amd64.zip](${BASE}/aether-proxy-windows-amd64.zip) |"
+
+          # Replace content between markers
+          if grep -q '<!-- DOWNLOAD_TABLE_START -->' README.md; then
+            awk -v table="$TABLE" '
+              /<!-- DOWNLOAD_TABLE_START -->/ { print; printf "%s\n", table; skip=1; next }
+              /<!-- DOWNLOAD_TABLE_END -->/ { skip=0 }
+              !skip { print }
+            ' README.md > README.tmp && mv README.tmp README.md
+          fi
+
+      - name: Commit and push
+        run: |
+          cd aether-proxy
+          git config user.name "github-actions[bot]"
+          git config user.email "github-actions[bot]@users.noreply.github.com"
+          git add README.md
+          git diff --cached --quiet && exit 0
+          TAG="${GITHUB_REF#refs/tags/}"
+          git commit -m "chore(proxy): update download links for ${TAG}"
+          git push
@@ -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
@@ -15,6 +15,7 @@ env:
  REGISTRY: ghcr.io
  BASE_IMAGE_NAME: fawney19/aether-base
  APP_IMAGE_NAME: fawney19/aether
+  GITHUB_REPO: fawney19/Aether
  # Base image hash inputs:
  # - Dockerfile.base
  # - pyproject.toml (dependency fingerprint only; ignores tool/optional deps)
@@ -29,7 +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
@@ -107,7 +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
@@ -163,7 +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
@@ -174,15 +175,36 @@ jobs:
          cache-to: type=gha,mode=max,scope=base
          platforms: linux/amd64,linux/arm64

+  download-hub:
+    runs-on: ubuntu-latest
+    permissions:
+      contents: read
+    outputs:
+      hub_tag: ${{ steps.hub-tag.outputs.tag }}
+    steps:
+      - name: Get latest hub release tag
+        id: hub-tag
+        env:
+          GH_TOKEN: ${{ secrets.GITHUB_TOKEN }}
+        run: |
+          TAG=$(gh release list --repo "${{ env.GITHUB_REPO }}" --limit 50 --json tagName,isDraft,isPrerelease \
+            --jq '[.[] | select(.tagName | startswith("hub-v")) | select(.isDraft == false and .isPrerelease == false)] | .[0].tagName')
+          if [ -z "$TAG" ] || [ "$TAG" = "null" ]; then
+            echo "No hub release found"
+            exit 1
+          fi
+          echo "tag=$TAG" >> $GITHUB_OUTPUT
+          echo "Hub release tag: $TAG"
+
  build-app:
-    needs: [check-base-changes, build-base]
-    if: always() && (needs.build-base.result == 'success' || needs.build-base.result == 'skipped')
+    needs: [check-base-changes, build-base, download-hub]
+    if: always() && (needs.build-base.result == 'success' || needs.build-base.result == 'skipped') && needs.download-hub.result == 'success'
    runs-on: ubuntu-latest
    permissions:
      contents: read
      packages: write
    steps:
-      - uses: actions/checkout@v4
+      - uses: actions/checkout@v5

      - name: Set up Docker Buildx
        uses: docker/setup-buildx-action@v3
@@ -243,15 +265,63 @@ jobs:
          version_tuple = __version_tuple__
          EOF

-      - name: Build and push app image
-        uses: docker/build-push-action@v5
+      - name: Resolve hub release for build args
+        run: |
+          echo "Hub release tag: ${{ needs.download-hub.outputs.hub_tag }}"
+
+      - name: Build and push app image (amd64)
+        id: build-amd64
+        uses: docker/build-push-action@v6
        with:
          context: .
          file: ./Dockerfile.app
-          push: true
-          tags: ${{ steps.meta.outputs.tags }}
          labels: ${{ steps.meta.outputs.labels }}
          no-cache-filters: builder
-          cache-from: type=gha,scope=app
-          cache-to: type=gha,mode=min,scope=app
-          platforms: linux/amd64,linux/arm64
+          cache-from: type=gha,scope=app-amd64
+          cache-to: type=gha,mode=min,scope=app-amd64
+          build-args: |
+            HUB_RELEASE_REPO=${{ env.GITHUB_REPO }}
+            HUB_TAG=${{ needs.download-hub.outputs.hub_tag }}
+          platforms: linux/amd64
+          outputs: type=image,"name=${{ env.REGISTRY }}/${{ env.APP_IMAGE_NAME }},docker.io/fawney19/aether",push-by-digest=true,name-canonical=true,push=true
+
+      - name: Build and push app image (arm64)
+        id: build-arm64
+        uses: docker/build-push-action@v6
+        with:
+          context: .
+          file: ./Dockerfile.app
+          labels: ${{ steps.meta.outputs.labels }}
+          no-cache-filters: builder
+          cache-from: type=gha,scope=app-arm64
+          cache-to: type=gha,mode=min,scope=app-arm64
+          build-args: |
+            HUB_RELEASE_REPO=${{ env.GITHUB_REPO }}
+            HUB_TAG=${{ needs.download-hub.outputs.hub_tag }}
+          platforms: linux/arm64
+          outputs: type=image,"name=${{ env.REGISTRY }}/${{ env.APP_IMAGE_NAME }},docker.io/fawney19/aether",push-by-digest=true,name-canonical=true,push=true
+
+      - name: Create multi-arch manifest and push
+        run: |
+          # Extract digests
+          AMD64_DIGEST="${{ steps.build-amd64.outputs.digest }}"
+          ARM64_DIGEST="${{ steps.build-arm64.outputs.digest }}"
+          echo "amd64 digest: $AMD64_DIGEST"
+          echo "arm64 digest: $ARM64_DIGEST"
+
+          # For each tag, create multi-arch manifest on each registry
+          TAGS=$(echo "${{ steps.meta.outputs.tags }}" | tr '\n' ' ')
+          for FULL_TAG in $TAGS; do
+            # Determine which registry this tag belongs to
+            if [[ "$FULL_TAG" == ghcr.io/* ]]; then
+              REPO="${{ env.REGISTRY }}/${{ env.APP_IMAGE_NAME }}"
+            elif [[ "$FULL_TAG" == docker.io/* ]]; then
+              REPO="docker.io/fawney19/aether"
+            else
+              continue
+            fi
+            echo "Creating manifest for $FULL_TAG"
+            docker buildx imagetools create -t "$FULL_TAG" \
+              "$REPO@$AMD64_DIGEST" \
+              "$REPO@$ARM64_DIGEST"
+          done
@@ -1,10 +1,8 @@
 # Created by https://www.toptal.com/developers/gitignore/api/python
 # Edit at https://www.toptal.com/developers/gitignore?templates=python

-CLIProxyAPI/
-sub2api/
-
 # AI Assistant Configuration
+.codex/
 .claude/
 .serena/
 .gemini*/
@@ -206,6 +204,7 @@ logs/

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

 # Database backups
 backups/
@@ -229,10 +228,15 @@ test.py
 .deps-hash
 .code-hash
 .migration-hash
+.hub-hash
+
+# Hub prebuilt binaries
+aether-hub/dist/

 # Version file (auto-generated by hatch-vcs)
 src/_version.py

 # Analysis folder (third-party code for reference)
 analysis/
+new-api/
 /aether-proxy/target/
@@ -1 +1 @@
-3.14
+3.13
@@ -2,14 +2,22 @@
 # 运行镜像：从 base 提取产物到精简运行时
 # 构建命令: docker build -f Dockerfile.app -t aether-app:latest .
 # 用于 GitHub Actions CI（官方源）
+
 FROM aether-base:latest AS builder
 WORKDIR /app
 # 复制前端源码并构建（CI 通过 no-cache-filters=builder 确保每次重建）
 COPY frontend/ ./frontend/
 RUN cd frontend && npm run build
+
 # ==================== 运行时镜像 ====================
-FROM python:3.14-slim
+FROM python:3.13-slim
 WORKDIR /app
+
+ARG HUB_RELEASE_REPO=fawney19/Aether
+ARG HUB_TAG
+ARG TARGETARCH
+ARG GITHUB_TOKEN
+
 # 运行时依赖（无 gcc/nodejs/npm，使用 BuildKit 缓存加速）
 RUN --mount=type=cache,target=/var/cache/apt,sharing=locked \
    --mount=type=cache,target=/var/lib/apt,sharing=locked \
@@ -17,13 +25,49 @@ 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 "$jemalloc_path" /usr/local/lib/libjemalloc.so.2
 # 从 base 镜像复制 Python 包
-COPY --from=builder /usr/local/lib/python3.14/site-packages /usr/local/lib/python3.14/site-packages
+COPY --from=builder /usr/local/lib/python3.13/site-packages /usr/local/lib/python3.13/site-packages
 # 只复制需要的 Python 可执行文件
 COPY --from=builder /usr/local/bin/gunicorn /usr/local/bin/
 COPY --from=builder /usr/local/bin/uvicorn /usr/local/bin/
 COPY --from=builder /usr/local/bin/alembic /usr/local/bin/
+# Hub 预编译二进制（构建时从 GitHub Release 下载）
+# GITHUB_TOKEN 可选：未认证 API 限流 60 次/小时，认证后 5000 次/小时
+RUN set -eux; \
+    auth_header=""; \
+    if [ -n "${GITHUB_TOKEN:-}" ]; then \
+        auth_header="Authorization: token ${GITHUB_TOKEN}"; \
+    fi; \
+    tag="${HUB_TAG:-}"; \
+    if [ -z "$tag" ]; then \
+        tag="$(curl -sL ${auth_header:+-H "$auth_header"} "https://api.github.com/repos/${HUB_RELEASE_REPO}/releases" | python3 -c "import json,sys;print(next((r['tag_name'] for r in json.load(sys.stdin) if r.get('tag_name','').startswith('hub-v') and not r.get('draft') and not r.get('prerelease')),''))")"; \
+    fi; \
+    if [ -z "$tag" ]; then \
+        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"; \
+    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; \
+    rm -f /tmp/aether-hub.tar.gz
 # 从 builder 阶段复制前端构建产物
 COPY --from=builder /app/frontend/dist /usr/share/nginx/html
 RUN chmod -R 755 /usr/share/nginx/html
@@ -36,108 +80,194 @@ COPY gunicorn_conf.py ./
 # 策略：白名单后端路由 → 后端代理，其余全部 → 前端 SPA（index.html）
 # 智能处理 IP：有外层代理头就透传，没有就用直连 IP
 RUN printf '%s\n' \
-'map $http_x_real_ip $real_ip {' \
-'    default $http_x_real_ip;' \
-'    ""      $remote_addr;' \
-'}' \
-'' \
-'map $http_x_forwarded_for $forwarded_for {' \
-'    default $http_x_forwarded_for;' \
-'    ""      $remote_addr;' \
-'}' \
-'' \
-'server {' \
-'    listen 80;' \
-'    server_name _;' \
-'    root /usr/share/nginx/html;' \
-'    index index.html;' \
-'    client_max_body_size 100M;' \
-'' \
-'    # gzip 压缩配置（对 base64 图片等非流式响应有效）' \
-'    gzip on;' \
-'    gzip_min_length 256;' \
-'    gzip_comp_level 5;' \
-'    gzip_vary on;' \
-'    gzip_proxied any;' \
-'    gzip_types application/json text/plain text/css text/javascript application/javascript application/octet-stream;' \
-'    gzip_disable "msie6";' \
-'' \
-'    # 静态资源：长期缓存' \
-'    location ~* \.(js|css|png|jpg|jpeg|gif|ico|svg|woff|woff2|ttf|eot)$ {' \
-'        expires 1y;' \
-'        add_header Cache-Control "public, no-transform";' \
-'        try_files $uri =404;' \
-'    }' \
-'' \
-'    # 安全：阻止访问源码目录' \
-'    location ~ ^/(src|node_modules)/ {' \
-'        deny all;' \
-'        return 404;' \
-'    }' \
-'' \
-'    # 后端 API 路由（白名单）→ 代理到后端' \
-'    location ~ ^/(api|v1|v1beta|upload|health)(/|$) {' \
-'        proxy_pass http://127.0.0.1:PORT_PLACEHOLDER;' \
-'        proxy_http_version 1.1;' \
-'        proxy_set_header Host $host;' \
-'        proxy_set_header X-Real-IP $real_ip;' \
-'        proxy_set_header X-Forwarded-For $forwarded_for;' \
-'        proxy_set_header X-Forwarded-Proto $scheme;' \
-'        proxy_set_header Connection "";' \
-'        proxy_set_header Accept $http_accept;' \
-'        proxy_set_header Content-Type $content_type;' \
-'        proxy_set_header Authorization $http_authorization;' \
-'        proxy_set_header X-Api-Key $http_x_api_key;' \
-'        proxy_buffering off;' \
-'        proxy_cache off;' \
-'        proxy_request_buffering off;' \
-'        chunked_transfer_encoding on;' \
-'        gzip off;' \
-'        add_header X-Accel-Buffering no;' \
-'        proxy_connect_timeout 600s;' \
-'        proxy_send_timeout 600s;' \
-'        proxy_read_timeout 600s;' \
-'    }' \
-'' \
-'    # API 文档路由 → 代理到后端' \
-'    location ~ ^/(docs|redoc|openapi\\.json)$ {' \
-'        proxy_pass http://127.0.0.1:PORT_PLACEHOLDER;' \
-'        proxy_http_version 1.1;' \
-'        proxy_set_header Host $host;' \
-'        proxy_set_header X-Real-IP $real_ip;' \
-'        proxy_set_header X-Forwarded-For $forwarded_for;' \
-'        proxy_set_header X-Forwarded-Proto $scheme;' \
-'    }' \
-'' \
-'    # 所有其他路由 → 前端 SPA（先尝试静态文件，再回退到 index.html）' \
-'    location / {' \
-'        try_files $uri $uri/ /index.html;' \
-'    }' \
-'}' > /etc/nginx/sites-available/default.template
+    'map $http_x_real_ip $real_ip {' \
+    '    default $http_x_real_ip;' \
+    '    ""      $remote_addr;' \
+    '}' \
+    '' \
+    'map $http_x_forwarded_for $forwarded_for {' \
+    '    default $http_x_forwarded_for;' \
+    '    ""      $remote_addr;' \
+    '}' \
+    '' \
+    'map $http_upgrade $connection_upgrade {' \
+    '    default upgrade;' \
+    '    ""      "";' \
+    '}' \
+    '' \
+    'server {' \
+    '    listen 80;' \
+    '    server_name _;' \
+    '    root /usr/share/nginx/html;' \
+    '    index index.html;' \
+    '    client_max_body_size 100M;' \
+    '' \
+    '    # gzip 压缩配置（对 base64 图片等非流式响应有效）' \
+    '    gzip on;' \
+    '    gzip_min_length 256;' \
+    '    gzip_comp_level 5;' \
+    '    gzip_vary on;' \
+    '    gzip_proxied any;' \
+    '    gzip_types application/json text/plain text/css text/javascript application/javascript application/octet-stream;' \
+    '    gzip_disable "msie6";' \
+    '' \
+    '    # 静态资源：长期缓存' \
+    '    location ~* \.(js|css|png|jpg|jpeg|gif|ico|svg|woff|woff2|ttf|eot)$ {' \
+    '        expires 1y;' \
+    '        add_header Cache-Control "public, no-transform";' \
+    '        try_files $uri =404;' \
+    '    }' \
+    '' \
+    '    # 安全：阻止访问源码目录' \
+    '    location ~ ^/(src|node_modules)/ {' \
+    '        deny all;' \
+    '        return 404;' \
+    '    }' \
+    '' \
+    '    # WebSocket 隧道端点（aether-proxy tunnel 模式）' \
+    '    location = /api/internal/proxy-tunnel {' \
+    '        proxy_pass http://127.0.0.1:8085/proxy;' \
+    '        proxy_http_version 1.1;' \
+    '        proxy_set_header Host $host;' \
+    '        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 Upgrade $http_upgrade;' \
+    '        proxy_set_header Connection "upgrade";' \
+    '        # 剥离 CF 头，防止泄露给上游或返回给客户端' \
+    '        proxy_hide_header CF-Connecting-IP;' \
+    '        proxy_hide_header CF-IPCountry;' \
+    '        proxy_hide_header CF-Ray;' \
+    '        proxy_hide_header CF-Visitor;' \
+    '        proxy_hide_header CDN-Loop;' \
+    '        proxy_hide_header True-Client-IP;' \
+    '        proxy_hide_header CF-Worker;' \
+    '        proxy_hide_header CF-EW-Via;' \
+    '        proxy_hide_header CF-Warp-Tag-ID;' \
+    '        proxy_set_header CF-Connecting-IP "";' \
+    '        proxy_set_header CF-IPCountry "";' \
+    '        proxy_set_header CF-Ray "";' \
+    '        proxy_set_header CF-Visitor "";' \
+    '        proxy_set_header CDN-Loop "";' \
+    '        proxy_set_header True-Client-IP "";' \
+    '        proxy_set_header CF-Worker "";' \
+    '        proxy_set_header CF-EW-Via "";' \
+    '        proxy_set_header CF-Warp-Tag-ID "";' \
+    '        proxy_read_timeout 86400s;' \
+    '        proxy_send_timeout 86400s;' \
+    '    }' \
+    '' \
+    '    # 后端 API 路由（白名单）→ 代理到后端' \
+    '    location ~ ^/(api|v1|v1beta|upload|health)(/|$) {' \
+    '        proxy_pass http://127.0.0.1:PORT_PLACEHOLDER;' \
+    '        proxy_http_version 1.1;' \
+    '        proxy_set_header Host $host;' \
+    '        proxy_set_header X-Real-IP $real_ip;' \
+    '        proxy_set_header X-Forwarded-For $forwarded_for;' \
+    '        proxy_set_header X-Forwarded-Proto $scheme;' \
+    '        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 $http_x_api_key;' \
+    '        # 剥离 CF 头，防止泄露给上游或返回给客户端' \
+    '        proxy_hide_header CF-Connecting-IP;' \
+    '        proxy_hide_header CF-IPCountry;' \
+    '        proxy_hide_header CF-Ray;' \
+    '        proxy_hide_header CF-Visitor;' \
+    '        proxy_hide_header CDN-Loop;' \
+    '        proxy_hide_header True-Client-IP;' \
+    '        proxy_hide_header CF-Worker;' \
+    '        proxy_hide_header CF-EW-Via;' \
+    '        proxy_hide_header CF-Warp-Tag-ID;' \
+    '        proxy_set_header CF-Connecting-IP "";' \
+    '        proxy_set_header CF-IPCountry "";' \
+    '        proxy_set_header CF-Ray "";' \
+    '        proxy_set_header CF-Visitor "";' \
+    '        proxy_set_header CDN-Loop "";' \
+    '        proxy_set_header True-Client-IP "";' \
+    '        proxy_set_header CF-Worker "";' \
+    '        proxy_set_header CF-EW-Via "";' \
+    '        proxy_set_header CF-Warp-Tag-ID "";' \
+    '        proxy_buffering off;' \
+    '        proxy_cache off;' \
+    '        proxy_request_buffering off;' \
+    '        chunked_transfer_encoding on;' \
+    '        gzip off;' \
+    '        add_header X-Accel-Buffering no;' \
+    '        proxy_connect_timeout 60s;' \
+    '        proxy_send_timeout 3600s;' \
+    '        proxy_read_timeout 3600s;' \
+    '    }' \
+    '' \
+    '    # API 文档路由 → 代理到后端' \
+    '    location ~ ^/(docs|redoc|openapi\\.json)$ {' \
+    '        proxy_pass http://127.0.0.1:PORT_PLACEHOLDER;' \
+    '        proxy_http_version 1.1;' \
+    '        proxy_set_header Host $host;' \
+    '        proxy_set_header X-Real-IP $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）' \
+    '    location / {' \
+    '        try_files $uri $uri/ /index.html;' \
+    '    }' \
+    '}' > /etc/nginx/sites-available/default.template
 # Supervisor 配置
 RUN printf '%s\n' \
-'[supervisord]' \
-'nodaemon=true' \
-'logfile=/var/log/supervisor/supervisord.log' \
-'pidfile=/var/run/supervisord.pid' \
-'' \
-'[program:nginx]' \
-'command=/bin/bash -c "sed \"s/PORT_PLACEHOLDER/8084/g\" /etc/nginx/sites-available/default.template > /etc/nginx/sites-available/default && /usr/sbin/nginx -g \"daemon off;\""' \
-'autostart=true' \
-'autorestart=true' \
-'stdout_logfile=/var/log/nginx/access.log' \
-'stderr_logfile=/var/log/nginx/error.log' \
-'' \
-'[program:app]' \
-'command=gunicorn src.main:app -c gunicorn_conf.py --preload -w %(ENV_GUNICORN_WORKERS)s -k uvicorn.workers.UvicornWorker --bind 127.0.0.1:8084 --timeout 120 --max-requests 2000 --max-requests-jitter 100 --access-logfile - --error-logfile - --log-level info' \
-'directory=/app' \
-'autostart=true' \
-'autorestart=true' \
-'stdout_logfile=/dev/stdout' \
-'stdout_logfile_maxbytes=0' \
-'stderr_logfile=/dev/stderr' \
-'stderr_logfile_maxbytes=0' \
-'environment=PYTHONUNBUFFERED=1,PYTHONIOENCODING=utf-8,LANG=C.UTF-8,LC_ALL=C.UTF-8,DOCKER_CONTAINER=true' > /etc/supervisor/conf.d/supervisord.conf
+    '[supervisord]' \
+    'nodaemon=true' \
+    'logfile=/var/log/supervisor/supervisord.log' \
+    'pidfile=/var/run/supervisord.pid' \
+    '' \
+    '[program:nginx]' \
+    'command=/bin/bash -c "sed \"s/PORT_PLACEHOLDER/8084/g\" /etc/nginx/sites-available/default.template > /etc/nginx/sites-available/default && /usr/sbin/nginx -g \"daemon off;\""' \
+    'autostart=true' \
+    'autorestart=true' \
+    'stdout_logfile=/var/log/nginx/access.log' \
+    'stderr_logfile=/var/log/nginx/error.log' \
+    '' \
+    '[program:app]' \
+    'command=/bin/bash -c "MAX_REQUESTS_JITTER=$((${MAX_REQUESTS:-50000}/20)); exec gunicorn src.main:app -c gunicorn_conf.py --preload -w %(ENV_GUNICORN_WORKERS)s -k uvicorn.workers.UvicornWorker --bind 127.0.0.1:8084 --max-requests ${MAX_REQUESTS:-50000} --max-requests-jitter $MAX_REQUESTS_JITTER --access-logfile - --error-logfile - --log-level info"' \
+    'directory=/app' \
+    'autostart=true' \
+    'autorestart=true' \
+    'stdout_logfile=/dev/stdout' \
+    'stdout_logfile_maxbytes=0' \
+    'stderr_logfile=/dev/stderr' \
+    'stderr_logfile_maxbytes=0' \
+    'environment=PYTHONUNBUFFERED=1,PYTHONIOENCODING=utf-8,LANG=C.UTF-8,LC_ALL=C.UTF-8,DOCKER_CONTAINER=true,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' \
+    'autostart=true' \
+    'autorestart=true' \
+    'stdout_logfile=/dev/stdout' \
+    'stdout_logfile_maxbytes=0' \
+    'stderr_logfile=/dev/stderr' \
+    'stderr_logfile_maxbytes=0' > /etc/supervisor/conf.d/supervisord.conf
 # 创建目录
 RUN mkdir -p /var/log/supervisor /app/logs /app/data
 # 入口脚本（启动前执行迁移）
@@ -149,8 +279,11 @@ 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=4
+    GUNICORN_WORKERS=2 \
+    MAX_REQUESTS=4000
 EXPOSE 80
 HEALTHCHECK --interval=30s --timeout=10s --start-period=5s --retries=3 \
    CMD curl -f http://localhost/health || exit 1
@@ -2,6 +2,7 @@
 # 运行镜像：从 base 提取产物到精简运行时（国内镜像源版本）
 # 构建命令: docker build -f Dockerfile.app.local -t aether-app:latest .
 # 用于本地/国内服务器部署
+
 FROM aether-base:latest AS builder

 WORKDIR /app
@@ -11,10 +12,20 @@ COPY frontend/ ./frontend/
 RUN cd frontend && npm run build

 # ==================== 运行时镜像 ====================
-FROM python:3.14-slim
+FROM python:3.13-slim

 WORKDIR /app

+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 \
    --mount=type=cache,target=/var/lib/apt,sharing=locked \
@@ -23,16 +34,60 @@ 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/python3.14/site-packages /usr/local/lib/python3.14/site-packages
+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
 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
+# 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 ${GITHUB_TOKEN}"; \
+    fi; \
+    tag="${HUB_TAG:-}"; \
+    if [ -z "$tag" ]; then \
+        tag="$(curl -sL ${auth_header:+-H "$auth_header"} "https://api.github.com/repos/${HUB_RELEASE_REPO}/releases" | python3 -c "import json,sys;print(next((r['tag_name'] for r in json.load(sys.stdin) if r.get('tag_name','').startswith('hub-v') and not r.get('draft') and not r.get('prerelease')),''))")"; \
+    fi; \
+    if [ -z "$tag" ]; then \
+        echo "Failed to resolve hub release tag"; \
+        exit 1; \
+    fi; \
+    echo "Using Hub release tag: $tag"; \
+    origin_url="https://github.com/${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; \
+    rm -f /tmp/aether-hub.tar.gz
+
 # 从 builder 阶段复制前端构建产物
 COPY --from=builder /app/frontend/dist /usr/share/nginx/html
 RUN chmod -R 755 /usr/share/nginx/html
@@ -47,109 +102,195 @@ COPY gunicorn_conf.py ./
 # 策略：白名单后端路由 → 后端代理，其余全部 → 前端 SPA（index.html）
 # 智能处理 IP：有外层代理头就透传，没有就用直连 IP
 RUN printf '%s\n' \
-'map $http_x_real_ip $real_ip {' \
-'    default $http_x_real_ip;' \
-'    ""      $remote_addr;' \
-'}' \
-'' \
-'map $http_x_forwarded_for $forwarded_for {' \
-'    default $http_x_forwarded_for;' \
-'    ""      $remote_addr;' \
-'}' \
-'' \
-'server {' \
-'    listen 80;' \
-'    server_name _;' \
-'    root /usr/share/nginx/html;' \
-'    index index.html;' \
-'    client_max_body_size 100M;' \
-'' \
-'    # gzip 压缩配置（对 base64 图片等非流式响应有效）' \
-'    gzip on;' \
-'    gzip_min_length 256;' \
-'    gzip_comp_level 5;' \
-'    gzip_vary on;' \
-'    gzip_proxied any;' \
-'    gzip_types application/json text/plain text/css text/javascript application/javascript application/octet-stream;' \
-'    gzip_disable "msie6";' \
-'' \
-'    # 静态资源：长期缓存' \
-'    location ~* \.(js|css|png|jpg|jpeg|gif|ico|svg|woff|woff2|ttf|eot)$ {' \
-'        expires 1y;' \
-'        add_header Cache-Control "public, no-transform";' \
-'        try_files $uri =404;' \
-'    }' \
-'' \
-'    # 安全：阻止访问源码目录' \
-'    location ~ ^/(src|node_modules)/ {' \
-'        deny all;' \
-'        return 404;' \
-'    }' \
-'' \
-'    # 后端 API 路由（白名单）→ 代理到后端' \
-'    location ~ ^/(api|v1|v1beta|upload|health)(/|$) {' \
-'        proxy_pass http://127.0.0.1:PORT_PLACEHOLDER;' \
-'        proxy_http_version 1.1;' \
-'        proxy_set_header Host $host;' \
-'        proxy_set_header X-Real-IP $real_ip;' \
-'        proxy_set_header X-Forwarded-For $forwarded_for;' \
-'        proxy_set_header X-Forwarded-Proto $scheme;' \
-'        proxy_set_header Connection "";' \
-'        proxy_set_header Accept $http_accept;' \
-'        proxy_set_header Content-Type $content_type;' \
-'        proxy_set_header Authorization $http_authorization;' \
-'        proxy_set_header X-Api-Key $http_x_api_key;' \
-'        proxy_buffering off;' \
-'        proxy_cache off;' \
-'        proxy_request_buffering off;' \
-'        chunked_transfer_encoding on;' \
-'        gzip off;' \
-'        add_header X-Accel-Buffering no;' \
-'        proxy_connect_timeout 600s;' \
-'        proxy_send_timeout 600s;' \
-'        proxy_read_timeout 600s;' \
-'    }' \
-'' \
-'    # API 文档路由 → 代理到后端' \
-'    location ~ ^/(docs|redoc|openapi\\.json)$ {' \
-'        proxy_pass http://127.0.0.1:PORT_PLACEHOLDER;' \
-'        proxy_http_version 1.1;' \
-'        proxy_set_header Host $host;' \
-'        proxy_set_header X-Real-IP $real_ip;' \
-'        proxy_set_header X-Forwarded-For $forwarded_for;' \
-'        proxy_set_header X-Forwarded-Proto $scheme;' \
-'    }' \
-'' \
-'    # 所有其他路由 → 前端 SPA（先尝试静态文件，再回退到 index.html）' \
-'    location / {' \
-'        try_files $uri $uri/ /index.html;' \
-'    }' \
-'}' > /etc/nginx/sites-available/default.template
+    'map $http_x_real_ip $real_ip {' \
+    '    default $http_x_real_ip;' \
+    '    ""      $remote_addr;' \
+    '}' \
+    '' \
+    'map $http_x_forwarded_for $forwarded_for {' \
+    '    default $http_x_forwarded_for;' \
+    '    ""      $remote_addr;' \
+    '}' \
+    '' \
+    'map $http_upgrade $connection_upgrade {' \
+    '    default upgrade;' \
+    '    ""      "";' \
+    '}' \
+    '' \
+    'server {' \
+    '    listen 80;' \
+    '    server_name _;' \
+    '    root /usr/share/nginx/html;' \
+    '    index index.html;' \
+    '    client_max_body_size 100M;' \
+    '' \
+    '    # gzip 压缩配置（对 base64 图片等非流式响应有效）' \
+    '    gzip on;' \
+    '    gzip_min_length 256;' \
+    '    gzip_comp_level 5;' \
+    '    gzip_vary on;' \
+    '    gzip_proxied any;' \
+    '    gzip_types application/json text/plain text/css text/javascript application/javascript application/octet-stream;' \
+    '    gzip_disable "msie6";' \
+    '' \
+    '    # 静态资源：长期缓存' \
+    '    location ~* \.(js|css|png|jpg|jpeg|gif|ico|svg|woff|woff2|ttf|eot)$ {' \
+    '        expires 1y;' \
+    '        add_header Cache-Control "public, no-transform";' \
+    '        try_files $uri =404;' \
+    '    }' \
+    '' \
+    '    # 安全：阻止访问源码目录' \
+    '    location ~ ^/(src|node_modules)/ {' \
+    '        deny all;' \
+    '        return 404;' \
+    '    }' \
+    '' \
+    '    # WebSocket 隧道端点（aether-proxy tunnel 模式）' \
+    '    location = /api/internal/proxy-tunnel {' \
+    '        proxy_pass http://127.0.0.1:8085/proxy;' \
+    '        proxy_http_version 1.1;' \
+    '        proxy_set_header Host $host;' \
+    '        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 Upgrade $http_upgrade;' \
+    '        proxy_set_header Connection "upgrade";' \
+    '        # 剥离 CF 头，防止泄露给上游或返回给客户端' \
+    '        proxy_hide_header CF-Connecting-IP;' \
+    '        proxy_hide_header CF-IPCountry;' \
+    '        proxy_hide_header CF-Ray;' \
+    '        proxy_hide_header CF-Visitor;' \
+    '        proxy_hide_header CDN-Loop;' \
+    '        proxy_hide_header True-Client-IP;' \
+    '        proxy_hide_header CF-Worker;' \
+    '        proxy_hide_header CF-EW-Via;' \
+    '        proxy_hide_header CF-Warp-Tag-ID;' \
+    '        proxy_set_header CF-Connecting-IP "";' \
+    '        proxy_set_header CF-IPCountry "";' \
+    '        proxy_set_header CF-Ray "";' \
+    '        proxy_set_header CF-Visitor "";' \
+    '        proxy_set_header CDN-Loop "";' \
+    '        proxy_set_header True-Client-IP "";' \
+    '        proxy_set_header CF-Worker "";' \
+    '        proxy_set_header CF-EW-Via "";' \
+    '        proxy_set_header CF-Warp-Tag-ID "";' \
+    '        proxy_read_timeout 86400s;' \
+    '        proxy_send_timeout 86400s;' \
+    '    }' \
+    '' \
+    '    # 后端 API 路由（白名单）→ 代理到后端' \
+    '    location ~ ^/(api|v1|v1beta|upload|health)(/|$) {' \
+    '        proxy_pass http://127.0.0.1:PORT_PLACEHOLDER;' \
+    '        proxy_http_version 1.1;' \
+    '        proxy_set_header Host $host;' \
+    '        proxy_set_header X-Real-IP $real_ip;' \
+    '        proxy_set_header X-Forwarded-For $forwarded_for;' \
+    '        proxy_set_header X-Forwarded-Proto $scheme;' \
+    '        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 $http_x_api_key;' \
+    '        # 剥离 CF 头，防止泄露给上游或返回给客户端' \
+    '        proxy_hide_header CF-Connecting-IP;' \
+    '        proxy_hide_header CF-IPCountry;' \
+    '        proxy_hide_header CF-Ray;' \
+    '        proxy_hide_header CF-Visitor;' \
+    '        proxy_hide_header CDN-Loop;' \
+    '        proxy_hide_header True-Client-IP;' \
+    '        proxy_hide_header CF-Worker;' \
+    '        proxy_hide_header CF-EW-Via;' \
+    '        proxy_hide_header CF-Warp-Tag-ID;' \
+    '        proxy_set_header CF-Connecting-IP "";' \
+    '        proxy_set_header CF-IPCountry "";' \
+    '        proxy_set_header CF-Ray "";' \
+    '        proxy_set_header CF-Visitor "";' \
+    '        proxy_set_header CDN-Loop "";' \
+    '        proxy_set_header True-Client-IP "";' \
+    '        proxy_set_header CF-Worker "";' \
+    '        proxy_set_header CF-EW-Via "";' \
+    '        proxy_set_header CF-Warp-Tag-ID "";' \
+    '        proxy_buffering off;' \
+    '        proxy_cache off;' \
+    '        proxy_request_buffering off;' \
+    '        chunked_transfer_encoding on;' \
+    '        gzip off;' \
+    '        add_header X-Accel-Buffering no;' \
+    '        proxy_connect_timeout 60s;' \
+    '        proxy_send_timeout 3600s;' \
+    '        proxy_read_timeout 3600s;' \
+    '    }' \
+    '' \
+    '    # API 文档路由 → 代理到后端' \
+    '    location ~ ^/(docs|redoc|openapi\\.json)$ {' \
+    '        proxy_pass http://127.0.0.1:PORT_PLACEHOLDER;' \
+    '        proxy_http_version 1.1;' \
+    '        proxy_set_header Host $host;' \
+    '        proxy_set_header X-Real-IP $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）' \
+    '    location / {' \
+    '        try_files $uri $uri/ /index.html;' \
+    '    }' \
+    '}' > /etc/nginx/sites-available/default.template

 # Supervisor 配置
 RUN printf '%s\n' \
-'[supervisord]' \
-'nodaemon=true' \
-'logfile=/var/log/supervisor/supervisord.log' \
-'pidfile=/var/run/supervisord.pid' \
-'' \
-'[program:nginx]' \
-'command=/bin/bash -c "sed \"s/PORT_PLACEHOLDER/${PORT:-8084}/g\" /etc/nginx/sites-available/default.template > /etc/nginx/sites-available/default && /usr/sbin/nginx -g \"daemon off;\""' \
-'autostart=true' \
-'autorestart=true' \
-'stdout_logfile=/var/log/nginx/access.log' \
-'stderr_logfile=/var/log/nginx/error.log' \
-'' \
-'[program:app]' \
-'command=gunicorn src.main:app -c gunicorn_conf.py --preload -w %(ENV_GUNICORN_WORKERS)s -k uvicorn.workers.UvicornWorker --bind 0.0.0.0:%(ENV_PORT)s --timeout 120 --max-requests 2000 --max-requests-jitter 100 --access-logfile - --error-logfile - --log-level info' \
-'directory=/app' \
-'autostart=true' \
-'autorestart=true' \
-'stdout_logfile=/dev/stdout' \
-'stdout_logfile_maxbytes=0' \
-'stderr_logfile=/dev/stderr' \
-'stderr_logfile_maxbytes=0' \
-'environment=PYTHONUNBUFFERED=1,PYTHONIOENCODING=utf-8,LANG=C.UTF-8,LC_ALL=C.UTF-8,DOCKER_CONTAINER=true' > /etc/supervisor/conf.d/supervisord.conf
+    '[supervisord]' \
+    'nodaemon=true' \
+    'logfile=/var/log/supervisor/supervisord.log' \
+    'pidfile=/var/run/supervisord.pid' \
+    '' \
+    '[program:nginx]' \
+    'command=/bin/bash -c "sed \"s/PORT_PLACEHOLDER/${PORT:-8084}/g\" /etc/nginx/sites-available/default.template > /etc/nginx/sites-available/default && /usr/sbin/nginx -g \"daemon off;\""' \
+    'autostart=true' \
+    'autorestart=true' \
+    'stdout_logfile=/var/log/nginx/access.log' \
+    'stderr_logfile=/var/log/nginx/error.log' \
+    '' \
+    '[program:app]' \
+    'command=/bin/bash -c "MAX_REQUESTS_JITTER=$((${MAX_REQUESTS:-50000}/20)); exec gunicorn src.main:app -c gunicorn_conf.py --preload -w %(ENV_GUNICORN_WORKERS)s -k uvicorn.workers.UvicornWorker --bind 0.0.0.0:%(ENV_PORT)s --max-requests ${MAX_REQUESTS:-50000} --max-requests-jitter $MAX_REQUESTS_JITTER --access-logfile - --error-logfile - --log-level info"' \
+    'directory=/app' \
+    'autostart=true' \
+    'autorestart=true' \
+    'stdout_logfile=/dev/stdout' \
+    'stdout_logfile_maxbytes=0' \
+    'stderr_logfile=/dev/stderr' \
+    'stderr_logfile_maxbytes=0' \
+    'environment=PYTHONUNBUFFERED=1,PYTHONIOENCODING=utf-8,LANG=C.UTF-8,LC_ALL=C.UTF-8,DOCKER_CONTAINER=true,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' \
+    'autostart=true' \
+    'autorestart=true' \
+    'stdout_logfile=/dev/stdout' \
+    'stdout_logfile_maxbytes=0' \
+    'stderr_logfile=/dev/stderr' \
+    'stderr_logfile_maxbytes=0' > /etc/supervisor/conf.d/supervisord.conf

 # 创建目录
 RUN mkdir -p /var/log/supervisor /app/logs /app/data
@@ -164,8 +305,11 @@ 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=4
+    GUNICORN_WORKERS=2 \
+    MAX_REQUESTS=4000

 EXPOSE 80

@@ -3,7 +3,7 @@
 # 用于 GitHub Actions CI 构建（不使用国内镜像源）
 # 构建命令: docker build -f Dockerfile.base -t aether-base:latest .
 # 只在 pyproject.toml 或 frontend/package*.json 变化时需要重建
-FROM python:3.14-slim
+FROM python:3.13-slim

 WORKDIR /app

@@ -2,7 +2,7 @@
 # 构建镜像：编译环境 + 预编译的依赖（国内镜像源版本）
 # 构建命令: docker build -f Dockerfile.base.local -t aether-base:latest .
 # 只在 pyproject.toml 或 frontend/package*.json 变化时需要重建
-FROM python:3.14-slim
+FROM python:3.13-slim

 WORKDIR /app

@@ -5,8 +5,8 @@
 <h1 align="center">Aether</h1>

 <p align="center">
-  <strong>开源 AI API 网关</strong><br>
-  支持 Claude / OpenAI / Gemini 及其 CLI 客户端的统一接入层
+  <strong>一站式 AI 基础设施平台</strong><br>
+  支持 Claude / OpenAI / Gemini 及其 CLI 客户端的统一接入、格式转换、正/反向代理, 致力于成为用户驱动AI服务的底座
 </p>
 <p align="center">
  <a href="#简介">简介</a> •
@@ -22,27 +22,15 @@

 Aether 是一个自托管的 AI API 网关，为团队和个人提供多租户管理、智能负载均衡、成本配额控制和健康监控能力。通过统一的 API 入口，可以无缝对接 Claude、OpenAI、Gemini 等主流 AI 服务及其 CLI 工具。

-### 页面预览
+<p align="center">
+  <picture>
+    <source media="(prefers-color-scheme: dark)" srcset="docs/architecture/architecture-dark.svg">
+    <source media="(prefers-color-scheme: light)" srcset="docs/architecture/architecture-light.svg">
+    <img src="docs/architecture/architecture-light.svg" width="680" alt="Aether Architecture">
+  </picture>
+</p>

-| 首页 | 仪表盘 |
-|:---:|:---:|
-| ![首页](docs/screenshots/home.png) | ![仪表盘](docs/screenshots/dashboard.png) |
-
-| 健康监控 | 用户管理 |
-|:---:|:---:|
-| ![健康监控](docs/screenshots/health.png) | ![用户管理](docs/screenshots/users.png) |
-
-| 提供商管理 | 使用记录 |
-|:---:|:---:|
-| ![提供商管理](docs/screenshots/providers.png) | ![使用记录](docs/screenshots/usage.png) |
-
-| 模型详情 | 关联提供商 |
-|:---:|:---:|
-| ![模型详情](docs/screenshots/model-detail.png) | ![关联提供商](docs/screenshots/model-providers.png) |
-
-| 链路追踪 | 系统设置 |
-|:---:|:---:|
-| ![链路追踪](docs/screenshots/tracing.png) | ![系统设置](docs/screenshots/settings.png) |
+页面预览: https://fawney19.github.io/Aether/

 ## 部署

@@ -60,7 +48,7 @@ python generate_keys.py  # 生成密钥, 并将生成的密钥填入 .env
 # 3. 部署 / 更新（自动执行数据库迁移）
 docker compose pull && docker compose up -d

-# 4. 升级前备份
+# 4. 升级前备份 (可选)
 docker compose exec postgres pg_dump -U postgres aether | gzip > backup_$(date +%Y%m%d_%H%M%S).sql.gz
 ```

@@ -76,6 +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
 ```

@@ -93,6 +82,14 @@ uv sync
 cd frontend && npm install && npm run dev
 ```

+## Aether Proxy (可选)
+
+Aether Proxy 是配套的正向代理节点，部署在海外 VPS 上，为墙内的 Aether 实例中转 API 流量。或者部署在其他服务器为指定的提供商、账号、Key使用不同的节点访问。支持 TUI 向导一键配置、systemd 服务管理、TLS 加密、DNS 缓存及连接池调优。
+
+- Docker Compose 部署或下载预编译二进制直接运行
+- 通过 `aether-proxy setup` 完成交互式配置，自动注册为系统服务
+- 详细文档见 [aether-proxy/README.md](aether-proxy/README.md)
+
 ## 环境变量

 ### 必需配置
@@ -114,7 +111,7 @@ cd frontend && npm install && npm run dev
 | `APP_PORT` | 8084 | 应用端口 |
 | `API_KEY_PREFIX` | sk | API Key 前缀 |
 | `LOG_LEVEL` | INFO | 日志级别 (DEBUG/INFO/WARNING/ERROR) |
-| `GUNICORN_WORKERS` | 4 | Gunicorn 工作进程数 |
+| `GUNICORN_WORKERS` | 2 | Gunicorn 工作进程数 |
 | `DB_PORT` | 5432 | PostgreSQL 端口 |
 | `REDIS_PORT` | 6379 | Redis 端口 |

@@ -130,27 +127,6 @@ cd frontend && npm install && npm run dev
 | Headers | Base + 请求头 |
 | Full | Headers + 请求体 |

-### Q: 管理员如何给模型配置 1M上下文 / 1H缓存 能力支持?
-
-1. **模型管理**: 给模型设置 1M上下文 / 1H缓存 的能力支持, 并配置好价格
-2. **提供商管理**: 给端点添加支持该能力的密钥, 并勾选对应的能力标签
-
-### Q: 用户如何使用 1H缓存?
-
- **模型级别**: 在模型管理中针对指定模型开启 1H缓存策略
- **密钥级别**: 在密钥管理中针对指定密钥使用 1H缓存策略
-
-> **注意**: 若对密钥设置强制 1H缓存, 则该密钥只能使用支持 1H缓存的模型, 匹配提供商Key, 将会导致这个Key无法同时用于Claude Code、Codex、GeminiCLI, 因为更推荐使用模型开启1H缓存.
-
-### Q: 如何配置负载均衡？
-
-在管理后台 **提供商管理** 中切换调度模式:
-
-| 模式 | 说明 | 适用场景 |
-|------|------|----------|
-| **提供商优先** | 按 Provider 优先级排序, 同优先级内按 Key 优先级排序, 相同优先级哈希分散 | 优先使用特定供应商 |
-| **全局 Key 优先** | 忽略 Provider 层级, 所有 Key 按全局优先级统一排序, 相同优先级哈希分散 | 跨 Provider 统一调度, 最大化利用所有 Key |
-
 ### Q: 更新出问题如何回滚？

 **有备份的情况（推荐）：**
@@ -205,4 +181,3 @@ docker compose up -d app
 ## Star History

 [![Star History Chart](https://api.star-history.com/svg?repos=fawney19/Aether&type=Date)](https://star-history.com/#fawney19/Aether&Date)
-
@@ -0,0 +1,3 @@
+target/
+.git/
+.DS_Store
@@ -0,0 +1,27 @@
+[package]
+name = "aether-hub"
+version = "0.2.0"
+edition = "2021"
+description = "Tunnel Hub for Aether - frame router between workers and proxies"
+
+[dependencies]
+tokio = { version = "1", features = ["full"] }
+axum = { version = "0.8", features = ["ws"] }
+serde = { version = "1", features = ["derive"] }
+serde_json = "1"
+tracing = "0.1"
+tracing-subscriber = { version = "0.3", features = ["env-filter"] }
+clap = { version = "4", features = ["derive", "env"] }
+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
+strip = true
+codegen-units = 1
@@ -0,0 +1,34 @@
+# syntax=docker/dockerfile:1
+
+FROM rust:1.85-slim AS builder
+WORKDIR /build/aether-hub
+
+# 可选：配置国内 Cargo 镜像源（本地构建时传 --build-arg CARGO_MIRROR=1）
+ARG CARGO_MIRROR
+RUN if [ -n "$CARGO_MIRROR" ]; then \
+    printf '[source.crates-io]\nreplace-with = "tuna"\n\n[source.tuna]\nregistry = "sparse+https://mirrors.tuna.tsinghua.edu.cn/crates.io-index/"\n' \
+    > /usr/local/cargo/config.toml; \
+    fi
+
+# 先构建依赖层，最大化后续代码变更时的缓存命中
+COPY Cargo.toml Cargo.lock ./
+RUN mkdir src && printf 'fn main() {}\n' > src/main.rs
+RUN --mount=type=cache,target=/usr/local/cargo/registry,sharing=locked \
+    --mount=type=cache,target=/build/aether-hub/target,sharing=locked \
+    cargo build --release --locked
+RUN rm -rf src
+
+COPY src ./src
+RUN --mount=type=cache,target=/usr/local/cargo/registry,sharing=locked \
+    --mount=type=cache,target=/build/aether-hub/target,sharing=locked \
+    cargo build --release --locked && \
+    cp target/release/aether-hub /tmp/aether-hub
+
+FROM debian:bookworm-slim
+RUN apt-get update && apt-get install -y --no-install-recommends ca-certificates && \
+    rm -rf /var/lib/apt/lists/*
+COPY --from=builder /tmp/aether-hub /usr/local/bin/aether-hub
+
+EXPOSE 8085
+ENTRYPOINT ["/usr/local/bin/aether-hub"]
+CMD ["--bind", "0.0.0.0:8085"]
@@ -0,0 +1,38 @@
+# aether-hub
+
+`aether-hub` 是 Tunnel Hub 服务，负责在 proxy 与 worker 之间路由帧。
+
+已集成在Docker镜像中, 无需单独部署。
+
+## 部署端指定 Hub 版本并构建
+
+```bash
+cd /path/to/Aether
+./deploy.sh --hub-tag hub-v0.1.0
+```
+
+不指定 `--hub-tag` 时，`./deploy.sh` 会自动解析最新 `hub-v*` release，并在构建 app 镜像时从 GitHub Release 下载对应架构的 Hub 二进制。
+
+## build.sh 模式说明
+
+- 默认是 `binary` 模式（`cross` 构建二进制）。
+- `--upload <hub-vX.Y.Z>` 会把构建产物上传到 GitHub Release。
+- 加 `--image` 后进入镜像模式（`docker buildx`，可选）。
+
+常用参数：
+
+- `--tag <tag>`: 镜像 tag
+- `--image-name <name>`: 镜像名（默认 `ghcr.io/fawney19/aether-hub`）
+- `--platforms <list>`: 例如 `linux/amd64,linux/arm64`
+- `--push`: 推送镜像
+- `--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 +1,272 @@
+#!/bin/bash
+# aether-hub 构建脚本
+#
+# 支持两种模式:
+# 1) binary 模式（默认）: 构建多架构二进制并可上传 GitHub Release
+# 2) image 模式: 构建并推送/加载 Docker 镜像（推荐生产发布用）
+#
+# 示例:
+#   # binary 模式（兼容旧行为）
+#   ./build.sh
+#   ./build.sh amd64
+#   ./build.sh --upload hub-v0.1.0
+#
+#   # image 模式（多架构推送）
+#   ./build.sh --image --tag v0.2.5 --push --latest
+#   ./build.sh --image --tag sha-abc123 --image-name ghcr.io/fawney19/aether-hub --push
+#   ./build.sh --image --tag local-test --platforms linux/amd64 --load
+
+set -euo pipefail
+
+SCRIPT_DIR="$(cd "$(dirname "$0")" && pwd)"
+PROJECT_DIR="$(cd "$SCRIPT_DIR/.." && pwd)"
+DIST_DIR="$SCRIPT_DIR/dist"
+
+# -------------------------------
+# Defaults
+# -------------------------------
+MODE="binary"                  # binary | image
+
+# binary mode options
+UPLOAD=false
+UPLOAD_TAG=""
+BINARY_TARGETS=""
+
+# image mode options
+IMAGE_NAME="${IMAGE_NAME:-ghcr.io/fawney19/aether-hub}"
+IMAGE_TAG=""
+IMAGE_PLATFORMS="linux/amd64,linux/arm64"
+IMAGE_PUSH=false
+IMAGE_LOAD=false
+IMAGE_LATEST=false
+
+usage() {
+    cat <<'EOF'
+用法:
+  ./build.sh [binary-args]
+  ./build.sh --image [image-args]
+
+binary 模式（默认）:
+  amd64|arm64              仅构建指定架构（可重复）
+  --upload <hub-vX.Y.Z>    上传到 GitHub Release（需要 gh CLI）
+
+image 模式:
+  --image                  启用镜像模式
+  --tag <tag>              镜像 tag（默认自动从 git describe 推导）
+  --image-name <name>      镜像名（默认 ghcr.io/fawney19/aether-hub）
+  --platforms <list>       平台列表，逗号分隔（默认 linux/amd64,linux/arm64）
+  --push                   推送镜像到仓库
+  --load                   加载到本地 Docker（仅单平台）
+  --latest                 额外打 latest tag
+
+通用:
+  -h, --help               显示帮助
+EOF
+}
+
+while [ $# -gt 0 ]; do
+    case "$1" in
+        --image)
+            MODE="image"
+            shift
+            ;;
+        --tag)
+            IMAGE_TAG="${2:-}"
+            shift 2
+            ;;
+        --image-name)
+            IMAGE_NAME="${2:-}"
+            shift 2
+            ;;
+        --platforms)
+            IMAGE_PLATFORMS="${2:-}"
+            shift 2
+            ;;
+        --push)
+            IMAGE_PUSH=true
+            shift
+            ;;
+        --load)
+            IMAGE_LOAD=true
+            shift
+            ;;
+        --latest)
+            IMAGE_LATEST=true
+            shift
+            ;;
+        --upload)
+            UPLOAD=true
+            UPLOAD_TAG="${2:-}"
+            shift 2
+            ;;
+        amd64|arm64)
+            BINARY_TARGETS="$BINARY_TARGETS $1"
+            shift
+            ;;
+        -h|--help)
+            usage
+            exit 0
+            ;;
+        *)
+            echo "❌ 未知参数: $1"
+            usage
+            exit 1
+            ;;
+    esac
+done
+
+build_binary() {
+    if [ -z "$BINARY_TARGETS" ]; then
+        BINARY_TARGETS="amd64 arm64"
+    fi
+
+    if ! command -v cross >/dev/null 2>&1; then
+        echo "❌ 需要安装 cross: cargo install cross --git https://github.com/cross-rs/cross"
+        exit 1
+    fi
+
+    mkdir -p "$DIST_DIR"
+
+    echo "🔨 开始构建 aether-hub 二进制..."
+    echo "   目标平台: $BINARY_TARGETS"
+    echo ""
+
+    ARTIFACTS=""
+    for arch in $BINARY_TARGETS; do
+        case "$arch" in
+            amd64) target="x86_64-unknown-linux-gnu" ;;
+            arm64) target="aarch64-unknown-linux-gnu" ;;
+            *) echo "❌ 未知架构: $arch"; exit 1 ;;
+        esac
+
+        echo ">>> 构建 $arch ($target)..."
+        cd "$SCRIPT_DIR"
+        cross build --release --target "$target" --locked
+
+        BIN="target/$target/release/aether-hub"
+        if [ ! -f "$BIN" ]; then
+            echo "❌ 未找到二进制文件: $BIN"
+            exit 1
+        fi
+
+        ARCHIVE="$DIST_DIR/aether-hub-linux-$arch.tar.gz"
+        tar czf "$ARCHIVE" -C "target/$target/release" aether-hub
+        ARTIFACTS="$ARTIFACTS $ARCHIVE"
+
+        SIZE=$(du -h "$ARCHIVE" | cut -f1)
+        echo "✅ $arch 构建完成: $ARCHIVE ($SIZE)"
+        echo ""
+    done
+
+    cd "$DIST_DIR"
+    shasum -a 256 aether-hub-*.tar.gz > SHA256SUMS.txt
+    echo "📋 SHA256 校验和:"
+    cat SHA256SUMS.txt
+    echo ""
+
+    if [ "$UPLOAD" = true ]; then
+        if [ -z "$UPLOAD_TAG" ]; then
+            echo "❌ --upload 需要指定 tag，例如: ./build.sh --upload hub-v0.1.0"
+            exit 1
+        fi
+        if ! command -v gh >/dev/null 2>&1; then
+            echo "❌ 需要安装 GitHub CLI: brew install gh"
+            exit 1
+        fi
+
+        echo "📦 上传到 GitHub Release: $UPLOAD_TAG"
+        cd "$PROJECT_DIR"
+
+        if ! git rev-parse "$UPLOAD_TAG" >/dev/null 2>&1; then
+            git tag "$UPLOAD_TAG"
+            git push origin "$UPLOAD_TAG"
+        fi
+
+        gh release create "$UPLOAD_TAG" \
+            --title "aether-hub ${UPLOAD_TAG#hub-}" \
+            --generate-notes \
+            $ARTIFACTS \
+            "$DIST_DIR/SHA256SUMS.txt"
+
+        echo "✅ 上传完成!"
+    fi
+
+    echo "🎉 binary 模式完成!"
+}
+
+build_image() {
+    if ! command -v docker >/dev/null 2>&1; then
+        echo "❌ 未找到 docker，请先安装 Docker"
+        exit 1
+    fi
+    if ! docker buildx version >/dev/null 2>&1; then
+        echo "❌ 未找到 docker buildx，请先启用 buildx"
+        exit 1
+    fi
+
+    if [ "$IMAGE_PUSH" = true ] && [ "$IMAGE_LOAD" = true ]; then
+        echo "❌ --push 与 --load 不能同时使用"
+        exit 1
+    fi
+
+    if [ "$IMAGE_PUSH" = false ] && [ "$IMAGE_LOAD" = false ]; then
+        # image 模式默认走 push，符合发布场景
+        IMAGE_PUSH=true
+    fi
+
+    if [ -z "$IMAGE_TAG" ]; then
+        IMAGE_TAG=$(git -C "$PROJECT_DIR" describe --tags --always 2>/dev/null | sed 's/^v//')
+        if [ -z "$IMAGE_TAG" ]; then
+            IMAGE_TAG=$(date +%Y%m%d%H%M%S)
+        fi
+    fi
+
+    if [ "$IMAGE_LOAD" = true ] && [[ "$IMAGE_PLATFORMS" == *,* ]]; then
+        echo "❌ --load 仅支持单平台，请用 --platforms linux/amd64（或 arm64）"
+        exit 1
+    fi
+
+    local ref="${IMAGE_NAME}:${IMAGE_TAG}"
+    local cmd=(docker buildx build
+        --platform "$IMAGE_PLATFORMS"
+        -f "$SCRIPT_DIR/Dockerfile"
+        -t "$ref"
+    )
+
+    if [ "$IMAGE_LATEST" = true ]; then
+        cmd+=(-t "${IMAGE_NAME}:latest")
+    fi
+
+    if [ "$IMAGE_PUSH" = true ]; then
+        cmd+=(--push)
+    else
+        cmd+=(--load)
+    fi
+
+    cmd+=("$SCRIPT_DIR")
+
+    echo "🔨 开始构建 aether-hub 镜像..."
+    echo "   image: $ref"
+    echo "   platforms: $IMAGE_PLATFORMS"
+    echo "   mode: $([ "$IMAGE_PUSH" = true ] && echo push || echo load)"
+    echo ""
+
+    "${cmd[@]}"
+
+    if [ "$IMAGE_PUSH" = true ]; then
+        echo "✅ 镜像已推送: $ref"
+        if [ "$IMAGE_LATEST" = true ]; then
+            echo "✅ 镜像已推送: ${IMAGE_NAME}:latest"
+        fi
+    else
+        echo "✅ 镜像已加载到本地: $ref"
+    fi
+
+    echo "🎉 image 模式完成!"
+}
+
+if [ "$MODE" = "image" ]; then
+    build_image
+else
+    build_binary
+fi
@@ -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 +1,878 @@
+use std::collections::HashMap;
+use std::sync::atomic::{AtomicBool, AtomicU32, AtomicU64, AtomicUsize, Ordering};
+use std::sync::Arc;
+use std::time::Duration;
+
+use axum::extract::ws::Message;
+use bytes::Bytes;
+use dashmap::DashMap;
+use parking_lot::{Mutex, RwLock};
+use tokio::sync::mpsc;
+use tokio::sync::mpsc::error::TrySendError;
+use tokio::sync::{watch, Notify};
+use tracing::{debug, info, warn};
+
+use crate::control_plane::ControlPlaneClient;
+use crate::protocol;
+
+const MAX_REQUEST_BODY_FRAME_SIZE: usize = 32 * 1024;
+
+#[derive(Debug, Clone, Copy, PartialEq, Eq)]
+pub enum SendStatus {
+    Queued,
+    Closed,
+    Congested,
+}
+
+#[derive(Debug, Clone, Copy)]
+pub struct ConnConfig {
+    pub ping_interval: Duration,
+    pub idle_timeout: Duration,
+    pub outbound_queue_capacity: usize,
+}
+
+pub struct BoundedOutbound {
+    tx: mpsc::Sender<Message>,
+    close_tx: watch::Sender<bool>,
+    closing: AtomicBool,
+}
+
+impl BoundedOutbound {
+    pub fn new(tx: mpsc::Sender<Message>, close_tx: watch::Sender<bool>) -> Self {
+        Self {
+            tx,
+            close_tx,
+            closing: AtomicBool::new(false),
+        }
+    }
+
+    pub fn send(&self, msg: Message) -> SendStatus {
+        if self.is_closing() {
+            return SendStatus::Closed;
+        }
+
+        match self.tx.try_send(msg) {
+            Ok(()) => SendStatus::Queued,
+            Err(TrySendError::Closed(_)) => {
+                self.mark_closing();
+                SendStatus::Closed
+            }
+            Err(TrySendError::Full(_)) => {
+                self.mark_closing();
+                SendStatus::Congested
+            }
+        }
+    }
+
+    pub fn is_closing(&self) -> bool {
+        self.closing.load(Ordering::Acquire)
+    }
+
+    pub fn mark_closing(&self) -> bool {
+        if self.closing.swap(true, Ordering::AcqRel) {
+            return false;
+        }
+        let _ = self.close_tx.send(true);
+        true
+    }
+}
+
+pub struct ProxyConn {
+    pub id: u64,
+    pub node_id: String,
+    pub node_name: String,
+    pub outbound: BoundedOutbound,
+    next_stream_id: AtomicU32,
+    pub stream_count: AtomicUsize,
+    pub max_streams: usize,
+}
+
+impl ProxyConn {
+    pub fn new(
+        id: u64,
+        node_id: String,
+        node_name: String,
+        tx: mpsc::Sender<Message>,
+        close_tx: watch::Sender<bool>,
+        max_streams: usize,
+    ) -> Self {
+        Self {
+            id,
+            node_id,
+            node_name,
+            outbound: BoundedOutbound::new(tx, close_tx),
+            next_stream_id: AtomicU32::new(2),
+            stream_count: AtomicUsize::new(0),
+            max_streams,
+        }
+    }
+
+    pub fn alloc_stream_id(&self) -> Option<u32> {
+        let mut current = self.stream_count.load(Ordering::Relaxed);
+        loop {
+            if current >= self.max_streams || !self.is_available() {
+                return None;
+            }
+            match self.stream_count.compare_exchange_weak(
+                current,
+                current + 1,
+                Ordering::AcqRel,
+                Ordering::Relaxed,
+            ) {
+                Ok(_) => break,
+                Err(observed) => current = observed,
+            }
+        }
+
+        let sid = loop {
+            let current_sid = self.next_stream_id.load(Ordering::Relaxed);
+            let next_sid = if current_sid >= 0xFFFF_FFFE {
+                2
+            } else {
+                current_sid + 2
+            };
+            if self
+                .next_stream_id
+                .compare_exchange_weak(current_sid, next_sid, Ordering::AcqRel, Ordering::Relaxed)
+                .is_ok()
+            {
+                break current_sid;
+            }
+        };
+
+        Some(sid)
+    }
+
+    pub fn release_stream(&self) {
+        let mut current = self.stream_count.load(Ordering::Relaxed);
+        while current > 0 {
+            match self.stream_count.compare_exchange_weak(
+                current,
+                current - 1,
+                Ordering::AcqRel,
+                Ordering::Relaxed,
+            ) {
+                Ok(_) => return,
+                Err(observed) => current = observed,
+            }
+        }
+    }
+
+    pub fn is_available(&self) -> bool {
+        !self.outbound.is_closing()
+    }
+
+    pub fn request_close(&self) {
+        self.outbound.mark_closing();
+    }
+
+    pub fn send(&self, msg: Message) -> SendStatus {
+        let was_closing = self.outbound.is_closing();
+        let status = self.outbound.send(msg);
+        if status == SendStatus::Congested && !was_closing {
+            warn!(
+                conn_id = self.id,
+                node_id = %self.node_id,
+                node_name = %self.node_name,
+                queued_streams = self.stream_count.load(Ordering::Relaxed),
+                "proxy outbound queue full, closing congested connection"
+            );
+        }
+        status
+    }
+}
+
+#[derive(Debug, Clone)]
+pub struct LocalResponseHead {
+    pub status: u16,
+    pub headers: Vec<(String, String)>,
+}
+
+#[derive(Debug)]
+pub enum LocalBodyEvent {
+    Chunk(Bytes),
+    End,
+    Error(String),
+}
+
+#[derive(Debug, Default)]
+struct LocalWaitState {
+    response: Option<LocalResponseHead>,
+    error: Option<String>,
+}
+
+pub struct LocalStream {
+    pub id: u64,
+    proxy_conn_id: u64,
+    proxy_stream_id: u32,
+    wait_state: Mutex<LocalWaitState>,
+    headers_notify: Notify,
+    body_tx: mpsc::Sender<LocalBodyEvent>,
+    body_rx: Mutex<Option<mpsc::Receiver<LocalBodyEvent>>>,
+    terminal: AtomicBool,
+}
+
+impl LocalStream {
+    fn new(id: u64, proxy_conn_id: u64, proxy_stream_id: u32) -> Self {
+        let (body_tx, body_rx) = mpsc::channel(128);
+        Self {
+            id,
+            proxy_conn_id,
+            proxy_stream_id,
+            wait_state: Mutex::new(LocalWaitState::default()),
+            headers_notify: Notify::new(),
+            body_tx,
+            body_rx: Mutex::new(Some(body_rx)),
+            terminal: AtomicBool::new(false),
+        }
+    }
+
+    pub async fn wait_headers(&self, timeout: Duration) -> Result<LocalResponseHead, String> {
+        tokio::time::timeout(timeout, async {
+            loop {
+                let outcome = {
+                    let state = self.wait_state.lock();
+                    if let Some(response) = &state.response {
+                        return Ok(response.clone());
+                    }
+                    state.error.clone()
+                };
+                if let Some(error) = outcome {
+                    return Err(error);
+                }
+                self.headers_notify.notified().await;
+            }
+        })
+        .await
+        .map_err(|_| "timed out waiting for response headers".to_string())?
+    }
+
+    pub fn take_body_receiver(&self) -> Option<mpsc::Receiver<LocalBodyEvent>> {
+        self.body_rx.lock().take()
+    }
+
+    fn set_response_headers(&self, meta: protocol::ResponseMeta) {
+        let mut notify = false;
+        {
+            let mut state = self.wait_state.lock();
+            if state.response.is_none() && state.error.is_none() {
+                state.response = Some(LocalResponseHead {
+                    status: meta.status,
+                    headers: meta.headers,
+                });
+                notify = true;
+            }
+        }
+        if notify {
+            self.headers_notify.notify_waiters();
+        }
+    }
+
+    fn push_body_chunk(&self, payload: Bytes) -> bool {
+        if self.terminal.load(Ordering::Acquire) {
+            return false;
+        }
+        self.body_tx
+            .try_send(LocalBodyEvent::Chunk(payload))
+            .is_ok()
+    }
+
+    fn finish(&self) {
+        if self.terminal.swap(true, Ordering::AcqRel) {
+            return;
+        }
+        let mut notify = false;
+        {
+            let mut state = self.wait_state.lock();
+            if state.response.is_none() && state.error.is_none() {
+                state.error = Some("stream ended before response headers".to_string());
+                notify = true;
+            }
+        }
+        if notify {
+            self.headers_notify.notify_waiters();
+        }
+        let _ = self.body_tx.try_send(LocalBodyEvent::End);
+    }
+
+    fn fail(&self, error: impl Into<String>) {
+        if self.terminal.swap(true, Ordering::AcqRel) {
+            return;
+        }
+
+        let error = error.into();
+        let mut notify = false;
+        {
+            let mut state = self.wait_state.lock();
+            if state.response.is_none() && state.error.is_none() {
+                state.error = Some(error.clone());
+                notify = true;
+            }
+        }
+        if notify {
+            self.headers_notify.notify_waiters();
+        }
+        let _ = self.body_tx.try_send(LocalBodyEvent::Error(error));
+    }
+}
+
+pub struct HubRouter {
+    proxy_conns: RwLock<HashMap<String, Vec<Arc<ProxyConn>>>>,
+    proxy_conns_by_id: DashMap<u64, Arc<ProxyConn>>,
+    local_streams: DashMap<u64, Arc<LocalStream>>,
+    proxy_to_local: DashMap<(u64, u32), u64>,
+    next_conn_id: AtomicU64,
+    next_local_stream_id: AtomicU64,
+    control_plane: ControlPlaneClient,
+}
+
+impl HubRouter {
+    pub fn new(control_plane: ControlPlaneClient) -> Arc<Self> {
+        Arc::new(Self {
+            proxy_conns: RwLock::new(HashMap::new()),
+            proxy_conns_by_id: DashMap::new(),
+            local_streams: DashMap::new(),
+            proxy_to_local: DashMap::new(),
+            next_conn_id: AtomicU64::new(1),
+            next_local_stream_id: AtomicU64::new(1),
+            control_plane,
+        })
+    }
+
+    pub fn alloc_conn_id(&self) -> u64 {
+        self.next_conn_id.fetch_add(1, Ordering::Relaxed)
+    }
+
+    pub fn register_proxy(&self, conn: Arc<ProxyConn>) {
+        let node_id = conn.node_id.clone();
+        let node_name = conn.node_name.clone();
+        let conn_id = conn.id;
+        self.proxy_conns_by_id.insert(conn_id, conn.clone());
+
+        let pool_size = {
+            let mut map = self.proxy_conns.write();
+            map.entry(node_id.clone()).or_default().push(conn);
+            map.get(&node_id).map(|v| v.len()).unwrap_or(0)
+        };
+
+        info!(
+            node_id = %node_id,
+            node_name = %node_name,
+            conn_id = conn_id,
+            pool_size = pool_size,
+            "proxy connected"
+        );
+
+        self.notify_node_status(node_id, true, pool_size);
+    }
+
+    pub fn unregister_proxy(&self, conn_id: u64, node_id: &str) {
+        self.proxy_conns_by_id.remove(&conn_id);
+
+        let pool_size = {
+            let mut map = self.proxy_conns.write();
+            if let Some(conns) = map.get_mut(node_id) {
+                conns.retain(|c| c.id != conn_id);
+                if conns.is_empty() {
+                    map.remove(node_id);
+                }
+            }
+            map.get(node_id).map(|v| v.len()).unwrap_or(0)
+        };
+
+        info!(
+            node_id = %node_id,
+            conn_id = conn_id,
+            remaining = pool_size,
+            "proxy disconnected"
+        );
+
+        self.cancel_streams_for_proxy(conn_id);
+        self.notify_node_status(node_id.to_string(), pool_size > 0, pool_size);
+    }
+
+    fn notify_node_status(&self, node_id: String, connected: bool, conn_count: usize) {
+        let control_plane = self.control_plane.clone();
+        tokio::spawn(async move {
+            if let Err(error) = control_plane
+                .push_node_status(&node_id, connected, conn_count)
+                .await
+            {
+                warn!(
+                    node_id = %node_id,
+                    connected = connected,
+                    conn_count = conn_count,
+                    error = %error,
+                    "failed to push node status to app control plane"
+                );
+            }
+        });
+    }
+
+    fn get_proxy_conn(&self, node_id: &str) -> Option<Arc<ProxyConn>> {
+        let map = self.proxy_conns.read();
+        let conns = map.get(node_id)?;
+        conns
+            .iter()
+            .filter(|c| c.is_available())
+            .min_by_key(|c| c.stream_count.load(Ordering::Relaxed))
+            .cloned()
+    }
+
+    pub fn open_local_stream(
+        &self,
+        node_id: &str,
+        meta: &protocol::RequestMeta,
+    ) -> Result<Arc<LocalStream>, String> {
+        let proxy_conn = self
+            .get_proxy_conn(node_id)
+            .ok_or_else(|| format!("no proxy connection for node {node_id}"))?;
+        let proxy_stream_id = proxy_conn
+            .alloc_stream_id()
+            .ok_or_else(|| format!("stream limit reached for node {node_id}"))?;
+
+        // Encode frames before registering the stream so that encoding failures
+        // (practically impossible but theoretically possible) don't leak a stream
+        // slot or orphan map entries.
+        let meta_json = match serde_json::to_vec(meta) {
+            Ok(json) => json,
+            Err(e) => {
+                proxy_conn.release_stream();
+                return Err(format!("failed to encode request metadata: {e}"));
+            }
+        };
+        let (meta_payload, meta_flags) = match protocol::compress_payload(&meta_json) {
+            Ok(result) => result,
+            Err(e) => {
+                proxy_conn.release_stream();
+                return Err(format!("failed to compress request metadata: {e}"));
+            }
+        };
+        let header_frame = protocol::encode_frame(
+            proxy_stream_id,
+            protocol::REQUEST_HEADERS,
+            meta_flags,
+            &meta_payload,
+        );
+
+        // Frames encoded successfully -- now register the stream.
+        let local_stream_id = self.next_local_stream_id.fetch_add(1, Ordering::Relaxed);
+        let local_stream = Arc::new(LocalStream::new(
+            local_stream_id,
+            proxy_conn.id,
+            proxy_stream_id,
+        ));
+        self.local_streams
+            .insert(local_stream_id, local_stream.clone());
+        self.proxy_to_local
+            .insert((proxy_conn.id, proxy_stream_id), local_stream_id);
+
+        match proxy_conn.send(Message::Binary(header_frame.into())) {
+            SendStatus::Queued => Ok(local_stream),
+            SendStatus::Closed | SendStatus::Congested => {
+                self.cleanup_local_stream(local_stream_id);
+                proxy_conn.release_stream();
+                Err("proxy connection congested".to_string())
+            }
+        }
+    }
+
+    pub fn push_local_request_body(
+        &self,
+        local_stream_id: u64,
+        payload: Bytes,
+        end_stream: bool,
+    ) -> Result<(), String> {
+        let stream = self
+            .local_streams
+            .get(&local_stream_id)
+            .map(|entry| entry.value().clone())
+            .ok_or_else(|| "local stream not found".to_string())?;
+        let proxy_conn = self
+            .proxy_conns_by_id
+            .get(&stream.proxy_conn_id)
+            .map(|entry| entry.value().clone())
+            .ok_or_else(|| "proxy connection unavailable".to_string())?;
+
+        let total_chunks = payload.len().div_ceil(MAX_REQUEST_BODY_FRAME_SIZE);
+        if total_chunks == 0 {
+            if end_stream {
+                self.send_request_body_frame(&proxy_conn, stream.proxy_stream_id, &[], true)?;
+            }
+        } else {
+            for (index, chunk) in payload.chunks(MAX_REQUEST_BODY_FRAME_SIZE).enumerate() {
+                let is_last_chunk = index + 1 == total_chunks;
+                self.send_request_body_frame(
+                    &proxy_conn,
+                    stream.proxy_stream_id,
+                    chunk,
+                    end_stream && is_last_chunk,
+                )?;
+            }
+        }
+
+        Ok(())
+    }
+
+    fn send_request_body_frame(
+        &self,
+        proxy_conn: &Arc<ProxyConn>,
+        proxy_stream_id: u32,
+        payload: &[u8],
+        end_stream: bool,
+    ) -> Result<(), String> {
+        let (body_payload, body_flags) = protocol::compress_payload(payload)
+            .map_err(|e| format!("failed to compress request body: {e}"))?;
+        let body_frame = protocol::encode_frame(
+            proxy_stream_id,
+            protocol::REQUEST_BODY,
+            body_flags
+                | if end_stream {
+                    protocol::FLAG_END_STREAM
+                } else {
+                    0
+                },
+            &body_payload,
+        );
+        match proxy_conn.send(Message::Binary(body_frame.into())) {
+            SendStatus::Queued => Ok(()),
+            SendStatus::Closed | SendStatus::Congested => {
+                Err("proxy connection congested".to_string())
+            }
+        }
+    }
+
+    pub fn cancel_local_stream(&self, local_stream_id: u64, reason: &str) {
+        let Some((_, stream)) = self.local_streams.remove(&local_stream_id) else {
+            return;
+        };
+
+        self.proxy_to_local
+            .remove(&(stream.proxy_conn_id, stream.proxy_stream_id));
+        if let Some(pc) = self.proxy_conns_by_id.get(&stream.proxy_conn_id) {
+            pc.release_stream();
+            let frame = protocol::encode_stream_error(stream.proxy_stream_id, reason);
+            let _ = pc.send(Message::Binary(frame.into()));
+        }
+        stream.fail(reason.to_string());
+    }
+
+    fn cleanup_local_stream(&self, local_stream_id: u64) {
+        let Some((_, stream)) = self.local_streams.remove(&local_stream_id) else {
+            return;
+        };
+        self.proxy_to_local
+            .remove(&(stream.proxy_conn_id, stream.proxy_stream_id));
+    }
+
+    pub async fn handle_proxy_frame(&self, proxy_conn_id: u64, data: &mut [u8]) {
+        let header = match protocol::FrameHeader::parse(data) {
+            Some(h) => h,
+            None => return,
+        };
+        let expected_len = protocol::HEADER_SIZE + header.payload_len as usize;
+        if data.len() < expected_len {
+            return;
+        }
+
+        match header.msg_type {
+            protocol::RESPONSE_HEADERS => {
+                self.route_response_headers(proxy_conn_id, header, data);
+            }
+            protocol::RESPONSE_BODY => {
+                self.route_response_body(proxy_conn_id, header, data);
+            }
+            protocol::STREAM_END => {
+                self.finish_proxy_stream(proxy_conn_id, header.stream_id);
+            }
+            protocol::STREAM_ERROR => {
+                let message = protocol::decode_payload(data, &header)
+                    .ok()
+                    .and_then(|payload| String::from_utf8(payload).ok())
+                    .unwrap_or_else(|| "stream error".to_string());
+                self.fail_proxy_stream(proxy_conn_id, header.stream_id, message);
+            }
+            protocol::HEARTBEAT_DATA => {
+                self.handle_heartbeat(proxy_conn_id, header.stream_id, data, &header)
+                    .await;
+            }
+            protocol::PING => {
+                let payload = protocol::frame_payload_by_header(data, &header).unwrap_or(&[]);
+                let pong = protocol::encode_pong(payload);
+                if let Some(pc) = self.proxy_conns_by_id.get(&proxy_conn_id) {
+                    let _ = pc.send(Message::Binary(pong.into()));
+                }
+            }
+            protocol::PONG => {}
+            protocol::GOAWAY => {
+                warn!(
+                    proxy_conn_id = proxy_conn_id,
+                    "received GOAWAY from proxy connection"
+                );
+            }
+            _ => {
+                debug!(
+                    msg_type = header.msg_type,
+                    proxy_conn_id = proxy_conn_id,
+                    "unexpected frame type from proxy"
+                );
+            }
+        }
+    }
+
+    fn route_response_headers(
+        &self,
+        proxy_conn_id: u64,
+        header: protocol::FrameHeader,
+        data: &[u8],
+    ) {
+        let Some(local_id) = self.lookup_local_stream(proxy_conn_id, header.stream_id) else {
+            return;
+        };
+        let Ok(payload) = protocol::decode_payload(data, &header) else {
+            self.fail_proxy_stream(
+                proxy_conn_id,
+                header.stream_id,
+                "failed to decode response headers",
+            );
+            return;
+        };
+        let Ok(meta) = serde_json::from_slice::<protocol::ResponseMeta>(&payload) else {
+            self.fail_proxy_stream(
+                proxy_conn_id,
+                header.stream_id,
+                "invalid response headers payload",
+            );
+            return;
+        };
+        if let Some(entry) = self.local_streams.get(&local_id) {
+            entry.value().set_response_headers(meta);
+        }
+    }
+
+    fn route_response_body(&self, proxy_conn_id: u64, header: protocol::FrameHeader, data: &[u8]) {
+        let Some(local_id) = self.lookup_local_stream(proxy_conn_id, header.stream_id) else {
+            return;
+        };
+        let Ok(payload) = protocol::decode_payload(data, &header) else {
+            self.fail_proxy_stream(
+                proxy_conn_id,
+                header.stream_id,
+                "failed to decode response body",
+            );
+            return;
+        };
+
+        let stream = match self.local_streams.get(&local_id) {
+            Some(entry) => entry.value().clone(),
+            None => return,
+        };
+
+        if !stream.push_body_chunk(Bytes::from(payload)) {
+            self.cancel_local_stream(local_id, "local relay response congested");
+        }
+    }
+
+    fn handle_stream_cleanup(
+        &self,
+        proxy_conn_id: u64,
+        proxy_stream_id: u32,
+    ) -> Option<Arc<LocalStream>> {
+        let local_id = self
+            .proxy_to_local
+            .remove(&(proxy_conn_id, proxy_stream_id))
+            .map(|(_, local_id)| local_id)?;
+
+        let stream = self
+            .local_streams
+            .remove(&local_id)
+            .map(|(_, stream)| stream)?;
+        if let Some(pc) = self.proxy_conns_by_id.get(&proxy_conn_id) {
+            pc.release_stream();
+        }
+        Some(stream)
+    }
+
+    fn finish_proxy_stream(&self, proxy_conn_id: u64, proxy_stream_id: u32) {
+        if let Some(stream) = self.handle_stream_cleanup(proxy_conn_id, proxy_stream_id) {
+            stream.finish();
+        }
+    }
+
+    fn fail_proxy_stream(
+        &self,
+        proxy_conn_id: u64,
+        proxy_stream_id: u32,
+        error: impl Into<String>,
+    ) {
+        if let Some(stream) = self.handle_stream_cleanup(proxy_conn_id, proxy_stream_id) {
+            stream.fail(error.into());
+        }
+    }
+
+    fn lookup_local_stream(&self, proxy_conn_id: u64, proxy_stream_id: u32) -> Option<u64> {
+        self.proxy_to_local
+            .get(&(proxy_conn_id, proxy_stream_id))
+            .map(|entry| *entry.value())
+    }
+
+    async fn handle_heartbeat(
+        &self,
+        proxy_conn_id: u64,
+        stream_id: u32,
+        data: &[u8],
+        header: &protocol::FrameHeader,
+    ) {
+        let payload = match protocol::decode_payload(data, header) {
+            Ok(payload) => payload,
+            Err(error) => {
+                warn!(proxy_conn_id = proxy_conn_id, error = %error, "failed to decode heartbeat payload");
+                return;
+            }
+        };
+        let ack_payload = match self.control_plane.heartbeat_ack(&payload).await {
+            Ok(payload) => payload,
+            Err(error) => {
+                warn!(proxy_conn_id = proxy_conn_id, error = %error, "control-plane heartbeat callback failed");
+                b"{}".to_vec()
+            }
+        };
+        if let Some(pc) = self.proxy_conns_by_id.get(&proxy_conn_id) {
+            let frame = protocol::encode_frame(stream_id, protocol::HEARTBEAT_ACK, 0, &ack_payload);
+            let _ = pc.send(Message::Binary(frame.into()));
+        }
+    }
+
+    fn cancel_streams_for_proxy(&self, proxy_conn_id: u64) {
+        let mut cancelled = 0usize;
+        self.proxy_to_local.retain(|key, local_id| {
+            if key.0 != proxy_conn_id {
+                return true;
+            }
+            if let Some((_, stream)) = self.local_streams.remove(local_id) {
+                stream.fail("proxy disconnected".to_string());
+            }
+            cancelled += 1;
+            false
+        });
+
+        if cancelled > 0 {
+            warn!(
+                proxy_conn_id = proxy_conn_id,
+                streams_cancelled = cancelled,
+                "cancelled in-flight streams due to proxy disconnect"
+            );
+        }
+    }
+
+    pub fn stats(&self) -> HubStats {
+        let proxy_conns = self.proxy_conns.read();
+        let total_proxy = proxy_conns.values().map(|v| v.len()).sum();
+        let nodes = proxy_conns.len();
+        drop(proxy_conns);
+
+        HubStats {
+            proxy_connections: total_proxy,
+            nodes,
+            active_streams: self.local_streams.len(),
+        }
+    }
+}
+
+#[derive(serde::Serialize)]
+pub struct HubStats {
+    pub proxy_connections: usize,
+    pub nodes: usize,
+    pub active_streams: usize,
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    fn build_meta() -> protocol::RequestMeta {
+        protocol::RequestMeta {
+            method: "GET".to_string(),
+            url: "https://example.com".to_string(),
+            headers: HashMap::new(),
+            timeout: 30,
+        }
+    }
+
+    #[tokio::test]
+    async fn cancel_local_stream_notifies_proxy() {
+        let hub = HubRouter::new(ControlPlaneClient::disabled());
+
+        let (proxy_tx, mut proxy_rx) = mpsc::channel(8);
+        let (proxy_close_tx, _) = watch::channel(false);
+        let proxy = Arc::new(ProxyConn::new(
+            100,
+            "node-1".to_string(),
+            "Node 1".to_string(),
+            proxy_tx,
+            proxy_close_tx,
+            16,
+        ));
+        hub.register_proxy(proxy);
+
+        let stream = hub
+            .open_local_stream("node-1", &build_meta())
+            .expect("open local stream");
+        let _ = proxy_rx.try_recv().expect("headers frame");
+        hub.push_local_request_body(stream.id, Bytes::new(), true)
+            .expect("finish empty body");
+        let _ = proxy_rx.try_recv().expect("body frame");
+
+        hub.cancel_local_stream(stream.id, "client dropped");
+
+        let cancelled = proxy_rx.try_recv().expect("cancel frame");
+        let cancelled_data = match cancelled {
+            Message::Binary(data) => data.to_vec(),
+            other => panic!("unexpected message: {other:?}"),
+        };
+        let header = protocol::FrameHeader::parse(&cancelled_data).expect("cancel frame header");
+        assert_eq!(header.msg_type, protocol::STREAM_ERROR);
+    }
+
+    #[tokio::test]
+    async fn push_local_request_body_splits_large_payload_and_marks_end() {
+        let hub = HubRouter::new(ControlPlaneClient::disabled());
+
+        let (proxy_tx, mut proxy_rx) = mpsc::channel(8);
+        let (proxy_close_tx, _) = watch::channel(false);
+        let proxy = Arc::new(ProxyConn::new(
+            200,
+            "node-2".to_string(),
+            "Node 2".to_string(),
+            proxy_tx,
+            proxy_close_tx,
+            16,
+        ));
+        hub.register_proxy(proxy);
+
+        let stream = hub
+            .open_local_stream("node-2", &build_meta())
+            .expect("open local stream");
+        let _ = proxy_rx.try_recv().expect("headers frame");
+
+        let payload = Bytes::from(vec![b'x'; MAX_REQUEST_BODY_FRAME_SIZE + 17]);
+        hub.push_local_request_body(stream.id, payload, true)
+            .expect("push request body");
+
+        let first = match proxy_rx.try_recv().expect("first body frame") {
+            Message::Binary(data) => data.to_vec(),
+            other => panic!("unexpected message: {other:?}"),
+        };
+        let first_header = protocol::FrameHeader::parse(&first).expect("first body header");
+        assert_eq!(first_header.msg_type, protocol::REQUEST_BODY);
+        assert_eq!(first_header.flags & protocol::FLAG_END_STREAM, 0);
+
+        let second = match proxy_rx.try_recv().expect("second body frame") {
+            Message::Binary(data) => data.to_vec(),
+            other => panic!("unexpected message: {other:?}"),
+        };
+        let second_header = protocol::FrameHeader::parse(&second).expect("second body header");
+        assert_eq!(second_header.msg_type, protocol::REQUEST_BODY);
+        assert_ne!(second_header.flags & protocol::FLAG_END_STREAM, 0);
+    }
+}
@@ -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")))
+}
@@ -0,0 +1,167 @@
+mod control_plane;
+mod hub;
+mod local_relay;
+mod protocol;
+mod proxy_conn;
+
+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, post};
+use axum::Router;
+use clap::Parser;
+use tracing::{info, warn};
+
+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")]
+struct Args {
+    /// Bind address
+    #[arg(long, default_value = "0.0.0.0:8085", env = "TUNNEL_HUB_BIND")]
+    bind: String,
+
+    /// 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,
+
+    /// Ping interval in seconds (for both sides)
+    #[arg(long, default_value_t = 15, env = "TUNNEL_HUB_PING_INTERVAL")]
+    ping_interval: u64,
+
+    /// 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)]
+pub struct AppState {
+    pub hub: std::sync::Arc<HubRouter>,
+    pub proxy_conn_cfg: ConnConfig,
+    pub max_streams: usize,
+}
+
+#[tokio::main]
+async fn main() -> Result<(), Box<dyn std::error::Error>> {
+    // Initialize tracing
+    tracing_subscriber::fmt()
+        .with_env_filter(
+            tracing_subscriber::EnvFilter::try_from_default_env()
+                .unwrap_or_else(|_| "aether_hub=info".into()),
+        )
+        .init();
+
+    let args = Args::parse();
+
+    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_conn_cfg: ConnConfig {
+            ping_interval,
+            idle_timeout: Duration::from_secs(args.proxy_idle_timeout),
+            outbound_queue_capacity,
+        },
+        max_streams: args.max_streams,
+    };
+
+    let app = Router::new()
+        .route("/health", get(health))
+        .route("/stats", get(stats))
+        .route("/proxy", get(ws_proxy))
+        .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.into_make_service_with_connect_info::<SocketAddr>(),
+    )
+    .await?;
+    Ok(())
+}
+
+// ---------------------------------------------------------------------------
+// HTTP endpoints
+// ---------------------------------------------------------------------------
+
+async fn health() -> impl IntoResponse {
+    Json(serde_json::json!({"status": "ok"}))
+}
+
+async fn stats(State(state): State<AppState>) -> impl IntoResponse {
+    Json(state.hub.stats())
+}
+
+// ---------------------------------------------------------------------------
+// WebSocket endpoints
+// ---------------------------------------------------------------------------
+
+async fn ws_proxy(
+    ws: WebSocketUpgrade,
+    State(state): State<AppState>,
+    headers: axum::http::HeaderMap,
+) -> impl IntoResponse {
+    let node_id = headers
+        .get("x-node-id")
+        .and_then(|v| v.to_str().ok())
+        .unwrap_or("")
+        .trim()
+        .to_string();
+
+    let node_name = headers
+        .get("x-node-name")
+        .and_then(|v| v.to_str().ok())
+        .unwrap_or(&node_id)
+        .trim()
+        .to_string();
+
+    let max_streams: usize = headers
+        .get("x-tunnel-max-streams")
+        .and_then(|v| v.to_str().ok())
+        .and_then(|v| v.parse().ok())
+        .unwrap_or(state.max_streams)
+        .clamp(64, 2048);
+
+    if node_id.is_empty() {
+        warn!("proxy connection rejected: missing X-Node-ID header");
+        return axum::http::StatusCode::BAD_REQUEST.into_response();
+    }
+
+    ws.max_frame_size(64 * 1024 * 1024)
+        .on_upgrade(move |socket| {
+            proxy_conn::handle_proxy_connection(
+                socket,
+                state.hub,
+                node_id,
+                node_name,
+                max_streams,
+                state.proxy_conn_cfg,
+            )
+        })
+        .into_response()
+}
@@ -0,0 +1,176 @@
+/// Tunnel binary frame protocol
+///
+/// Frame format (10-byte header + payload):
+/// | stream_id (4B) | msg_type (1B) | flags (1B) | payload_len (4B) | payload (NB) |
+use std::io::Read;
+
+use flate2::read::GzDecoder;
+use flate2::write::GzEncoder;
+use flate2::Compression;
+
+pub const HEADER_SIZE: usize = 10;
+
+// Message types
+pub const REQUEST_HEADERS: u8 = 0x01;
+pub const REQUEST_BODY: u8 = 0x02;
+pub const RESPONSE_HEADERS: u8 = 0x03;
+pub const RESPONSE_BODY: u8 = 0x04;
+pub const STREAM_END: u8 = 0x05;
+pub const STREAM_ERROR: u8 = 0x06;
+pub const PING: u8 = 0x10;
+pub const PONG: u8 = 0x11;
+pub const GOAWAY: u8 = 0x12;
+pub const HEARTBEAT_DATA: u8 = 0x13;
+pub const HEARTBEAT_ACK: u8 = 0x14;
+// Flags
+pub const FLAG_END_STREAM: u8 = 0x01;
+pub const FLAG_GZIP_COMPRESSED: u8 = 0x02;
+
+#[derive(Debug, Clone, Copy)]
+pub struct FrameHeader {
+    pub stream_id: u32,
+    pub msg_type: u8,
+    pub flags: u8,
+    pub payload_len: u32,
+}
+
+impl FrameHeader {
+    /// Parse frame header from raw bytes (must be >= HEADER_SIZE)
+    #[inline]
+    pub fn parse(data: &[u8]) -> Option<Self> {
+        if data.len() < HEADER_SIZE {
+            return None;
+        }
+        Some(Self {
+            stream_id: u32::from_be_bytes([data[0], data[1], data[2], data[3]]),
+            msg_type: data[4],
+            flags: data[5],
+            payload_len: u32::from_be_bytes([data[6], data[7], data[8], data[9]]),
+        })
+    }
+}
+
+#[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, 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> {
+    encode_frame(stream_id, STREAM_ERROR, 0, msg.as_bytes())
+}
+
+/// Encode a PING frame (stream_id=0)
+pub fn encode_ping() -> Vec<u8> {
+    encode_frame(0, PING, 0, &[])
+}
+
+/// Encode a PONG frame (stream_id=0, echo payload)
+pub fn encode_pong(payload: &[u8]) -> Vec<u8> {
+    encode_frame(0, PONG, 0, payload)
+}
+
+/// Encode a GOAWAY frame (stream_id=0)
+pub fn encode_goaway() -> Vec<u8> {
+    encode_frame(0, GOAWAY, 0, &[])
+}
+
+#[inline]
+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 {
+        return None;
+    }
+    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,
+) -> Result<(Vec<u8>, u8), std::io::Error> {
+    if !prefer_gzip {
+        return Ok((payload.to_vec(), 0));
+    }
+    let mut encoder = GzEncoder::new(Vec::new(), Compression::default());
+    std::io::Write::write_all(&mut encoder, payload)?;
+    let compressed = encoder.finish()?;
+    if compressed.len() < payload.len() {
+        Ok((compressed, FLAG_GZIP_COMPRESSED))
+    } else {
+        Ok((payload.to_vec(), 0))
+    }
+}
@@ -0,0 +1,156 @@
+/// Proxy-side WebSocket connection handler
+///
+/// Handles the lifecycle of a single aether-proxy connection:
+/// accept -> authenticate (headers) -> read loop -> cleanup
+use std::sync::Arc;
+use std::time::Duration;
+
+use axum::extract::ws::{Message, WebSocket};
+use futures_util::{SinkExt, StreamExt};
+use tokio::sync::{mpsc, watch};
+use tracing::{debug, info, warn};
+
+use crate::hub::{ConnConfig, HubRouter, ProxyConn, SendStatus};
+use crate::protocol;
+
+/// Maximum single frame size: 64 MB
+const MAX_FRAME_SIZE: usize = 64 * 1024 * 1024;
+
+pub async fn handle_proxy_connection(
+    ws: WebSocket,
+    hub: Arc<HubRouter>,
+    node_id: String,
+    node_name: String,
+    max_streams: usize,
+    cfg: ConnConfig,
+) {
+    let conn_id = hub.alloc_conn_id();
+    let (mut ws_tx, ws_rx) = ws.split();
+
+    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());
+
+    let writer = tokio::spawn(async move {
+        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;
+    });
+
+    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 !matches!(
+                ping_conn.send(Message::Binary(ping.into())),
+                SendStatus::Queued
+            ) {
+                break;
+            }
+        }
+    });
+
+    let reader_hub = hub.clone();
+    let reader_conn = conn.clone();
+    let reader = tokio::spawn(async move {
+        run_proxy_reader(ws_rx, reader_hub, reader_conn, cfg.idle_timeout).await;
+    });
+
+    let _ = reader.await;
+    ping_task.abort();
+    conn.request_close();
+    hub.unregister_proxy(conn_id, &node_id);
+    drop(conn);
+    tokio::time::sleep(Duration::from_millis(100)).await;
+    writer.abort();
+    let _ = writer.await;
+}
+
+async fn run_proxy_reader(
+    mut ws_rx: futures_util::stream::SplitStream<WebSocket>,
+    hub: Arc<HubRouter>,
+    conn: Arc<ProxyConn>,
+    idle_timeout: Duration,
+) {
+    let idle_enabled = !idle_timeout.is_zero();
+    let mut oversized_count = 0u32;
+    loop {
+        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 {
+            Some(Ok(Message::Binary(data))) => {
+                let mut data = data.to_vec();
+                if data.len() > MAX_FRAME_SIZE {
+                    oversized_count += 1;
+                    warn!(
+                        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");
+                        conn.request_close();
+                        break;
+                    }
+                    continue;
+                }
+                oversized_count = 0;
+
+                if data.len() < protocol::HEADER_SIZE {
+                    debug!(conn_id = conn.id, "frame too small, skipping");
+                    continue;
+                }
+
+                hub.handle_proxy_frame(conn.id, &mut data).await;
+            }
+            Some(Ok(Message::Close(_))) | None => {
+                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");
+                break;
+            }
+            _ => {}
+        }
+    }
+}
@@ -4,28 +4,5 @@ AETHER_PROXY_AETHER_URL=https://aether.example.com
 # Management Token (ae_xxx, must belong to an ADMIN user)
 AETHER_PROXY_MANAGEMENT_TOKEN=ae_xxxxx

-# HMAC key (must match Aether's PROXY_HMAC_KEY)
-AETHER_PROXY_HMAC_KEY=
-
-# Proxy listen port
-AETHER_PROXY_LISTEN_PORT=18080
-
-# Public IP (auto-detected if omitted)
-# AETHER_PROXY_PUBLIC_IP=203.0.113.42
-
 # Node identification
 AETHER_PROXY_NODE_NAME=proxy-01
-# AETHER_PROXY_NODE_REGION=ap-northeast-1
-
-# Heartbeat interval in seconds
-AETHER_PROXY_HEARTBEAT_INTERVAL=30
-
-# Allowed destination ports (comma-separated)
-AETHER_PROXY_ALLOWED_PORTS=80,443,8080,8443
-
-# HMAC timestamp tolerance in seconds
-AETHER_PROXY_TIMESTAMP_TOLERANCE=300
-
-# Logging
-AETHER_PROXY_LOG_LEVEL=info
-AETHER_PROXY_LOG_JSON=false
@@ -10,9 +10,10 @@ checksum = "320119579fcad9c21884f5c4861d16174d0e06250625266f50fe6898340abefa"

 [[package]]
 name = "aether-proxy"
-version = "0.1.0"
+version = "0.2.4"
 dependencies = [
 "anyhow",
+ "arc-swap",
 "base64",
 "bytes",
 "clap",
@@ -20,30 +21,29 @@ dependencies = [
 "flate2",
 "futures-util",
 "hex",
- "hmac",
 "http-body-util",
 "hyper",
 "hyper-util",
 "libc",
 "ratatui",
- "rcgen",
 "reqwest",
 "rustls",
- "rustls-pemfile",
- "rustls-pki-types",
 "serde",
 "serde_json",
 "sha2",
- "subtle",
+ "socket2 0.5.10",
 "sysinfo",
 "tar",
 "thiserror 2.0.18",
 "tokio",
 "tokio-rustls",
+ "tokio-tungstenite",
 "toml",
+ "tower-service",
 "tracing",
 "tracing-subscriber",
 "url",
+ "webpki-roots 0.26.11",
 ]

 [[package]]
@@ -117,6 +117,15 @@ version = "1.0.101"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "5f0e0fee31ef5ed1ba1316088939cea399010ed7731dba877ed44aeb407a75ea"

+[[package]]
+name = "arc-swap"
+version = "1.8.2"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "f9f3647c145568cec02c42054e07bdf9a5a698e15b466fb2341bfc393cd24aa5"
+dependencies = [
+ "rustversion",
+]
+
 [[package]]
 name = "atomic"
 version = "0.6.1"
@@ -214,6 +223,12 @@ version = "1.25.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "c8efb64bd706a16a1bdde310ae86b351e4d21550d98d056f22f8a7f7a2183fec"

+[[package]]
+name = "byteorder"
+version = "1.5.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "1fd0f2584146f6f2ef48085050886acf353beff7305ebd1ae69500e27c67f64b"
+
 [[package]]
 name = "bytes"
 version = "1.11.1"
@@ -477,6 +492,12 @@ dependencies = [
 "syn 2.0.114",
 ]

+[[package]]
+name = "data-encoding"
+version = "2.10.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "d7a1e2f27636f116493b8b860f5546edb47c8d8f8ea73e1d2a20be88e28d1fea"
+
 [[package]]
 name = "deltae"
 version = "0.3.2"
@@ -522,7 +543,6 @@ checksum = "9ed9a281f7bc9b7576e61468ba615a66a5c8cfdff42420a70aa82701a3b1e292"
 dependencies = [
 "block-buffer",
 "crypto-common",
- "subtle",
 ]

 [[package]]
@@ -767,6 +787,25 @@ dependencies = [
 "wasm-bindgen",
 ]

+[[package]]
+name = "h2"
+version = "0.4.13"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "2f44da3a8150a6703ed5d34e164b875fd14c2cdab9af1252a9a1020bde2bdc54"
+dependencies = [
+ "atomic-waker",
+ "bytes",
+ "fnv",
+ "futures-core",
+ "futures-sink",
+ "http",
+ "indexmap",
+ "slab",
+ "tokio",
+ "tokio-util",
+ "tracing",
+]
+
 [[package]]
 name = "hashbrown"
 version = "0.16.1"
@@ -790,15 +829,6 @@ version = "0.4.3"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "7f24254aa9a54b5c858eaee2f5bccdb46aaf0e486a595ed5fd8f86ba55232a70"

-[[package]]
-name = "hmac"
-version = "0.12.1"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "6c49c37c09c17a53d937dfbb742eb3a961d65a994e6bcdcf37e7399d0cc8ab5e"
-dependencies = [
- "digest",
-]
-
 [[package]]
 name = "http"
 version = "1.4.0"
@@ -838,12 +868,6 @@ version = "1.10.1"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "6dbf3de79e51f3d586ab4cb9d5c3e2c14aa28ed23d180cf89b4df0454a69cc87"

-[[package]]
-name = "httpdate"
-version = "1.0.3"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "df3b46402a9d5adb4c86a0cf463f42e19994e3ee891101b1841f30a545cb49a9"
-
 [[package]]
 name = "hyper"
 version = "1.8.1"
@@ -854,10 +878,10 @@ dependencies = [
 "bytes",
 "futures-channel",
 "futures-core",
+ "h2",
 "http",
 "http-body",
 "httparse",
- "httpdate",
 "itoa",
 "pin-project-lite",
 "pin-utils",
@@ -880,7 +904,7 @@ dependencies = [
 "tokio",
 "tokio-rustls",
 "tower-service",
- "webpki-roots",
+ "webpki-roots 1.0.6",
 ]

 [[package]]
@@ -900,7 +924,7 @@ dependencies = [
 "libc",
 "percent-encoding",
 "pin-project-lite",
- "socket2",
+ "socket2 0.6.2",
 "tokio",
 "tower-service",
 "tracing",
@@ -1394,16 +1418,6 @@ dependencies = [
 "windows-link",
 ]

-[[package]]
-name = "pem"
-version = "3.0.6"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "1d30c53c26bc5b31a98cd02d20f25a7c8567146caf63ed593a9d87b2775291be"
-dependencies = [
- "base64",
- "serde_core",
-]
-
 [[package]]
 name = "percent-encoding"
 version = "2.3.2"
@@ -1569,7 +1583,7 @@ dependencies = [
 "quinn-udp",
 "rustc-hash",
 "rustls",
- "socket2",
+ "socket2 0.6.2",
 "thiserror 2.0.18",
 "tokio",
 "tracing",
@@ -1606,7 +1620,7 @@ dependencies = [
 "cfg_aliases",
 "libc",
 "once_cell",
- "socket2",
+ "socket2 0.6.2",
 "tracing",
 "windows-sys 0.60.2",
 ]
@@ -1632,6 +1646,8 @@ version = "0.8.5"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "34af8d1a0e25924bc5b7c43c079c942339d8f0a8b57c39049bef581b46327404"
 dependencies = [
+ "libc",
+ "rand_chacha 0.3.1",
 "rand_core 0.6.4",
 ]

@@ -1641,10 +1657,20 @@ version = "0.9.2"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "6db2770f06117d490610c7488547d543617b21bfa07796d7a12f6f1bd53850d1"
 dependencies = [
- "rand_chacha",
+ "rand_chacha 0.9.0",
 "rand_core 0.9.5",
 ]

+[[package]]
+name = "rand_chacha"
+version = "0.3.1"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "e6c10a63a0fa32252be49d21e7709d4d4baf8d231c2dbce1eaa8141b9b127d88"
+dependencies = [
+ "ppv-lite86",
+ "rand_core 0.6.4",
+]
+
 [[package]]
 name = "rand_chacha"
 version = "0.9.0"
@@ -1660,6 +1686,9 @@ name = "rand_core"
 version = "0.6.4"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "ec0be4795e2f6a28069bec0b5ff3e2ac9bafc99e6a9a7dc3547996c5c816922c"
+dependencies = [
+ "getrandom 0.2.17",
+]

 [[package]]
 name = "rand_core"
@@ -1775,19 +1804,6 @@ dependencies = [
 "crossbeam-utils",
 ]

-[[package]]
-name = "rcgen"
-version = "0.13.2"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "75e669e5202259b5314d1ea5397316ad400819437857b90861765f24c4cf80a2"
-dependencies = [
- "pem",
- "ring",
- "rustls-pki-types",
- "time",
- "yasna",
-]
-
 [[package]]
 name = "redox_syscall"
 version = "0.5.18"
@@ -1845,6 +1861,7 @@ dependencies = [
 "bytes",
 "futures-core",
 "futures-util",
+ "h2",
 "http",
 "http-body",
 "http-body-util",
@@ -1873,7 +1890,7 @@ dependencies = [
 "wasm-bindgen-futures",
 "wasm-streams",
 "web-sys",
- "webpki-roots",
+ "webpki-roots 1.0.6",
 ]

 [[package]]
@@ -1947,15 +1964,6 @@ dependencies = [
 "zeroize",
 ]

-[[package]]
-name = "rustls-pemfile"
-version = "2.2.0"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "dce314e5fee3f39953d46bb63bb8a46d40c2f8fb7cc5a3b6cab2bde9721d6e50"
-dependencies = [
- "rustls-pki-types",
-]
-
 [[package]]
 name = "rustls-pki-types"
 version = "1.14.0"
@@ -2066,6 +2074,17 @@ dependencies = [
 "serde",
 ]

+[[package]]
+name = "sha1"
+version = "0.10.6"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "e3bf829a2d51ab4a5ddf1352d8470c140cadc8301b2ae1789db023f01cedd6ba"
+dependencies = [
+ "cfg-if",
+ "cpufeatures",
+ "digest",
+]
+
 [[package]]
 name = "sha2"
 version = "0.10.9"
@@ -2147,6 +2166,16 @@ version = "1.15.1"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "67b1b7a3b5fe4f1376887184045fcf45c69e92af734b7aaddc05fb777b6fbd03"

+[[package]]
+name = "socket2"
+version = "0.5.10"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "e22376abed350d73dd1cd119b57ffccad95b4e585a7cda43e286245ce23c0678"
+dependencies = [
+ "libc",
+ "windows-sys 0.52.0",
+]
+
 [[package]]
 name = "socket2"
 version = "0.6.2"
@@ -2439,7 +2468,7 @@ dependencies = [
 "parking_lot",
 "pin-project-lite",
 "signal-hook-registry",
- "socket2",
+ "socket2 0.6.2",
 "tokio-macros",
 "windows-sys 0.61.2",
 ]
@@ -2465,6 +2494,22 @@ dependencies = [
 "tokio",
 ]

+[[package]]
+name = "tokio-tungstenite"
+version = "0.24.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "edc5f74e248dc973e0dbb7b74c7e0d6fcc301c694ff50049504004ef4d0cdcd9"
+dependencies = [
+ "futures-util",
+ "log",
+ "rustls",
+ "rustls-pki-types",
+ "tokio",
+ "tokio-rustls",
+ "tungstenite",
+ "webpki-roots 0.26.11",
+]
+
 [[package]]
 name = "tokio-util"
 version = "0.7.18"
@@ -2644,6 +2689,26 @@ version = "0.2.5"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "e421abadd41a4225275504ea4d6566923418b7f05506fbc9c0fe86ba7396114b"

+[[package]]
+name = "tungstenite"
+version = "0.24.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "18e5b8366ee7a95b16d32197d0b2604b43a0be89dc5fac9f8e96ccafbaedda8a"
+dependencies = [
+ "byteorder",
+ "bytes",
+ "data-encoding",
+ "http",
+ "httparse",
+ "log",
+ "rand 0.8.5",
+ "rustls",
+ "rustls-pki-types",
+ "sha1",
+ "thiserror 1.0.69",
+ "utf-8",
+]
+
 [[package]]
 name = "typenum"
 version = "1.19.0"
@@ -2703,6 +2768,12 @@ dependencies = [
 "serde",
 ]

+[[package]]
+name = "utf-8"
+version = "0.7.6"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "09cc8ee72d2a9becf2f2febe0205bbed8fc6615b7cb429ad062dc7b7ddd036a9"
+
 [[package]]
 name = "utf8_iter"
 version = "1.0.4"
@@ -2864,6 +2935,15 @@ dependencies = [
 "wasm-bindgen",
 ]

+[[package]]
+name = "webpki-roots"
+version = "0.26.11"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "521bc38abb08001b01866da9f51eb7c5d647a19260e00054a8c7fd5f9e57f7a9"
+dependencies = [
+ "webpki-roots 1.0.6",
+]
+
 [[package]]
 name = "webpki-roots"
 version = "1.0.6"
@@ -3213,15 +3293,6 @@ dependencies = [
 "rustix 1.1.3",
 ]

-[[package]]
-name = "yasna"
-version = "0.5.2"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "e17bb3549cc1321ae1296b9cdc2698e2b6cb1992adfa19a8c72e5b7a738f44cd"
-dependencies = [
- "time",
-]
-
 [[package]]
 name = "yoke"
 version = "0.8.1"
@@ -1,19 +1,18 @@
 [package]
 name = "aether-proxy"
-version = "0.1.0"
+version = "0.2.5"
 edition = "2021"
-description = "Forward proxy for Aether with HMAC authentication"
+description = "Tunnel proxy for Aether"

 [dependencies]
 tokio = { version = "1", features = ["full"] }
-hyper = { version = "1", features = ["http1", "server"] }
-hyper-util = { version = "0.1", features = ["tokio", "http1", "server"] }
+reqwest = { version = "0.12", default-features = false, features = ["json", "rustls-tls", "stream", "http2"] }
+hyper = { version = "1", features = ["client", "http1", "http2"] }
+hyper-util = { version = "0.1", features = ["client", "client-legacy", "http1", "http2", "tokio"] }
 http-body-util = "0.1"
-reqwest = { version = "0.12", default-features = false, features = ["json", "rustls-tls", "stream"] }
+tokio-tungstenite = { version = "0.24", features = ["rustls-tls-webpki-roots"] }
+tokio-rustls = "0.26"
 futures-util = "0.3"
-hmac = "0.12"
-sha2 = "0.10"
-subtle = "2"
 base64 = "0.22"
 clap = { version = "4", features = ["derive", "env"] }
 tracing = "0.1"
@@ -22,14 +21,12 @@ serde = { version = "1", features = ["derive"] }
 serde_json = "1"
 thiserror = "2"
 bytes = "1"
+sha2 = "0.10"
 hex = "0.4"
 anyhow = "1"
+arc-swap = "1"
 toml = "0.8"
-tokio-rustls = "0.26"
 rustls = { version = "0.23", features = ["ring"] }
-rustls-pki-types = "1"
-rustls-pemfile = "2"
-rcgen = "0.13"
 ratatui = "0.30"
 crossterm = "0.28"
 url = "2"
@@ -37,6 +34,9 @@ sysinfo = "0.32"
 libc = "0.2"
 flate2 = "1"
 tar = "0.4"
+socket2 = { version = "0.5", features = ["all"] }
+tower-service = "0.3"
+webpki-roots = "0.26"

 [profile.release]
 lto = true
@@ -1,25 +1,10 @@
-FROM rust:1.83-slim AS builder
-
-WORKDIR /app
-
-RUN apt-get update && apt-get install -y pkg-config libssl-dev && rm -rf /var/lib/apt/lists/*
-
-COPY Cargo.toml Cargo.lock* ./
-# Create dummy main.rs for dependency caching
-RUN mkdir src && echo "fn main() {}" > src/main.rs
-RUN cargo build --release 2>/dev/null || true
-
-COPY src/ src/
-# Touch main.rs to force rebuild with real source
-RUN touch src/main.rs
-RUN cargo build --release
-
 FROM debian:bookworm-slim

-RUN apt-get update && apt-get install -y ca-certificates && rm -rf /var/lib/apt/lists/*
+ARG TARGETARCH

-COPY --from=builder /app/target/release/aether-proxy /usr/local/bin/aether-proxy
+RUN apt-get update && apt-get install -y --no-install-recommends ca-certificates \
+    && rm -rf /var/lib/apt/lists/*

-EXPOSE 18080
+COPY build/linux-${TARGETARCH}/aether-proxy /usr/local/bin/aether-proxy

 ENTRYPOINT ["aether-proxy"]
@@ -1,13 +1,30 @@
 # aether-proxy

-Aether 正向代理节点，部署在海外 VPS 上，为墙内的 Aether 实例中转 API 流量。
+Aether Tunnel 代理节点，部署在海外 VPS 上，通过 WebSocket 隧道为 Aether 实例中转 API 流量。
+
+Tunnel 模式下代理节点**无需对外监听端口**，仅需出站连接到 Aether 服务器。

 ## 安装

+### Docker Compose 部署
+
+```bash
+cp .env.example .env
+# 编辑 .env 填入 AETHER_PROXY_AETHER_URL 和 AETHER_PROXY_MANAGEMENT_TOKEN
+docker compose up -d
+```
+
 ### 下载预编译二进制

-在 [GitHub Releases](../../releases) 页面下载对应平台的预编译文件，无需安装 Rust 环境。
-
+<!-- DOWNLOAD_TABLE_START -->
+| Platform | Download |
+|----------|----------|
+| 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

@@ -29,7 +46,8 @@ sudo aether-proxy setup
 # 4. 彻底卸载
 sudo aether-proxy uninstall
 ```
-保存后配置写入 `aether-proxy.toml`，如果启用了 Install Service，将自动注册并启动 systemd 服务。
+
+完成向导后, 配置自动保存到 `aether-proxy.toml`，如果启用了 Install Service，将自动注册并启动 systemd 服务。

 ### 直接运行

@@ -49,29 +67,88 @@ sudo aether-proxy uninstall

 ### 参数一览

+#### 基础配置
+
 | 参数 | 环境变量 | 默认值 | 说明 |
 |------|----------|--------|------|
 | `--aether-url` | `AETHER_PROXY_AETHER_URL` | **必填** | Aether 服务器地址 |
 | `--management-token` | `AETHER_PROXY_MANAGEMENT_TOKEN` | **必填** | 管理员 Token（`ae_xxx` 格式） |
-| `--hmac-key` | `AETHER_PROXY_HMAC_KEY` | **必填** | HMAC 密钥，需与 Aether 端一致 |
-| `--listen-port` | `AETHER_PROXY_LISTEN_PORT` | `18080` | 监听端口 |
 | `--public-ip` | `AETHER_PROXY_PUBLIC_IP` | 自动检测 | 公网 IP |
 | `--node-name` | `AETHER_PROXY_NODE_NAME` | `proxy-01` | 节点名称标识 |
 | `--node-region` | `AETHER_PROXY_NODE_REGION` | 自动检测 | 地区标识 |
 | `--heartbeat-interval` | `AETHER_PROXY_HEARTBEAT_INTERVAL` | `30` | 心跳间隔（秒） |
 | `--allowed-ports` | `AETHER_PROXY_ALLOWED_PORTS` | `80,443,8080,8443` | 允许代理的目标端口 |
-| `--timestamp-tolerance` | `AETHER_PROXY_TIMESTAMP_TOLERANCE` | `300` | HMAC 时间戳容差（秒） |
+
+#### Tunnel 连接
+
+| 参数 | 环境变量 | 默认值 | 说明 |
+|------|----------|--------|------|
+| `--tunnel-connections` | `AETHER_PROXY_TUNNEL_CONNECTIONS` | `3` | 到 Aether 的连接池大小 |
+| `--tunnel-max-streams` | `AETHER_PROXY_TUNNEL_MAX_STREAMS` | 自动（硬件估算） | 单连接最大并发 stream 数 |
+| `--tunnel-connect-timeout-secs` | `AETHER_PROXY_TUNNEL_CONNECT_TIMEOUT_SECS` | `15` | TCP + TLS 握手超时（秒） |
+| `--tunnel-tcp-keepalive-secs` | `AETHER_PROXY_TUNNEL_TCP_KEEPALIVE_SECS` | `30` | TCP keepalive 初始延迟（秒） |
+| `--tunnel-tcp-nodelay` | `AETHER_PROXY_TUNNEL_TCP_NODELAY` | `true` | 禁用 Nagle 算法 |
+| `--tunnel-ping-interval-secs` | `AETHER_PROXY_TUNNEL_PING_INTERVAL_SECS` | `15` | WebSocket Ping 频率（秒） |
+| `--tunnel-stale-timeout-secs` | `AETHER_PROXY_TUNNEL_STALE_TIMEOUT_SECS` | `45` | 无数据断连阈值（秒） |
+| `--tunnel-reconnect-base-ms` | `AETHER_PROXY_TUNNEL_RECONNECT_BASE_MS` | `500` | 指数退避基础延迟（毫秒） |
+| `--tunnel-reconnect-max-ms` | `AETHER_PROXY_TUNNEL_RECONNECT_MAX_MS` | `30000` | 指数退避上限（毫秒） |
+
+#### 上游 HTTP 请求
+
+| 参数 | 环境变量 | 默认值 | 说明 |
+|------|----------|--------|------|
+| `--upstream-connect-timeout-secs` | `AETHER_PROXY_UPSTREAM_CONNECT_TIMEOUT_SECS` | `30` | 上游建连超时（秒） |
+| `--upstream-pool-max-idle-per-host` | `AETHER_PROXY_UPSTREAM_POOL_MAX_IDLE_PER_HOST` | `64` | 每 Host 最大空闲连接数 |
+| `--upstream-pool-idle-timeout-secs` | `AETHER_PROXY_UPSTREAM_POOL_IDLE_TIMEOUT_SECS` | `300` | 连接池空闲超时（秒） |
+| `--upstream-tcp-keepalive-secs` | `AETHER_PROXY_UPSTREAM_TCP_KEEPALIVE_SECS` | `60` | TCP keepalive（秒，0 关闭） |
+| `--upstream-tcp-nodelay` | `AETHER_PROXY_UPSTREAM_TCP_NODELAY` | `true` | 启用 TCP_NODELAY |
+
+#### Aether API 客户端
+
+| 参数 | 环境变量 | 默认值 | 说明 |
+|------|----------|--------|------|
+| `--aether-request-timeout-secs` | `AETHER_PROXY_AETHER_REQUEST_TIMEOUT_SECS` | `10` | 请求总超时（秒） |
+| `--aether-connect-timeout-secs` | `AETHER_PROXY_AETHER_CONNECT_TIMEOUT_SECS` | `10` | 建连超时（秒） |
+| `--aether-retry-max-attempts` | `AETHER_PROXY_AETHER_RETRY_MAX_ATTEMPTS` | `3` | 最大重试次数 |
+
+#### DNS 与安全
+
+| 参数 | 环境变量 | 默认值 | 说明 |
+|------|----------|--------|------|
+| `--dns-cache-ttl-secs` | `AETHER_PROXY_DNS_CACHE_TTL_SECS` | `60` | DNS 缓存 TTL（秒） |
+| `--dns-cache-capacity` | `AETHER_PROXY_DNS_CACHE_CAPACITY` | `1024` | DNS 缓存容量（条目数） |
+
+#### 日志
+
+| 参数 | 环境变量 | 默认值 | 说明 |
+|------|----------|--------|------|
 | `--log-level` | `AETHER_PROXY_LOG_LEVEL` | `info` | 日志级别 |
 | `--log-json` | `AETHER_PROXY_LOG_JSON` | `false` | JSON 格式日志 |
-| `--enable-tls` | `AETHER_PROXY_ENABLE_TLS` | `true` | 启用 TLS |
-| `--tls-cert` | `AETHER_PROXY_TLS_CERT` | `aether-proxy-cert.pem` | TLS 证书路径 |
-| `--tls-key` | `AETHER_PROXY_TLS_KEY` | `aether-proxy-key.pem` | TLS 私钥路径 |
+
+### 多服务器配置
+
+在 `aether-proxy.toml` 中使用 `[[servers]]` 配置多个 Aether 服务器：
+
+```toml
+[[servers]]
+aether_url = "https://aether-1.example.com"
+management_token = "ae_xxx"
+node_name = "jp-proxy-01"
+
+[[servers]]
+aether_url = "https://aether-2.example.com"
+management_token = "ae_yyy"
+node_name = "jp-proxy-02"
+```

 ## 发布新版本

-推送 `proxy-v*` 格式的 tag，GitHub Actions 会自动编译所有平台并发布到 Releases：
+推送 `proxy-v*` 格式的 tag，GitHub Actions 会自动：
+- 编译所有平台二进制并发布到 Releases
+- 构建 Docker 镜像并推送到 GHCR 和 Docker Hub
+- 更新 README 中的下载链接表格

 ```bash
-git tag proxy-v0.1.0
-git push origin proxy-v0.1.0
+git tag proxy-v0.2.0
+git push origin proxy-v0.2.0
 ```
@@ -0,0 +1,14 @@
+services:
+  aether-proxy:
+    image: ghcr.io/fawney19/aether-proxy:latest
+    container_name: aether-proxy
+    restart: unless-stopped
+    env_file:
+      - .env
+    environment:
+      AETHER_PROXY_LOG_JSON: "true"
+    logging:
+      driver: json-file
+      options:
+        max-size: "50m"
+        max-file: "3"
@@ -1,39 +1,41 @@
 //! Application lifecycle: initialization, task orchestration, and shutdown.
-//!
-//! Extracted from `main.rs` to keep the entry point minimal and consolidate
-//! the startup sequence, tracing init, and graceful shutdown logic.

 use std::sync::atomic::AtomicU64;
 use std::sync::{Arc, RwLock};
+use std::time::Duration;

+use arc_swap::ArcSwap;
 use tokio::signal;
-use tokio::sync::watch;
-use tracing::{error, info};
+use tokio::sync::{watch, Mutex};
+use tracing::{error, info, warn};

-use crate::config::Config;
+use crate::config::{Config, ServerEntry};
 use crate::net;
 use crate::registration::client::AetherClient;
 use crate::runtime::{self, DynamicConfig};
-use crate::state::AppState;
-use crate::{hardware, proxy};
+use crate::state::{AppState, ProxyMetrics, ServerContext};
+use crate::upstream_client;
+use crate::{hardware, target_filter, tunnel};

 /// Run the full application lifecycle after config has been parsed.
-pub async fn run(mut config: Config) -> anyhow::Result<()> {
+pub async fn run(mut config: Config, servers: Vec<ServerEntry>) -> anyhow::Result<()> {
+    config.validate()?;
    init_tracing(&config);

    info!(
        version = env!("CARGO_PKG_VERSION"),
-        port = config.listen_port,
        node_name = %config.node_name,
-        "aether-proxy starting"
+        server_count = servers.len(),
+        "aether-proxy starting (tunnel mode)"
    );

-    // Resolve public IP
+    // Resolve public IP (best-effort for region info)
    let public_ip = match &config.public_ip {
        Some(ip) => ip.clone(),
-        None => net::detect_public_ip().await?,
+        None => net::detect_public_ip()
+            .await
+            .unwrap_or_else(|_| "0.0.0.0".to_string()),
    };
-    info!(public_ip = %public_ip, "using public IP");

    // Auto-detect region if not configured
    if config.node_region.is_none() {
@@ -42,111 +44,270 @@ pub async fn run(mut config: Config) -> anyhow::Result<()> {
        }
    }

-    // Initialize TLS if enabled
-    let (tls_acceptor, tls_fingerprint) = if config.enable_tls {
-        let cert_path = std::path::PathBuf::from(&config.tls_cert);
-        let key_path = std::path::PathBuf::from(&config.tls_key);
-
-        proxy::tls::ensure_self_signed_cert(&cert_path, &key_path)?;
-        let acceptor = proxy::tls::build_tls_acceptor(&cert_path, &key_path)?;
-        let fingerprint = proxy::tls::cert_sha256_fingerprint(&cert_path)?;
-
-        info!(fingerprint = %fingerprint, "TLS enabled");
-        (Some(acceptor), Some(fingerprint))
-    } else {
-        info!("TLS disabled");
-        (None, None)
-    };
-
-    // Collect hardware info (once at startup)
+    // Collect hardware info (once at startup, sent during registration)
    let hw_info = hardware::collect();

-    // Register with Aether
-    let aether_client = Arc::new(AetherClient::new(&config));
-    let node_id = aether_client
-        .register(
+    // Auto-detect tunnel_max_streams from hardware if not explicitly set
+    if config.tunnel_max_streams.is_none() {
+        let auto = (hw_info.estimated_max_concurrency / 10).clamp(64, 1024) as u32;
+        config.tunnel_max_streams = Some(auto);
+        info!(
+            tunnel_max_streams = auto,
+            "auto-detected tunnel_max_streams from hardware"
+        );
+    }
+
+    info!(
+        max_concurrency = hw_info.estimated_max_concurrency,
+        "hardware info collected"
+    );
+
+    let dns_cache = Arc::new(target_filter::DnsCache::new(
+        Duration::from_secs(config.dns_cache_ttl_secs),
+        config.dns_cache_capacity,
+    ));
+
+    // Build Hyper client for tunnel upstream requests (shared).
+    // DNS still flows through validated addresses from DnsCache, while the
+    // custom connector exposes per-request connect/TLS timing when available.
+    let upstream_client = upstream_client::build_upstream_client(&config, Arc::clone(&dns_cache));
+
+    // Register with each Aether server and build per-server contexts.
+    // Wrapped in Arc<Mutex> so retry_failed_registrations can append later.
+    let server_contexts: Arc<Mutex<Vec<Arc<ServerContext>>>> = Arc::new(Mutex::new(Vec::new()));
+    let mut failed_entries: Vec<(String, ServerEntry)> = Vec::new();
+    for (i, entry) in servers.iter().enumerate() {
+        let label = if servers.len() == 1 {
+            "server".to_string()
+        } else {
+            format!("server-{}", i)
+        };
+        let node_name = entry
+            .node_name
+            .clone()
+            .unwrap_or_else(|| config.node_name.clone());
+        let client = Arc::new(AetherClient::new(
            &config,
-            &public_ip,
-            config.enable_tls,
-            tls_fingerprint.as_deref(),
-            Some(&hw_info),
-        )
-        .await?;
+            &entry.aether_url,
+            &entry.management_token,
+        ));
+        match client
+            .register(&config, &node_name, &public_ip, Some(&hw_info))
+            .await
+        {
+            Ok(node_id) => {
+                info!(server = %label, node_id = %node_id, url = %entry.aether_url, node_name = %node_name, "registered");
+                // Initialize dynamic config with per-server node_name (not global),
+                // so that the heartbeat and reconnect use the correct name.
+                let mut dynamic = DynamicConfig::from_config(&config);
+                dynamic.node_name = node_name.clone();
+                server_contexts.lock().await.push(Arc::new(ServerContext {
+                    server_label: label,
+                    aether_url: entry.aether_url.clone(),
+                    management_token: entry.management_token.clone(),
+                    node_name,
+                    node_id: Arc::new(RwLock::new(node_id)),
+                    aether_client: client,
+                    dynamic: Arc::new(ArcSwap::from_pointee(dynamic)),
+                    active_connections: Arc::new(AtomicU64::new(0)),
+                    metrics: Arc::new(ProxyMetrics::new()),
+                }));
+            }
+            Err(e) => {
+                warn!(
+                    server = %label,
+                    url = %entry.aether_url,
+                    error = %e,
+                    "registration failed, will retry in background"
+                );
+                failed_entries.push((label, entry.clone()));
+            }
+        }
+    }

-    info!(node_id = %node_id, "node registered");
-
-    // Build DynamicConfig before moving config into Arc
-    let dynamic = Arc::new(RwLock::new(DynamicConfig::from_config(&config)));
-
-    // Build delegate HTTP client (for proxy-initiated upstream requests).
-    // No overall timeout — SSE streams can last indefinitely.
-    // Connect timeout limits connection establishment; Aether controls
-    // first-byte / idle timeouts on its own side.
-    let delegate_client = reqwest::Client::builder()
-        .connect_timeout(std::time::Duration::from_secs(30))
-        .pool_max_idle_per_host(20)
-        .pool_idle_timeout(std::time::Duration::from_secs(90))
-        .build()
-        .expect("failed to create delegate HTTP client");
+    {
+        let ctx_count = server_contexts.lock().await.len();
+        if ctx_count == 0 && failed_entries.is_empty() {
+            anyhow::bail!("no servers configured");
+        }
+        if ctx_count == 0 {
+            anyhow::bail!(
+                "no servers registered successfully (all {} failed)",
+                failed_entries.len()
+            );
+        }
+    }

    // Build shared application state
+    let tunnel_tls_config = Arc::new(crate::tunnel::client::build_tls_config());
    let state = Arc::new(AppState {
        config: Arc::new(config),
-        node_id: Arc::new(RwLock::new(node_id)),
-        dynamic,
-        aether_client,
-        hardware_info: Arc::new(hw_info),
-        public_ip,
-        tls_fingerprint,
-        tls_acceptor,
-        delegate_client,
-        active_connections: Arc::new(AtomicU64::new(0)),
+        dns_cache,
+        upstream_client,
+        tunnel_tls_config,
    });

    // Shutdown signal channel
    let (shutdown_tx, shutdown_rx) = watch::channel(false);

-    // Start heartbeat task
-    let heartbeat_handle = {
-        let state = Arc::clone(&state);
-        let rx = shutdown_rx.clone();
-        tokio::spawn(async move {
-            crate::registration::heartbeat::run(&state, rx).await;
-        })
-    };
+    info!(
+        active_servers = server_contexts.lock().await.len(),
+        "running in tunnel mode"
+    );

-    // Start proxy server
-    let server_handle = {
-        let state = Arc::clone(&state);
-        let rx = shutdown_rx.clone();
-        tokio::spawn(async move {
-            if let Err(e) = proxy::server::run(&state, rx).await {
-                error!(error = %e, "proxy server error");
-            }
-        })
-    };
-
-    // Wait for shutdown signal (SIGTERM or SIGINT)
-    wait_for_shutdown().await;
-
-    info!("shutdown signal received, cleaning up...");
-
-    // Signal all tasks to stop
-    let _ = shutdown_tx.send(true);
-
-    // Graceful unregister (best-effort)
-    let current_node_id = state.node_id.read().unwrap().clone();
-    if let Err(e) = state.aether_client.unregister(&current_node_id).await {
-        error!(error = %e, "unregister failed during shutdown");
+    // Spawn tunnel connections per server (pool_size connections each)
+    let pool_size = state.config.tunnel_connections.max(1) as usize;
+    let mut tunnel_handles = Vec::new();
+    for server in server_contexts.lock().await.iter() {
+        for conn_idx in 0..pool_size {
+            let s = Arc::clone(&state);
+            let srv = Arc::clone(server);
+            let rx = shutdown_rx.clone();
+            tunnel_handles.push(tokio::spawn(async move {
+                tunnel::run(&s, &srv, conn_idx, rx).await;
+            }));
+        }
    }

-    // Wait for tasks to finish
-    let _ = tokio::join!(heartbeat_handle, server_handle);
+    // Spawn background retry for failed server registrations
+    if !failed_entries.is_empty() {
+        let retry_state = Arc::clone(&state);
+        let retry_contexts = Arc::clone(&server_contexts);
+        let retry_public_ip = public_ip.clone();
+        let retry_hw_info = hw_info.clone();
+        let retry_shutdown = shutdown_rx.clone();
+        let retry_pool_size = pool_size;
+        tokio::spawn(async move {
+            retry_failed_registrations(
+                retry_state,
+                retry_contexts,
+                failed_entries,
+                retry_public_ip,
+                retry_hw_info,
+                retry_pool_size,
+                retry_shutdown,
+            )
+            .await;
+        });
+    }
+
+    // Wait for shutdown signal
+    wait_for_shutdown().await;
+    info!("shutdown signal received, cleaning up...");
+    let _ = shutdown_tx.send(true);
+
+    // Graceful unregister from all servers (including retry-registered ones)
+    for server in server_contexts.lock().await.iter() {
+        let node_id = server.node_id.read().unwrap().clone();
+        if let Err(e) = server.aether_client.unregister(&node_id).await {
+            error!(
+                server = %server.server_label,
+                error = %e,
+                "unregister failed during shutdown"
+            );
+        }
+    }
+
+    // Wait for all tunnel tasks
+    for h in tunnel_handles {
+        let _ = h.await;
+    }

    info!("aether-proxy stopped");
    Ok(())
 }

+/// Retry interval for failed server registrations (5 minutes).
+const REGISTRATION_RETRY_INTERVAL: Duration = Duration::from_secs(300);
+/// Max registration retry attempts before giving up.
+const REGISTRATION_RETRY_MAX: u32 = 12;
+
+/// Background task that retries registration for servers that failed at startup.
+async fn retry_failed_registrations(
+    state: Arc<AppState>,
+    server_contexts: Arc<Mutex<Vec<Arc<ServerContext>>>>,
+    failed: Vec<(String, ServerEntry)>,
+    public_ip: String,
+    hw_info: crate::hardware::HardwareInfo,
+    pool_size: usize,
+    mut shutdown: watch::Receiver<bool>,
+) {
+    for (label, entry) in &failed {
+        let node_name = entry
+            .node_name
+            .clone()
+            .unwrap_or_else(|| state.config.node_name.clone());
+        let client = Arc::new(AetherClient::new(
+            &state.config,
+            &entry.aether_url,
+            &entry.management_token,
+        ));
+
+        let mut attempt = 0u32;
+        let node_id = loop {
+            attempt += 1;
+
+            tokio::select! {
+                _ = tokio::time::sleep(REGISTRATION_RETRY_INTERVAL) => {}
+                _ = shutdown.changed() => {
+                    info!(server = %label, "shutdown during registration retry");
+                    return;
+                }
+            }
+
+            match client
+                .register(&state.config, &node_name, &public_ip, Some(&hw_info))
+                .await
+            {
+                Ok(id) => {
+                    info!(server = %label, node_id = %id, attempt, "registration retry succeeded");
+                    break id;
+                }
+                Err(e) => {
+                    warn!(
+                        server = %label,
+                        attempt,
+                        max = REGISTRATION_RETRY_MAX,
+                        error = %e,
+                        "registration retry failed"
+                    );
+                    if attempt >= REGISTRATION_RETRY_MAX {
+                        error!(server = %label, "giving up registration after {} attempts", attempt);
+                        return;
+                    }
+                }
+            }
+        };
+
+        // Build server context and spawn tunnels
+        let mut dynamic = DynamicConfig::from_config(&state.config);
+        dynamic.node_name = node_name.clone();
+        let server = Arc::new(ServerContext {
+            server_label: label.clone(),
+            aether_url: entry.aether_url.clone(),
+            management_token: entry.management_token.clone(),
+            node_name,
+            node_id: Arc::new(RwLock::new(node_id)),
+            aether_client: client,
+            dynamic: Arc::new(ArcSwap::from_pointee(dynamic)),
+            active_connections: Arc::new(AtomicU64::new(0)),
+            metrics: Arc::new(ProxyMetrics::new()),
+        });
+
+        // Add to shared list so shutdown can unregister this server
+        server_contexts.lock().await.push(Arc::clone(&server));
+
+        for conn_idx in 0..pool_size {
+            let s = Arc::clone(&state);
+            let srv = Arc::clone(&server);
+            let rx = shutdown.clone();
+            tokio::spawn(async move {
+                tunnel::run(&s, &srv, conn_idx, rx).await;
+            });
+        }
+    }
+}
+
 fn init_tracing(config: &Config) {
    use tracing_subscriber::prelude::*;
    use tracing_subscriber::{reload, EnvFilter};
@@ -155,7 +316,6 @@ fn init_tracing(config: &Config) {

    let (filter_layer, reload_handle) = reload::Layer::new(filter);

-    // Register log-level hot-reloader
    runtime::set_log_reloader(Box::new(move |level: &str| {
        if let Ok(new_filter) = EnvFilter::try_new(level) {
            let _ = reload_handle.modify(|f| *f = new_filter);
@@ -1,176 +0,0 @@
-use base64::Engine;
-use hmac::{Hmac, Mac};
-use sha2::Sha256;
-use subtle::ConstantTimeEq;
-
-use crate::config::Config;
-
-type HmacSha256 = Hmac<Sha256>;
-
-#[derive(Debug)]
-pub enum AuthError {
-    MissingHeader,
-    InvalidBasicAuth,
-    InvalidUsername,
-    InvalidPasswordFormat,
-    TimestampParseError,
-    TimestampExpired,
-    SignatureMismatch,
-}
-
-impl std::fmt::Display for AuthError {
-    fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
-        match self {
-            Self::MissingHeader => write!(f, "missing Proxy-Authorization header"),
-            Self::InvalidBasicAuth => write!(f, "invalid Basic auth encoding"),
-            Self::InvalidUsername => write!(f, "username must be 'hmac'"),
-            Self::InvalidPasswordFormat => {
-                write!(f, "password format must be 'timestamp.signature'")
-            }
-            Self::TimestampParseError => write!(f, "invalid timestamp"),
-            Self::TimestampExpired => write!(f, "timestamp outside tolerance window"),
-            Self::SignatureMismatch => write!(f, "HMAC signature mismatch"),
-        }
-    }
-}
-
-/// Validate Proxy-Authorization header.
-///
-/// Expected format: `Basic base64(hmac:{timestamp}.{signature})`
-/// where signature = hex(HMAC-SHA256(hmac_key, "{timestamp}"))
-///
-/// The signature no longer includes `node_id`, eliminating race conditions
-/// during re-registration where the Aether server's cached `node_id` could
-/// differ from the proxy's freshly assigned `node_id`.
-///
-/// `timestamp_tolerance` is accepted separately so the caller can supply
-/// the value from [`DynamicConfig`](crate::runtime::DynamicConfig) (which
-/// may be updated remotely).
-pub fn validate_proxy_auth(
-    proxy_auth_header: Option<&str>,
-    config: &Config,
-    timestamp_tolerance: u64,
-) -> Result<(), AuthError> {
-    let header = proxy_auth_header.ok_or(AuthError::MissingHeader)?;
-
-    let encoded = header
-        .strip_prefix("Basic ")
-        .or_else(|| header.strip_prefix("basic "))
-        .ok_or(AuthError::InvalidBasicAuth)?;
-
-    let decoded_bytes = base64::engine::general_purpose::STANDARD
-        .decode(encoded.trim())
-        .map_err(|_| AuthError::InvalidBasicAuth)?;
-
-    let decoded = String::from_utf8(decoded_bytes).map_err(|_| AuthError::InvalidBasicAuth)?;
-
-    // format: hmac:{timestamp}.{signature}
-    let (username, password) = decoded.split_once(':').ok_or(AuthError::InvalidBasicAuth)?;
-
-    if username != "hmac" {
-        return Err(AuthError::InvalidUsername);
-    }
-
-    let (timestamp_str, signature_hex) = password
-        .split_once('.')
-        .ok_or(AuthError::InvalidPasswordFormat)?;
-
-    // Validate timestamp window
-    let timestamp: u64 = timestamp_str
-        .parse()
-        .map_err(|_| AuthError::TimestampParseError)?;
-
-    let now = std::time::SystemTime::now()
-        .duration_since(std::time::UNIX_EPOCH)
-        .expect("system clock before epoch")
-        .as_secs();
-
-    let diff = now.abs_diff(timestamp);
-
-    if diff > timestamp_tolerance {
-        return Err(AuthError::TimestampExpired);
-    }
-
-    // Recompute signature: HMAC-SHA256(key, timestamp)
-    let mut mac =
-        HmacSha256::new_from_slice(config.hmac_key.as_bytes()).expect("HMAC accepts any key size");
-    mac.update(timestamp_str.as_bytes());
-    let expected = mac.finalize().into_bytes();
-    let expected_hex = hex::encode(expected);
-
-    // Constant-time comparison
-    let sig_bytes = signature_hex.as_bytes();
-    let exp_bytes = expected_hex.as_bytes();
-
-    if sig_bytes.len() != exp_bytes.len() || sig_bytes.ct_eq(exp_bytes).unwrap_u8() != 1 {
-        return Err(AuthError::SignatureMismatch);
-    }
-
-    Ok(())
-}
-
-#[cfg(test)]
-mod tests {
-    use super::*;
-
-    fn make_config() -> Config {
-        Config {
-            aether_url: String::new(),
-            management_token: String::new(),
-            hmac_key: "test-hmac-key".to_string(),
-            listen_port: 18080,
-            public_ip: None,
-            node_name: "test".to_string(),
-            node_region: None,
-            heartbeat_interval: 30,
-            allowed_ports: vec![80, 443],
-            timestamp_tolerance: 300,
-            log_level: "info".to_string(),
-            log_json: false,
-            enable_tls: false,
-            tls_cert: String::new(),
-            tls_key: String::new(),
-        }
-    }
-
-    fn make_valid_auth(config: &Config) -> String {
-        let now = std::time::SystemTime::now()
-            .duration_since(std::time::UNIX_EPOCH)
-            .unwrap()
-            .as_secs();
-        let mut mac = HmacSha256::new_from_slice(config.hmac_key.as_bytes()).unwrap();
-        mac.update(now.to_string().as_bytes());
-        let sig = hex::encode(mac.finalize().into_bytes());
-        let cred = format!("hmac:{}.{}", now, sig);
-        let encoded = base64::engine::general_purpose::STANDARD.encode(cred);
-        format!("Basic {}", encoded)
-    }
-
-    #[test]
-    fn test_valid_auth() {
-        let config = make_config();
-        let header = make_valid_auth(&config);
-        assert!(validate_proxy_auth(Some(&header), &config, config.timestamp_tolerance).is_ok());
-    }
-
-    #[test]
-    fn test_missing_header() {
-        let config = make_config();
-        assert!(matches!(
-            validate_proxy_auth(None, &config, config.timestamp_tolerance),
-            Err(AuthError::MissingHeader)
-        ));
-    }
-
-    #[test]
-    fn test_wrong_username() {
-        let cred = "user:12345.abc";
-        let encoded = base64::engine::general_purpose::STANDARD.encode(cred);
-        let header = format!("Basic {}", encoded);
-        let config = make_config();
-        assert!(matches!(
-            validate_proxy_auth(Some(&header), &config, config.timestamp_tolerance),
-            Err(AuthError::InvalidUsername)
-        ));
-    }
-}
@@ -1,3 +0,0 @@
-pub mod hmac;
-
-pub use self::hmac::validate_proxy_auth;
@@ -3,11 +3,41 @@ use std::path::Path;
 use clap::Parser;
 use serde::{Deserialize, Serialize};

-/// Aether forward proxy with HMAC authentication.
+/// Fields that existed in 0.1.x but were removed in 0.2.0.
+const LEGACY_ONLY_KEYS: &[&str] = &[
+    "hmac_key",
+    "listen_port",
+    "timestamp_tolerance",
+    "connect_timeout_secs",
+    "tls_handshake_timeout_secs",
+    "enable_tls",
+    "tls_cert",
+    "tls_key",
+];
+
+/// Fields renamed from 0.1.x `delegate_*` to 0.2.0 `upstream_*`.
+const DELEGATE_TO_UPSTREAM: &[(&str, &str)] = &[
+    (
+        "delegate_connect_timeout_secs",
+        "upstream_connect_timeout_secs",
+    ),
+    (
+        "delegate_pool_max_idle_per_host",
+        "upstream_pool_max_idle_per_host",
+    ),
+    (
+        "delegate_pool_idle_timeout_secs",
+        "upstream_pool_idle_timeout_secs",
+    ),
+    ("delegate_tcp_keepalive_secs", "upstream_tcp_keepalive_secs"),
+    ("delegate_tcp_nodelay", "upstream_tcp_nodelay"),
+];
+
+/// Aether tunnel proxy.
 ///
 /// Deployed on overseas VPS to relay API traffic for Aether instances
-/// behind the GFW. Registers with Aether, sends heartbeats, and validates
-/// incoming proxy requests via HMAC-SHA256 signatures in Basic Auth.
+/// behind the GFW. Connects to Aether via WebSocket tunnel, registers
+/// with Aether, and relays upstream requests.
 #[derive(Parser, Debug, Clone)]
 #[command(version, about)]
 pub struct Config {
@@ -19,14 +49,6 @@ pub struct Config {
    #[arg(long, env = "AETHER_PROXY_MANAGEMENT_TOKEN")]
    pub management_token: String,

-    /// HMAC-SHA256 key for proxy authentication
-    #[arg(long, env = "AETHER_PROXY_HMAC_KEY")]
-    pub hmac_key: String,
-
-    /// Port to listen on for proxy connections
-    #[arg(long, env = "AETHER_PROXY_LISTEN_PORT", default_value_t = 18080)]
-    pub listen_port: u16,
-
    /// Public IP address of this node (auto-detected if omitted)
    #[arg(long, env = "AETHER_PROXY_PUBLIC_IP")]
    pub public_ip: Option<String>,
@@ -44,12 +66,133 @@ pub struct Config {
    pub heartbeat_interval: u64,

    /// Allowed destination ports (default: 80,443,8080,8443)
-    #[arg(long, env = "AETHER_PROXY_ALLOWED_PORTS", value_delimiter = ',', default_values_t = vec![80, 443, 8080, 8443])]
+    #[arg(
+        long,
+        env = "AETHER_PROXY_ALLOWED_PORTS",
+        value_delimiter = ',',
+        default_values_t = vec![80, 443, 8080, 8443]
+    )]
    pub allowed_ports: Vec<u16>,

-    /// Timestamp tolerance window in seconds for HMAC validation
-    #[arg(long, env = "AETHER_PROXY_TIMESTAMP_TOLERANCE", default_value_t = 300)]
-    pub timestamp_tolerance: u64,
+    /// Aether API request timeout in seconds
+    #[arg(
+        long,
+        env = "AETHER_PROXY_AETHER_REQUEST_TIMEOUT",
+        default_value_t = 10
+    )]
+    pub aether_request_timeout_secs: u64,
+
+    /// Aether API connect timeout in seconds
+    #[arg(
+        long,
+        env = "AETHER_PROXY_AETHER_CONNECT_TIMEOUT",
+        default_value_t = 10
+    )]
+    pub aether_connect_timeout_secs: u64,
+
+    /// Aether API max idle connections per host
+    #[arg(
+        long,
+        env = "AETHER_PROXY_AETHER_POOL_MAX_IDLE_PER_HOST",
+        default_value_t = 8
+    )]
+    pub aether_pool_max_idle_per_host: usize,
+
+    /// Aether API idle timeout in seconds
+    #[arg(
+        long,
+        env = "AETHER_PROXY_AETHER_POOL_IDLE_TIMEOUT",
+        default_value_t = 90
+    )]
+    pub aether_pool_idle_timeout_secs: u64,
+
+    /// Aether API TCP keepalive in seconds (0 disables)
+    #[arg(long, env = "AETHER_PROXY_AETHER_TCP_KEEPALIVE", default_value_t = 60)]
+    pub aether_tcp_keepalive_secs: u64,
+
+    /// Aether API TCP_NODELAY
+    #[arg(long, env = "AETHER_PROXY_AETHER_TCP_NODELAY", default_value_t = true)]
+    pub aether_tcp_nodelay: bool,
+
+    /// Enable HTTP/2 when talking to Aether API
+    #[arg(long, env = "AETHER_PROXY_AETHER_HTTP2", default_value_t = true)]
+    pub aether_http2: bool,
+
+    /// Aether API retry attempts (including initial)
+    #[arg(
+        long,
+        env = "AETHER_PROXY_AETHER_RETRY_MAX_ATTEMPTS",
+        default_value_t = 3
+    )]
+    pub aether_retry_max_attempts: u32,
+
+    /// Aether API retry base delay in milliseconds
+    #[arg(
+        long,
+        env = "AETHER_PROXY_AETHER_RETRY_BASE_DELAY_MS",
+        default_value_t = 200
+    )]
+    pub aether_retry_base_delay_ms: u64,
+
+    /// Aether API retry max delay in milliseconds
+    #[arg(
+        long,
+        env = "AETHER_PROXY_AETHER_RETRY_MAX_DELAY_MS",
+        default_value_t = 2000
+    )]
+    pub aether_retry_max_delay_ms: u64,
+
+    /// Maximum concurrent TCP connections (defaults to hardware estimate)
+    #[arg(long, env = "AETHER_PROXY_MAX_CONCURRENT_CONNECTIONS")]
+    pub max_concurrent_connections: Option<u64>,
+
+    /// DNS cache TTL in seconds
+    #[arg(long, env = "AETHER_PROXY_DNS_CACHE_TTL", default_value_t = 60)]
+    pub dns_cache_ttl_secs: u64,
+
+    /// DNS cache capacity (entries)
+    #[arg(long, env = "AETHER_PROXY_DNS_CACHE_CAPACITY", default_value_t = 1024)]
+    pub dns_cache_capacity: usize,
+
+    /// Upstream HTTP client connect timeout in seconds
+    #[arg(
+        long,
+        env = "AETHER_PROXY_UPSTREAM_CONNECT_TIMEOUT",
+        default_value_t = 30
+    )]
+    pub upstream_connect_timeout_secs: u64,
+
+    /// Upstream HTTP client max idle connections per host
+    #[arg(
+        long,
+        env = "AETHER_PROXY_UPSTREAM_POOL_MAX_IDLE_PER_HOST",
+        default_value_t = 64
+    )]
+    pub upstream_pool_max_idle_per_host: usize,
+
+    /// Upstream HTTP client idle timeout in seconds
+    #[arg(
+        long,
+        env = "AETHER_PROXY_UPSTREAM_POOL_IDLE_TIMEOUT",
+        default_value_t = 300
+    )]
+    pub upstream_pool_idle_timeout_secs: u64,
+
+    /// Upstream TCP keepalive in seconds (0 disables)
+    #[arg(
+        long,
+        env = "AETHER_PROXY_UPSTREAM_TCP_KEEPALIVE",
+        default_value_t = 60
+    )]
+    pub upstream_tcp_keepalive_secs: u64,
+
+    /// Upstream TCP_NODELAY
+    #[arg(
+        long,
+        env = "AETHER_PROXY_UPSTREAM_TCP_NODELAY",
+        default_value_t = true
+    )]
+    pub upstream_tcp_nodelay: bool,

    /// Log level (trace, debug, info, warn, error)
    #[arg(long, env = "AETHER_PROXY_LOG_LEVEL", default_value = "info")]
@@ -59,25 +202,106 @@ pub struct Config {
    #[arg(long, env = "AETHER_PROXY_LOG_JSON", default_value_t = false)]
    pub log_json: bool,

-    /// Enable TLS encryption (dual-stack: accepts both HTTP and TLS on same port)
-    #[arg(long, env = "AETHER_PROXY_ENABLE_TLS", default_value_t = true)]
-    pub enable_tls: bool,
-
-    /// Path to TLS certificate PEM file
+    /// Tunnel reconnect base delay in milliseconds (used by exponential backoff)
    #[arg(
        long,
-        env = "AETHER_PROXY_TLS_CERT",
-        default_value = "aether-proxy-cert.pem"
+        env = "AETHER_PROXY_TUNNEL_RECONNECT_BASE_MS",
+        default_value_t = 500
    )]
-    pub tls_cert: String,
+    pub tunnel_reconnect_base_ms: u64,

-    /// Path to TLS private key PEM file
+    /// Tunnel reconnect max delay in milliseconds (cap for exponential backoff)
    #[arg(
        long,
-        env = "AETHER_PROXY_TLS_KEY",
-        default_value = "aether-proxy-key.pem"
+        env = "AETHER_PROXY_TUNNEL_RECONNECT_MAX_MS",
+        default_value_t = 30000
    )]
-    pub tls_key: String,
+    pub tunnel_reconnect_max_ms: u64,
+
+    /// WebSocket tunnel ping interval in seconds
+    #[arg(long, env = "AETHER_PROXY_TUNNEL_PING_INTERVAL", default_value_t = 15)]
+    pub tunnel_ping_interval_secs: u64,
+
+    /// Maximum concurrent streams over tunnel (auto-detected from hardware if omitted)
+    #[arg(long, env = "AETHER_PROXY_TUNNEL_MAX_STREAMS")]
+    pub tunnel_max_streams: Option<u32>,
+
+    /// WebSocket tunnel TCP connect timeout in seconds
+    #[arg(
+        long,
+        env = "AETHER_PROXY_TUNNEL_CONNECT_TIMEOUT",
+        default_value_t = 15
+    )]
+    pub tunnel_connect_timeout_secs: u64,
+
+    /// WebSocket tunnel TCP keepalive in seconds (0 disables)
+    #[arg(long, env = "AETHER_PROXY_TUNNEL_TCP_KEEPALIVE", default_value_t = 30)]
+    pub tunnel_tcp_keepalive_secs: u64,
+
+    /// WebSocket tunnel TCP_NODELAY
+    #[arg(long, env = "AETHER_PROXY_TUNNEL_TCP_NODELAY", default_value_t = true)]
+    pub tunnel_tcp_nodelay: bool,
+
+    /// Tunnel connection staleness timeout in seconds (triggers reconnect if no data received)
+    #[arg(long, env = "AETHER_PROXY_TUNNEL_STALE_TIMEOUT", default_value_t = 45)]
+    pub tunnel_stale_timeout_secs: u64,
+
+    /// Number of parallel WebSocket tunnel connections per server (connection pool)
+    #[arg(long, env = "AETHER_PROXY_TUNNEL_CONNECTIONS", default_value_t = 3)]
+    pub tunnel_connections: u32,
+}
+
+impl Config {
+    /// Validate configuration values are within sane ranges.
+    /// Called after parsing to catch misconfigurations early.
+    pub fn validate(&self) -> anyhow::Result<()> {
+        if self.heartbeat_interval == 0 {
+            anyhow::bail!("heartbeat_interval must be > 0");
+        }
+        if self.heartbeat_interval > 3600 {
+            anyhow::bail!("heartbeat_interval must be <= 3600");
+        }
+        if self.allowed_ports.is_empty() {
+            anyhow::bail!("allowed_ports must not be empty");
+        }
+        for &port in &self.allowed_ports {
+            if port == 0 {
+                anyhow::bail!("allowed_ports: port 0 is not valid");
+            }
+        }
+        if self.tunnel_connect_timeout_secs == 0 {
+            anyhow::bail!("tunnel_connect_timeout_secs must be > 0");
+        }
+        if self.tunnel_ping_interval_secs == 0 {
+            anyhow::bail!("tunnel_ping_interval_secs must be > 0");
+        }
+        if self.tunnel_stale_timeout_secs <= self.tunnel_ping_interval_secs {
+            anyhow::bail!(
+                "tunnel_stale_timeout_secs ({}) must be > tunnel_ping_interval_secs ({})",
+                self.tunnel_stale_timeout_secs,
+                self.tunnel_ping_interval_secs
+            );
+        }
+        if self.tunnel_connections == 0 {
+            anyhow::bail!("tunnel_connections must be > 0");
+        }
+        if self.aether_retry_max_attempts == 0 {
+            anyhow::bail!("aether_retry_max_attempts must be >= 1");
+        }
+        if self.upstream_connect_timeout_secs == 0 {
+            anyhow::bail!("upstream_connect_timeout_secs must be > 0");
+        }
+        Ok(())
+    }
+}
+
+/// Per-server connection config (used in multi-server TOML `[[servers]]`).
+#[derive(Debug, Clone, Serialize, Deserialize)]
+pub struct ServerEntry {
+    pub aether_url: String,
+    pub management_token: String,
+    /// Per-server node name override. Falls back to the global `node_name`.
+    pub node_name: Option<String>,
 }

 // ---------------------------------------------------------------------------
@@ -85,7 +309,7 @@ pub struct Config {
 // ---------------------------------------------------------------------------

 /// Serializable config for TOML file persistence.
-/// All fields are optional — only populated values are written.
+/// All fields are optional -- only populated values are written.
 #[derive(Debug, Default, Serialize, Deserialize)]
 pub struct ConfigFile {
    #[serde(skip_serializing_if = "Option::is_none")]
@@ -93,10 +317,6 @@ pub struct ConfigFile {
    #[serde(skip_serializing_if = "Option::is_none")]
    pub management_token: Option<String>,
    #[serde(skip_serializing_if = "Option::is_none")]
-    pub hmac_key: Option<String>,
-    #[serde(skip_serializing_if = "Option::is_none")]
-    pub listen_port: Option<u16>,
-    #[serde(skip_serializing_if = "Option::is_none")]
    pub public_ip: Option<String>,
    #[serde(skip_serializing_if = "Option::is_none")]
    pub node_name: Option<String>,
@@ -107,17 +327,69 @@ pub struct ConfigFile {
    #[serde(skip_serializing_if = "Option::is_none")]
    pub allowed_ports: Option<Vec<u16>>,
    #[serde(skip_serializing_if = "Option::is_none")]
-    pub timestamp_tolerance: Option<u64>,
+    pub aether_request_timeout_secs: Option<u64>,
+    #[serde(skip_serializing_if = "Option::is_none")]
+    pub aether_connect_timeout_secs: Option<u64>,
+    #[serde(skip_serializing_if = "Option::is_none")]
+    pub aether_pool_max_idle_per_host: Option<usize>,
+    #[serde(skip_serializing_if = "Option::is_none")]
+    pub aether_pool_idle_timeout_secs: Option<u64>,
+    #[serde(skip_serializing_if = "Option::is_none")]
+    pub aether_tcp_keepalive_secs: Option<u64>,
+    #[serde(skip_serializing_if = "Option::is_none")]
+    pub aether_tcp_nodelay: Option<bool>,
+    #[serde(skip_serializing_if = "Option::is_none")]
+    pub aether_http2: Option<bool>,
+    #[serde(skip_serializing_if = "Option::is_none")]
+    pub aether_retry_max_attempts: Option<u32>,
+    #[serde(skip_serializing_if = "Option::is_none")]
+    pub aether_retry_base_delay_ms: Option<u64>,
+    #[serde(skip_serializing_if = "Option::is_none")]
+    pub aether_retry_max_delay_ms: Option<u64>,
+    #[serde(skip_serializing_if = "Option::is_none")]
+    pub max_concurrent_connections: Option<u64>,
+    #[serde(skip_serializing_if = "Option::is_none")]
+    pub dns_cache_ttl_secs: Option<u64>,
+    #[serde(skip_serializing_if = "Option::is_none")]
+    pub dns_cache_capacity: Option<usize>,
+    #[serde(skip_serializing_if = "Option::is_none")]
+    pub upstream_connect_timeout_secs: Option<u64>,
+    #[serde(skip_serializing_if = "Option::is_none")]
+    pub upstream_pool_max_idle_per_host: Option<usize>,
+    #[serde(skip_serializing_if = "Option::is_none")]
+    pub upstream_pool_idle_timeout_secs: Option<u64>,
+    #[serde(skip_serializing_if = "Option::is_none")]
+    pub upstream_tcp_keepalive_secs: Option<u64>,
+    #[serde(skip_serializing_if = "Option::is_none")]
+    pub upstream_tcp_nodelay: Option<bool>,
    #[serde(skip_serializing_if = "Option::is_none")]
    pub log_level: Option<String>,
    #[serde(skip_serializing_if = "Option::is_none")]
    pub log_json: Option<bool>,
    #[serde(skip_serializing_if = "Option::is_none")]
-    pub enable_tls: Option<bool>,
+    pub tunnel_reconnect_base_ms: Option<u64>,
    #[serde(skip_serializing_if = "Option::is_none")]
-    pub tls_cert: Option<String>,
+    pub tunnel_reconnect_max_ms: Option<u64>,
    #[serde(skip_serializing_if = "Option::is_none")]
-    pub tls_key: Option<String>,
+    pub tunnel_ping_interval_secs: Option<u64>,
+    #[serde(skip_serializing_if = "Option::is_none")]
+    pub tunnel_max_streams: Option<u32>,
+    #[serde(skip_serializing_if = "Option::is_none")]
+    pub tunnel_connect_timeout_secs: Option<u64>,
+    #[serde(skip_serializing_if = "Option::is_none")]
+    pub tunnel_tcp_keepalive_secs: Option<u64>,
+    #[serde(skip_serializing_if = "Option::is_none")]
+    pub tunnel_tcp_nodelay: Option<bool>,
+    #[serde(skip_serializing_if = "Option::is_none")]
+    pub tunnel_stale_timeout_secs: Option<u64>,
+    #[serde(skip_serializing_if = "Option::is_none")]
+    pub tunnel_connections: Option<u32>,
+
+    /// Multi-server config: each entry connects to a separate Aether instance.
+    /// When present, top-level aether_url/management_token are ignored for
+    /// tunnel connections (but still injected as env for clap compatibility).
+    #[serde(default, skip_serializing_if = "Vec::is_empty")]
+    pub servers: Vec<ServerEntry>,
 }

 impl ConfigFile {
@@ -134,6 +406,102 @@ impl ConfigFile {
        Ok(())
    }

+    /// Detect and migrate a 0.1.x config file to 0.2.0 format in-place.
+    ///
+    /// Returns `true` if migration was performed, `false` if already current.
+    /// The original file is backed up as `<name>.v1.bak` before rewriting.
+    pub fn migrate_legacy(path: &Path) -> anyhow::Result<bool> {
+        let content = match std::fs::read_to_string(path) {
+            Ok(c) => c,
+            Err(_) => return Ok(false),
+        };
+        let mut table: toml::map::Map<String, toml::Value> = toml::from_str(&content)?;
+
+        // Detect legacy format: presence of any 0.1.x-only key.
+        let is_legacy = LEGACY_ONLY_KEYS.iter().any(|k| table.contains_key(*k))
+            || DELEGATE_TO_UPSTREAM
+                .iter()
+                .any(|(old, _)| table.contains_key(*old));
+
+        if !is_legacy {
+            return Ok(false);
+        }
+
+        // 1. Rename delegate_* -> upstream_* (carry over user-customized values)
+        for &(old, new) in DELEGATE_TO_UPSTREAM {
+            if let Some(val) = table.remove(old) {
+                table.entry(new.to_string()).or_insert(val);
+            }
+        }
+
+        // 2. Build [[servers]] from top-level aether_url + management_token + node_name
+        if !table.contains_key("servers") {
+            let aether_url = table.get("aether_url").and_then(|v| v.as_str());
+            let management_token = table.get("management_token").and_then(|v| v.as_str());
+            if let (Some(url), Some(token)) = (aether_url, management_token) {
+                let mut entry = toml::map::Map::new();
+                entry.insert("aether_url".into(), toml::Value::String(url.to_string()));
+                entry.insert(
+                    "management_token".into(),
+                    toml::Value::String(token.to_string()),
+                );
+                if let Some(name) = table.get("node_name").and_then(|v| v.as_str()) {
+                    entry.insert("node_name".into(), toml::Value::String(name.to_string()));
+                }
+                table.insert(
+                    "servers".into(),
+                    toml::Value::Array(vec![toml::Value::Table(entry)]),
+                );
+            }
+        }
+
+        // 3. Remove top-level fields that are now in [[servers]] or obsolete
+        table.remove("aether_url");
+        table.remove("management_token");
+        table.remove("node_name");
+        for &key in LEGACY_ONLY_KEYS {
+            table.remove(key);
+        }
+
+        // 4. Backup original file (abort migration if backup fails)
+        let backup_path = path.with_extension("v1.bak");
+        std::fs::copy(path, &backup_path).map_err(|e| {
+            anyhow::anyhow!(
+                "failed to backup config before migration: {} -> {}: {}",
+                path.display(),
+                backup_path.display(),
+                e
+            )
+        })?;
+
+        // 5. Write migrated config
+        let new_content = toml::to_string_pretty(&table)?;
+        std::fs::write(path, &new_content)?;
+
+        eprintln!("  Config migrated from 0.1.x to 0.2.0 format.");
+        eprintln!("  Backup saved: {}", backup_path.display());
+
+        Ok(true)
+    }
+
+    /// Resolve the effective server list.
+    ///
+    /// If `[[servers]]` is present, use it. Otherwise fall back to the
+    /// top-level `aether_url` + `management_token` as a single server.
+    pub fn effective_servers(&self) -> Vec<ServerEntry> {
+        if !self.servers.is_empty() {
+            return self.servers.clone();
+        }
+        match (&self.aether_url, &self.management_token) {
+            (Some(url), Some(token)) => vec![ServerEntry {
+                aether_url: url.clone(),
+                management_token: token.clone(),
+                node_name: None,
+            }],
+            _ => vec![],
+        }
+    }
+
    /// Inject values as environment variables so clap picks them up.
    ///
    /// Only sets variables that are **not** already present in the
@@ -159,20 +527,119 @@ impl ConfigFile {
                }
            };
        }
-        set!("AETHER_PROXY_AETHER_URL", self.aether_url);
-        set!("AETHER_PROXY_MANAGEMENT_TOKEN", self.management_token);
-        set!("AETHER_PROXY_HMAC_KEY", self.hmac_key);
-        set!("AETHER_PROXY_LISTEN_PORT", self.listen_port);
+
+        // When top-level fields are absent, fall back to the first [[servers]]
+        // entry so that clap's required `aether_url` / `management_token` are
+        // satisfied even with the new config format.
+        let first_server = self.servers.first();
+        let aether_url = self
+            .aether_url
+            .as_deref()
+            .or(first_server.map(|s| s.aether_url.as_str()));
+        let management_token = self
+            .management_token
+            .as_deref()
+            .or(first_server.map(|s| s.management_token.as_str()));
+        let node_name = self
+            .node_name
+            .as_deref()
+            .or(first_server.and_then(|s| s.node_name.as_deref()));
+
+        set!("AETHER_PROXY_AETHER_URL", aether_url);
+        set!("AETHER_PROXY_MANAGEMENT_TOKEN", management_token);
        set!("AETHER_PROXY_PUBLIC_IP", self.public_ip);
-        set!("AETHER_PROXY_NODE_NAME", self.node_name);
+        set!("AETHER_PROXY_NODE_NAME", node_name);
        set!("AETHER_PROXY_NODE_REGION", self.node_region);
        set!("AETHER_PROXY_HEARTBEAT_INTERVAL", self.heartbeat_interval);
-        set!("AETHER_PROXY_TIMESTAMP_TOLERANCE", self.timestamp_tolerance);
+        set!(
+            "AETHER_PROXY_AETHER_REQUEST_TIMEOUT",
+            self.aether_request_timeout_secs
+        );
+        set!(
+            "AETHER_PROXY_AETHER_CONNECT_TIMEOUT",
+            self.aether_connect_timeout_secs
+        );
+        set!(
+            "AETHER_PROXY_AETHER_POOL_MAX_IDLE_PER_HOST",
+            self.aether_pool_max_idle_per_host
+        );
+        set!(
+            "AETHER_PROXY_AETHER_POOL_IDLE_TIMEOUT",
+            self.aether_pool_idle_timeout_secs
+        );
+        set!(
+            "AETHER_PROXY_AETHER_TCP_KEEPALIVE",
+            self.aether_tcp_keepalive_secs
+        );
+        set!("AETHER_PROXY_AETHER_TCP_NODELAY", self.aether_tcp_nodelay);
+        set!("AETHER_PROXY_AETHER_HTTP2", self.aether_http2);
+        set!(
+            "AETHER_PROXY_AETHER_RETRY_MAX_ATTEMPTS",
+            self.aether_retry_max_attempts
+        );
+        set!(
+            "AETHER_PROXY_AETHER_RETRY_BASE_DELAY_MS",
+            self.aether_retry_base_delay_ms
+        );
+        set!(
+            "AETHER_PROXY_AETHER_RETRY_MAX_DELAY_MS",
+            self.aether_retry_max_delay_ms
+        );
+        set!(
+            "AETHER_PROXY_MAX_CONCURRENT_CONNECTIONS",
+            self.max_concurrent_connections
+        );
+        set!("AETHER_PROXY_DNS_CACHE_TTL", self.dns_cache_ttl_secs);
+        set!("AETHER_PROXY_DNS_CACHE_CAPACITY", self.dns_cache_capacity);
+        set!(
+            "AETHER_PROXY_UPSTREAM_CONNECT_TIMEOUT",
+            self.upstream_connect_timeout_secs
+        );
+        set!(
+            "AETHER_PROXY_UPSTREAM_POOL_MAX_IDLE_PER_HOST",
+            self.upstream_pool_max_idle_per_host
+        );
+        set!(
+            "AETHER_PROXY_UPSTREAM_POOL_IDLE_TIMEOUT",
+            self.upstream_pool_idle_timeout_secs
+        );
+        set!(
+            "AETHER_PROXY_UPSTREAM_TCP_KEEPALIVE",
+            self.upstream_tcp_keepalive_secs
+        );
+        set!(
+            "AETHER_PROXY_UPSTREAM_TCP_NODELAY",
+            self.upstream_tcp_nodelay
+        );
        set!("AETHER_PROXY_LOG_LEVEL", self.log_level);
        set!("AETHER_PROXY_LOG_JSON", self.log_json);
-        set!("AETHER_PROXY_ENABLE_TLS", self.enable_tls);
-        set!("AETHER_PROXY_TLS_CERT", self.tls_cert);
-        set!("AETHER_PROXY_TLS_KEY", self.tls_key);
+        set!(
+            "AETHER_PROXY_TUNNEL_RECONNECT_BASE_MS",
+            self.tunnel_reconnect_base_ms
+        );
+        set!(
+            "AETHER_PROXY_TUNNEL_RECONNECT_MAX_MS",
+            self.tunnel_reconnect_max_ms
+        );
+        set!(
+            "AETHER_PROXY_TUNNEL_PING_INTERVAL",
+            self.tunnel_ping_interval_secs
+        );
+        set!("AETHER_PROXY_TUNNEL_MAX_STREAMS", self.tunnel_max_streams);
+        set!(
+            "AETHER_PROXY_TUNNEL_CONNECT_TIMEOUT",
+            self.tunnel_connect_timeout_secs
+        );
+        set!(
+            "AETHER_PROXY_TUNNEL_TCP_KEEPALIVE",
+            self.tunnel_tcp_keepalive_secs
+        );
+        set!("AETHER_PROXY_TUNNEL_TCP_NODELAY", self.tunnel_tcp_nodelay);
+        set!(
+            "AETHER_PROXY_TUNNEL_STALE_TIMEOUT",
+            self.tunnel_stale_timeout_secs
+        );
+        set!("AETHER_PROXY_TUNNEL_CONNECTIONS", self.tunnel_connections);

        // allowed_ports needs special handling (comma-separated)
        if let Some(ref ports) = self.allowed_ports {
@@ -1,13 +1,14 @@
 mod app;
-mod auth;
 mod config;
 mod hardware;
 mod net;
-mod proxy;
 mod registration;
 mod runtime;
 mod setup;
 mod state;
+mod target_filter;
+mod tunnel;
+mod upstream_client;

 use std::path::PathBuf;

@@ -56,8 +57,13 @@ async fn main() -> anyhow::Result<()> {
    // Load config file as env-var defaults (before clap parsing)
    let config_file_path =
        std::env::var("AETHER_PROXY_CONFIG").unwrap_or_else(|_| DEFAULT_CONFIG.to_string());
-    if std::path::Path::new(&config_file_path).exists() {
-        if let Ok(file_cfg) = config::ConfigFile::load(std::path::Path::new(&config_file_path)) {
+    let config_path = std::path::Path::new(&config_file_path);
+    if config_path.exists() {
+        // Migrate legacy 0.1.x config to 0.2.0 format if needed
+        if let Err(e) = config::ConfigFile::migrate_legacy(config_path) {
+            eprintln!("  WARNING: config migration failed: {}", e);
+        }
+        if let Ok(file_cfg) = config::ConfigFile::load(config_path) {
            file_cfg.inject_env();
        }
    }
@@ -130,10 +136,33 @@ async fn run_proxy(config: Config) -> anyhow::Result<()> {
    // Skip this check when we ARE the systemd service (INVOCATION_ID is set by systemd).
    if std::env::var_os("INVOCATION_ID").is_none() && setup::service::is_service_active() {
        eprintln!("Warning: systemd service is already running.");
-        eprintln!("Use `aether-proxy stop` to stop it first, or manage via subcommands:");
-        eprintln!("  aether-proxy status / logs / restart / stop");
+        eprintln!("Use `./aether-proxy stop` to stop it first, or manage via subcommands:");
+        eprintln!("  ./aether-proxy status / logs / restart / stop");
        std::process::exit(1);
    }

-    app::run(config).await
+    // Resolve server list: prefer [[servers]] from TOML, fall back to CLI/env single server.
+    let config_path =
+        std::env::var("AETHER_PROXY_CONFIG").unwrap_or_else(|_| DEFAULT_CONFIG.to_string());
+    let servers = if std::path::Path::new(&config_path).exists() {
+        config::ConfigFile::load(std::path::Path::new(&config_path))
+            .ok()
+            .map(|f| f.effective_servers())
+            .filter(|s| !s.is_empty())
+            .unwrap_or_else(|| {
+                vec![config::ServerEntry {
+                    aether_url: config.aether_url.clone(),
+                    management_token: config.management_token.clone(),
+                    node_name: None,
+                }]
+            })
+    } else {
+        vec![config::ServerEntry {
+            aether_url: config.aether_url.clone(),
+            management_token: config.management_token.clone(),
+            node_name: None,
+        }]
+    };
+
+    app::run(config, servers).await
 }
@@ -1,135 +0,0 @@
-use std::collections::HashSet;
-use std::sync::Arc;
-
-use hyper::body::Incoming;
-use hyper::{Request, Response};
-use tokio::net::TcpStream;
-use tracing::{debug, warn};
-
-use crate::auth;
-use crate::config::Config;
-use crate::proxy::target_filter;
-
-/// Handle HTTP CONNECT tunnel requests.
-///
-/// Flow: validate auth -> check target filter -> TCP connect -> 200 -> bidirectional copy
-pub async fn handle_connect(
-    req: Request<Incoming>,
-    config: Arc<Config>,
-    allowed_ports: &HashSet<u16>,
-    timestamp_tolerance: u64,
-) -> Response<http_body_util::Empty<bytes::Bytes>> {
-    // Extract Proxy-Authorization header
-    let proxy_auth = req
-        .headers()
-        .get("proxy-authorization")
-        .and_then(|v| v.to_str().ok());
-
-    // HMAC authentication
-    if let Err(e) = auth::validate_proxy_auth(proxy_auth, &config, timestamp_tolerance) {
-        warn!(error = %e, "CONNECT auth failed");
-        return proxy_auth_required(&e.to_string());
-    }
-
-    // Parse target host:port from CONNECT URI
-    let authority = match req.uri().authority() {
-        Some(auth) => auth.clone(),
-        None => {
-            warn!("CONNECT request missing authority");
-            return bad_request("missing target authority");
-        }
-    };
-
-    let host = authority.host().to_string();
-    let port = authority.port_u16().unwrap_or(443);
-
-    // Target filter: private IP + port whitelist
-    let target_addr = match target_filter::validate_target(&host, port, allowed_ports).await {
-        Ok(addr) => addr,
-        Err(e) => {
-            warn!(host = %host, port, error = %e, "CONNECT target rejected");
-            return forbidden(&e.to_string());
-        }
-    };
-
-    debug!(target = %target_addr, "CONNECT tunnel establishing");
-
-    // Connect to target
-    let target_stream = match TcpStream::connect(target_addr).await {
-        Ok(s) => s,
-        Err(e) => {
-            warn!(target = %target_addr, error = %e, "CONNECT target connection failed");
-            return bad_gateway(&e.to_string());
-        }
-    };
-
-    // Respond 200 and upgrade connection to raw TCP tunnel
-    let target_display = target_addr.to_string();
-    tokio::task::spawn(async move {
-        match hyper::upgrade::on(req).await {
-            Ok(upgraded) => {
-                let mut upgraded = hyper_util::rt::TokioIo::new(upgraded);
-                let mut target = target_stream;
-
-                match tokio::io::copy_bidirectional(&mut upgraded, &mut target).await {
-                    Ok((from_client, from_target)) => {
-                        debug!(
-                            target = %target_display,
-                            from_client,
-                            from_target,
-                            "CONNECT tunnel closed"
-                        );
-                    }
-                    Err(e) => {
-                        debug!(target = %target_display, error = %e, "CONNECT tunnel error");
-                    }
-                }
-            }
-            Err(e) => {
-                warn!(target = %target_display, error = %e, "CONNECT upgrade failed");
-            }
-        }
-    });
-
-    Response::builder()
-        .status(200)
-        .body(http_body_util::Empty::new())
-        .unwrap()
-}
-
-fn proxy_auth_required(msg: &str) -> Response<http_body_util::Empty<bytes::Bytes>> {
-    Response::builder()
-        .status(407)
-        .header("Proxy-Authenticate", "HMAC-SHA256")
-        .header("Content-Length", "0")
-        .header("X-Error", msg)
-        .body(http_body_util::Empty::new())
-        .unwrap()
-}
-
-fn forbidden(msg: &str) -> Response<http_body_util::Empty<bytes::Bytes>> {
-    Response::builder()
-        .status(403)
-        .header("Content-Length", "0")
-        .header("X-Error", msg)
-        .body(http_body_util::Empty::new())
-        .unwrap()
-}
-
-fn bad_request(msg: &str) -> Response<http_body_util::Empty<bytes::Bytes>> {
-    Response::builder()
-        .status(400)
-        .header("Content-Length", "0")
-        .header("X-Error", msg)
-        .body(http_body_util::Empty::new())
-        .unwrap()
-}
-
-fn bad_gateway(msg: &str) -> Response<http_body_util::Empty<bytes::Bytes>> {
-    Response::builder()
-        .status(502)
-        .header("Content-Length", "0")
-        .header("X-Error", msg)
-        .body(http_body_util::Empty::new())
-        .unwrap()
-}
@@ -1,214 +0,0 @@
-use std::collections::HashMap;
-use std::collections::HashSet;
-use std::sync::Arc;
-
-use futures_util::TryStreamExt;
-use http_body_util::{BodyExt, Full, Limited, StreamBody};
-use hyper::body::{Frame, Incoming};
-use hyper::{Request, Response};
-use serde::Deserialize;
-use tracing::{debug, warn};
-use url::Url;
-
-use super::BoxBody;
-use crate::auth;
-use crate::config::Config;
-use crate::proxy::target_filter;
-
-/// Delegation request payload sent by Aether.
-#[derive(Debug, Deserialize)]
-struct DelegateRequest {
-    method: String,
-    url: String,
-    headers: HashMap<String, String>,
-    body: Option<String>,
-    /// Accepted but not used on the proxy side — Aether controls timeouts.
-    #[allow(dead_code)]
-    timeout: Option<u64>,
-}
-
-/// Handle delegation requests: Aether sends a full request description,
-/// and the proxy issues the actual upstream HTTP call using its own TLS stack.
-///
-/// Endpoint: POST /_aether/delegate
-pub async fn handle_delegate(
-    req: Request<Incoming>,
-    config: Arc<Config>,
-    allowed_ports: &HashSet<u16>,
-    timestamp_tolerance: u64,
-    http_client: &reqwest::Client,
-) -> Response<BoxBody> {
-    // Authenticate via Authorization header (same HMAC scheme as Proxy-Authorization)
-    let auth_header = req
-        .headers()
-        .get("authorization")
-        .and_then(|v| v.to_str().ok());
-
-    if let Err(e) = auth::validate_proxy_auth(auth_header, &config, timestamp_tolerance) {
-        warn!(error = %e, "delegate auth failed");
-        return error_response(401, "authentication_failed", &e.to_string());
-    }
-
-    // Read and parse request body (limit to 10 MB to prevent OOM)
-    const MAX_BODY: usize = 10 * 1024 * 1024;
-    let body_bytes = match Limited::new(req.into_body(), MAX_BODY).collect().await {
-        Ok(collected) => collected.to_bytes(),
-        Err(e) => {
-            warn!(error = %e, "failed to read delegate request body");
-            return error_response(413, "payload_too_large", "request body exceeds 10MB limit");
-        }
-    };
-
-    let delegate_req: DelegateRequest = match serde_json::from_slice(&body_bytes) {
-        Ok(r) => r,
-        Err(e) => {
-            warn!(error = %e, "invalid delegate request JSON");
-            return error_response(400, "bad_request", &format!("invalid JSON: {}", e));
-        }
-    };
-
-    // Target filter: validate the upstream URL against allowed ports and private IP checks
-    let parsed_url = match Url::parse(&delegate_req.url) {
-        Ok(u) => u,
-        Err(e) => {
-            warn!(url = %delegate_req.url, error = %e, "invalid delegate target URL");
-            return error_response(400, "bad_request", &format!("invalid URL: {}", e));
-        }
-    };
-
-    let host = match parsed_url.host_str() {
-        Some(h) => h.to_string(),
-        None => {
-            warn!(url = %delegate_req.url, "delegate target URL missing host");
-            return error_response(400, "bad_request", "URL missing host");
-        }
-    };
-
-    let port = parsed_url.port_or_known_default().unwrap_or(443);
-
-    if let Err(e) = target_filter::validate_target(&host, port, allowed_ports).await {
-        warn!(host = %host, port, error = %e, "delegate target rejected");
-        return error_response(403, "target_not_allowed", &e.to_string());
-    }
-
-    debug!(
-        method = %delegate_req.method,
-        url = %delegate_req.url,
-        "delegate request"
-    );
-
-    // Build upstream request
-    let method = match delegate_req.method.parse::<reqwest::Method>() {
-        Ok(m) => m,
-        Err(e) => {
-            warn!(error = %e, method = %delegate_req.method, "invalid HTTP method");
-            return error_response(400, "bad_request", &format!("invalid method: {}", e));
-        }
-    };
-
-    let mut upstream_req = http_client.request(method, &delegate_req.url);
-
-    // NOTE: We intentionally do NOT set a per-request timeout here.
-    // reqwest's `.timeout()` caps the *entire* request including body streaming,
-    // which would truncate long-lived SSE streams.  The delegate_client already
-    // has a 30s connect_timeout for connection establishment, and Aether controls
-    // first-byte / idle timeouts on its own side via asyncio.
-
-    // Set headers (skip `host` — reqwest sets it from the URL automatically,
-    // and a duplicate Host header can confuse certain upstreams)
-    for (name, value) in &delegate_req.headers {
-        if name.eq_ignore_ascii_case("host") {
-            continue;
-        }
-        upstream_req = upstream_req.header(name.as_str(), value.as_str());
-    }
-
-    // Set body
-    if let Some(body) = delegate_req.body {
-        upstream_req = upstream_req.body(body);
-    }
-
-    // Send upstream request
-    let upstream_resp = match upstream_req.send().await {
-        Ok(resp) => resp,
-        Err(e) => {
-            warn!(url = %delegate_req.url, error = %e, "delegate upstream request failed");
-            // Sanitize: strip URL details from error message to avoid leaking
-            // API keys or paths that may appear in query strings / paths.
-            let safe_detail = sanitize_upstream_error(&e.to_string());
-            if e.is_timeout() {
-                return error_response(504, "upstream_timeout", &safe_detail);
-            }
-            return error_response(502, "upstream_connection_failed", &safe_detail);
-        }
-    };
-
-    // Build response: pass through upstream status + headers, stream body back
-    let status = upstream_resp.status().as_u16();
-    let upstream_headers = upstream_resp.headers().clone();
-
-    debug!(url = %delegate_req.url, status, "delegate upstream response");
-
-    // Stream the response body
-    let body_stream = upstream_resp
-        .bytes_stream()
-        .map_ok(Frame::data)
-        .map_err(|e| -> Box<dyn std::error::Error + Send + Sync> { Box::new(e) });
-
-    let stream_body: BoxBody = StreamBody::new(body_stream).boxed();
-
-    let mut builder = Response::builder().status(status);
-    for (name, value) in upstream_headers.iter() {
-        builder = builder.header(name, value);
-    }
-
-    builder
-        .body(stream_body)
-        .unwrap_or_else(|_| Response::builder().status(500).body(super::empty_box_body()).unwrap())
-}
-
-// ── Sanitisation ─────────────────────────────────────────────────────────────
-
-/// Strip full URLs from error messages to prevent leaking upstream API keys,
-/// paths, or query parameters in the delegate error response.
-///
-/// Replaces `https://api.example.com/v1/chat?key=xxx` with `api.example.com`.
-fn sanitize_upstream_error(msg: &str) -> String {
-    // Simple regex-free approach: find "https://..." or "http://..." spans and
-    // replace them with just the host portion.
-    let mut result = msg.to_string();
-    for scheme in &["https://", "http://"] {
-        while let Some(start) = result.find(scheme) {
-            let after_scheme = start + scheme.len();
-            // Host ends at '/', '?', '#', ' ', or end of string
-            let host_end = result[after_scheme..]
-                .find(['/', '?', '#', ' '])
-                .map(|i| after_scheme + i)
-                .unwrap_or(result.len());
-            let host = &result[after_scheme..host_end];
-            result = format!("{}{}{}", &result[..start], host, &result[host_end..]);
-        }
-    }
-    result
-}
-
-// ── Error response helpers ───────────────────────────────────────────────────
-
-fn error_response(status: u16, error: &str, detail: &str) -> Response<BoxBody> {
-    let body = serde_json::json!({
-        "error": error,
-        "detail": detail,
-    });
-    let body_bytes = bytes::Bytes::from(body.to_string());
-
-    Response::builder()
-        .status(status)
-        .header("Content-Type", "application/json")
-        .header("X-Delegate-Error", "true")
-        .body(
-            Full::new(body_bytes)
-                .map_err(|e| -> Box<dyn std::error::Error + Send + Sync> { match e {} })
-                .boxed(),
-        )
-        .unwrap()
-}
@@ -1,18 +0,0 @@
-pub mod connect;
-pub mod delegate;
-pub mod server;
-pub mod target_filter;
-pub mod tls;
-
-use http_body_util::BodyExt;
-
-/// Boxed body type used across proxy handlers.
-pub type BoxBody =
-    http_body_util::combinators::BoxBody<bytes::Bytes, Box<dyn std::error::Error + Send + Sync>>;
-
-/// Create an empty [`BoxBody`] (for error responses, 405, etc.).
-pub fn empty_box_body() -> BoxBody {
-    http_body_util::Full::new(bytes::Bytes::new())
-        .map_err(|e| -> Box<dyn std::error::Error + Send + Sync> { match e {} })
-        .boxed()
-}
@@ -1,167 +0,0 @@
-use std::net::SocketAddr;
-use std::sync::atomic::Ordering;
-use std::sync::Arc;
-
-use http_body_util::BodyExt;
-use hyper::body::Incoming;
-use hyper::rt::{Read, Write};
-use hyper::server::conn::http1;
-use hyper::service::service_fn;
-use hyper::{Method, Request, Response};
-use hyper_util::rt::TokioIo;
-use tokio::net::TcpListener;
-use tokio::sync::watch;
-use tracing::{debug, info, warn};
-
-use crate::proxy::{connect, delegate, tls, BoxBody};
-use crate::state::AppState;
-
-/// Start the proxy server.
-///
-/// Listens for incoming TCP connections and dispatches:
-/// - CONNECT requests -> tunnel handler
-/// - POST /_aether/delegate -> delegate handler
-/// - Other requests -> 405 Method Not Allowed
-///
-/// When TLS is configured, the server operates in dual-stack mode:
-/// it peeks at the first byte of each connection to distinguish TLS ClientHello
-/// (0x16) from plain HTTP, and handles both on the same port.
-pub async fn run(
-    state: &Arc<AppState>,
-    mut shutdown_rx: watch::Receiver<bool>,
-) -> anyhow::Result<()> {
-    let addr = SocketAddr::from(([0, 0, 0, 0], state.config.listen_port));
-    let listener = TcpListener::bind(addr).await?;
-
-    if state.tls_acceptor.is_some() {
-        info!(addr = %addr, "proxy server listening (HTTP+TLS dual-stack)");
-    } else {
-        info!(addr = %addr, "proxy server listening (HTTP only)");
-    }
-
-    loop {
-        tokio::select! {
-            result = listener.accept() => {
-                let (stream, peer_addr) = match result {
-                    Ok(v) => v,
-                    Err(e) => {
-                        warn!(error = %e, "failed to accept connection");
-                        continue;
-                    }
-                };
-
-                debug!(peer = %peer_addr, "new connection");
-
-                let state = Arc::clone(state);
-                state.active_connections.fetch_add(1, Ordering::Relaxed);
-
-                tokio::task::spawn(async move {
-                    // Dual-stack: peek first byte to decide TLS vs plain HTTP
-                    if let Some(ref acceptor) = state.tls_acceptor {
-                        if tls::is_tls_client_hello(&stream).await {
-                            match acceptor.clone().accept(stream).await {
-                                Ok(tls_stream) => {
-                                    debug!(peer = %peer_addr, "TLS handshake ok");
-                                    serve_connection(
-                                        TokioIo::new(tls_stream),
-                                        peer_addr,
-                                        &state,
-                                    )
-                                    .await;
-                                }
-                                Err(e) => {
-                                    debug!(peer = %peer_addr, error = %e, "TLS handshake failed");
-                                }
-                            }
-                            state.active_connections.fetch_sub(1, Ordering::Relaxed);
-                            return;
-                        }
-                    }
-
-                    // Plain HTTP
-                    serve_connection(
-                        TokioIo::new(stream),
-                        peer_addr,
-                        &state,
-                    )
-                    .await;
-
-                    state.active_connections.fetch_sub(1, Ordering::Relaxed);
-                });
-            }
-            _ = shutdown_rx.changed() => {
-                info!("proxy server shutting down");
-                break;
-            }
-        }
-    }
-
-    Ok(())
-}
-
-/// Serve a single HTTP/1.1 connection (works over both plain TCP and TLS).
-async fn serve_connection<I>(io: I, peer_addr: SocketAddr, state: &Arc<AppState>)
-where
-    I: Read + Write + Unpin + Send + 'static,
-{
-    let config = Arc::clone(&state.config);
-    let dynamic = Arc::clone(&state.dynamic);
-    let delegate_client = state.delegate_client.clone();
-
-    let service = service_fn(move |req: Request<Incoming>| {
-        let config = Arc::clone(&config);
-        let dynamic = Arc::clone(&dynamic);
-        let delegate_client = delegate_client.clone();
-
-        async move {
-            // Snapshot current dynamic values (may be updated by remote config)
-            let (allowed_ports, timestamp_tolerance) = {
-                let d = dynamic.read().unwrap();
-                (d.allowed_ports.clone(), d.timestamp_tolerance)
-            };
-
-            if req.method() == Method::CONNECT {
-                let resp =
-                    connect::handle_connect(req, config, &allowed_ports, timestamp_tolerance).await;
-                let resp = resp.map(|_| -> BoxBody {
-                    http_body_util::Empty::new()
-                        .map_err(|e| -> Box<dyn std::error::Error + Send + Sync> { match e {} })
-                        .boxed()
-                });
-                Ok::<_, hyper::Error>(resp)
-            } else if req.uri().path() == "/_aether/delegate" && req.method() == hyper::Method::POST
-            {
-                let resp = delegate::handle_delegate(
-                    req,
-                    config,
-                    &allowed_ports,
-                    timestamp_tolerance,
-                    &delegate_client,
-                )
-                .await;
-                Ok(resp)
-            } else {
-                // Only CONNECT tunnels and /_aether/delegate are supported;
-                // plain HTTP forward proxy was removed (all API traffic is HTTPS).
-                Ok(Response::builder()
-                    .status(405)
-                    .header("Allow", "CONNECT")
-                    .header("Content-Length", "0")
-                    .body(crate::proxy::empty_box_body())
-                    .unwrap())
-            }
-        }
-    });
-
-    if let Err(e) = http1::Builder::new()
-        .preserve_header_case(true)
-        .title_case_headers(false)
-        .serve_connection(io, service)
-        .with_upgrades()
-        .await
-    {
-        if !e.to_string().contains("connection closed") {
-            debug!(peer = %peer_addr, error = %e, "connection error");
-        }
-    }
-}
@@ -1,180 +0,0 @@
-use std::collections::HashSet;
-use std::net::{IpAddr, Ipv4Addr, Ipv6Addr, SocketAddr};
-
-/// Check if an IP address belongs to a private/reserved network.
-fn is_private_ip(ip: &IpAddr) -> bool {
-    match ip {
-        IpAddr::V4(v4) => is_private_ipv4(v4),
-        IpAddr::V6(v6) => is_private_ipv6(v6),
-    }
-}
-
-fn is_private_ipv4(ip: &Ipv4Addr) -> bool {
-    let octets = ip.octets();
-    // 10.0.0.0/8
-    if octets[0] == 10 {
-        return true;
-    }
-    // 172.16.0.0/12
-    if octets[0] == 172 && (16..=31).contains(&octets[1]) {
-        return true;
-    }
-    // 192.168.0.0/16
-    if octets[0] == 192 && octets[1] == 168 {
-        return true;
-    }
-    // 127.0.0.0/8
-    if octets[0] == 127 {
-        return true;
-    }
-    // 169.254.0.0/16 (link-local)
-    if octets[0] == 169 && octets[1] == 254 {
-        return true;
-    }
-    // 0.0.0.0/8
-    if octets[0] == 0 {
-        return true;
-    }
-    false
-}
-
-fn is_private_ipv6(ip: &Ipv6Addr) -> bool {
-    // ::1 loopback
-    if ip.is_loopback() {
-        return true;
-    }
-    // :: unspecified
-    if ip.is_unspecified() {
-        return true;
-    }
-    let segments = ip.segments();
-    // fc00::/7 (ULA) - first byte is 0xfc or 0xfd
-    if segments[0] & 0xfe00 == 0xfc00 {
-        return true;
-    }
-    // fe80::/10 (link-local)
-    if segments[0] & 0xffc0 == 0xfe80 {
-        return true;
-    }
-    // IPv4-mapped IPv6 (::ffff:x.x.x.x) - check the embedded IPv4
-    if let Some(v4) = ip.to_ipv4_mapped() {
-        return is_private_ipv4(&v4);
-    }
-    false
-}
-
-#[derive(Debug)]
-pub enum FilterError {
-    PrivateIp(IpAddr),
-    PortNotAllowed(u16),
-    DnsResolutionFailed(String),
-}
-
-impl std::fmt::Display for FilterError {
-    fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
-        match self {
-            Self::PrivateIp(ip) => write!(f, "target IP {} is in private/reserved range", ip),
-            Self::PortNotAllowed(port) => write!(f, "port {} not in allowed list", port),
-            Self::DnsResolutionFailed(host) => write!(f, "DNS resolution failed for {}", host),
-        }
-    }
-}
-
-/// Validate that the target host:port is allowed.
-///
-/// Uses async DNS resolution (via `tokio::net::lookup_host`) to avoid
-/// blocking the async runtime on potentially slow DNS lookups.
-///
-/// Returns the resolved socket address to connect to.
-pub async fn validate_target(
-    host: &str,
-    port: u16,
-    allowed_ports: &HashSet<u16>,
-) -> Result<SocketAddr, FilterError> {
-    // Port whitelist check
-    if !allowed_ports.contains(&port) {
-        return Err(FilterError::PortNotAllowed(port));
-    }
-
-    // Try parsing as IP directly (no DNS needed)
-    if let Ok(ip) = host.parse::<IpAddr>() {
-        if is_private_ip(&ip) {
-            return Err(FilterError::PrivateIp(ip));
-        }
-        return Ok(SocketAddr::new(ip, port));
-    }
-
-    // Async DNS resolution with private IP check (DNS rebinding protection)
-    let addr_str = format!("{}:{}", host, port);
-    let addrs: Vec<SocketAddr> = tokio::net::lookup_host(&addr_str)
-        .await
-        .map_err(|_| FilterError::DnsResolutionFailed(host.to_string()))?
-        .collect();
-
-    if addrs.is_empty() {
-        return Err(FilterError::DnsResolutionFailed(host.to_string()));
-    }
-
-    // All resolved addresses must be non-private
-    for addr in &addrs {
-        if is_private_ip(&addr.ip()) {
-            return Err(FilterError::PrivateIp(addr.ip()));
-        }
-    }
-
-    // Return the first valid address
-    Ok(addrs[0])
-}
-
-#[cfg(test)]
-mod tests {
-    use super::*;
-
-    fn ports() -> HashSet<u16> {
-        [80, 443, 8080, 8443].into_iter().collect()
-    }
-
-    #[test]
-    fn test_private_ipv4() {
-        assert!(is_private_ip(&IpAddr::V4(Ipv4Addr::new(10, 0, 0, 1))));
-        assert!(is_private_ip(&IpAddr::V4(Ipv4Addr::new(172, 16, 0, 1))));
-        assert!(is_private_ip(&IpAddr::V4(Ipv4Addr::new(192, 168, 1, 1))));
-        assert!(is_private_ip(&IpAddr::V4(Ipv4Addr::new(127, 0, 0, 1))));
-        assert!(is_private_ip(&IpAddr::V4(Ipv4Addr::new(169, 254, 1, 1))));
-        assert!(is_private_ip(&IpAddr::V4(Ipv4Addr::new(0, 0, 0, 0))));
-        assert!(!is_private_ip(&IpAddr::V4(Ipv4Addr::new(8, 8, 8, 8))));
-        assert!(!is_private_ip(&IpAddr::V4(Ipv4Addr::new(203, 0, 113, 1))));
-    }
-
-    #[test]
-    fn test_private_ipv6() {
-        assert!(is_private_ip(&IpAddr::V6(Ipv6Addr::LOCALHOST)));
-        assert!(is_private_ip(&IpAddr::V6(Ipv6Addr::UNSPECIFIED)));
-        // fc00::1 (ULA)
-        assert!(is_private_ip(&IpAddr::V6(Ipv6Addr::new(
-            0xfc00, 0, 0, 0, 0, 0, 0, 1
-        ))));
-        // fe80::1 (link-local)
-        assert!(is_private_ip(&IpAddr::V6(Ipv6Addr::new(
-            0xfe80, 0, 0, 0, 0, 0, 0, 1
-        ))));
-    }
-
-    #[tokio::test]
-    async fn test_port_not_allowed() {
-        let result = validate_target("8.8.8.8", 22, &ports()).await;
-        assert!(matches!(result, Err(FilterError::PortNotAllowed(22))));
-    }
-
-    #[tokio::test]
-    async fn test_private_ip_blocked() {
-        let result = validate_target("127.0.0.1", 80, &ports()).await;
-        assert!(matches!(result, Err(FilterError::PrivateIp(_))));
-    }
-
-    #[tokio::test]
-    async fn test_public_ip_allowed() {
-        let result = validate_target("8.8.8.8", 443, &ports()).await;
-        assert!(result.is_ok());
-    }
-}
@@ -1,112 +0,0 @@
-use std::fs;
-use std::io::BufReader;
-use std::path::Path;
-use std::sync::Arc;
-
-use rcgen::{CertificateParams, KeyPair};
-use rustls_pki_types::{CertificateDer, PrivateKeyDer};
-use sha2::{Digest, Sha256};
-use tokio_rustls::TlsAcceptor;
-use tracing::{info, warn};
-
-/// Generate a self-signed certificate if the files do not already exist.
-///
-/// The certificate includes SANs: `localhost` and `aether-proxy`.
-/// The private key file is set to mode 0600 on unix.
-pub fn ensure_self_signed_cert(cert_path: &Path, key_path: &Path) -> anyhow::Result<()> {
-    if cert_path.exists() && key_path.exists() {
-        info!(
-            cert = %cert_path.display(),
-            key = %key_path.display(),
-            "using existing TLS certificate"
-        );
-        return Ok(());
-    }
-
-    info!("generating self-signed TLS certificate");
-
-    let mut params = CertificateParams::new(vec!["localhost".into(), "aether-proxy".into()])?;
-    params.distinguished_name = rcgen::DistinguishedName::new();
-    params
-        .distinguished_name
-        .push(rcgen::DnType::CommonName, "aether-proxy");
-
-    let key_pair = KeyPair::generate()?;
-    let cert = params.self_signed(&key_pair)?;
-
-    let cert_pem = cert.pem();
-    let key_pem = key_pair.serialize_pem();
-
-    fs::write(cert_path, &cert_pem)?;
-    fs::write(key_path, &key_pem)?;
-
-    // Set key file permissions to 0600 on unix
-    #[cfg(unix)]
-    {
-        use std::os::unix::fs::PermissionsExt;
-        let perms = fs::Permissions::from_mode(0o600);
-        fs::set_permissions(key_path, perms)?;
-    }
-
-    info!(
-        cert = %cert_path.display(),
-        key = %key_path.display(),
-        "self-signed TLS certificate generated"
-    );
-
-    Ok(())
-}
-
-/// Build a `TlsAcceptor` from PEM certificate and key files.
-pub fn build_tls_acceptor(cert_path: &Path, key_path: &Path) -> anyhow::Result<TlsAcceptor> {
-    let cert_file = fs::File::open(cert_path)?;
-    let key_file = fs::File::open(key_path)?;
-
-    let certs: Vec<CertificateDer<'static>> =
-        rustls_pemfile::certs(&mut BufReader::new(cert_file)).collect::<Result<Vec<_>, _>>()?;
-
-    if certs.is_empty() {
-        anyhow::bail!("no certificates found in {}", cert_path.display());
-    }
-
-    let key: PrivateKeyDer<'static> =
-        rustls_pemfile::private_key(&mut BufReader::new(key_file))?
-            .ok_or_else(|| anyhow::anyhow!("no private key found in {}", key_path.display()))?;
-
-    let config = rustls::ServerConfig::builder()
-        .with_no_client_auth()
-        .with_single_cert(certs, key)?;
-
-    Ok(TlsAcceptor::from(Arc::new(config)))
-}
-
-/// Compute the SHA-256 fingerprint of the first certificate in a PEM file.
-///
-/// Returns the hex-encoded fingerprint (lowercase, no separators).
-pub fn cert_sha256_fingerprint(cert_path: &Path) -> anyhow::Result<String> {
-    let cert_file = fs::File::open(cert_path)?;
-    let certs: Vec<CertificateDer<'static>> =
-        rustls_pemfile::certs(&mut BufReader::new(cert_file)).collect::<Result<Vec<_>, _>>()?;
-
-    let cert = certs
-        .first()
-        .ok_or_else(|| anyhow::anyhow!("no certificates found in {}", cert_path.display()))?;
-
-    let digest = Sha256::digest(cert.as_ref());
-    Ok(hex::encode(digest))
-}
-
-/// Peek at the first byte of a TCP stream to determine if it is a TLS ClientHello.
-///
-/// Returns `true` if the first byte is 0x16 (TLS record type: Handshake).
-pub async fn is_tls_client_hello(stream: &tokio::net::TcpStream) -> bool {
-    let mut buf = [0u8; 1];
-    match stream.peek(&mut buf).await {
-        Ok(1) => buf[0] == 0x16,
-        Ok(_) => false,
-        Err(e) => {
-            warn!(error = %e, "failed to peek first byte");
-            false
-        }
-    }
-}
@@ -1,29 +1,13 @@
+use std::time::{Duration, SystemTime, UNIX_EPOCH};
+
 use reqwest::{Client, StatusCode};
 use serde::{Deserialize, Serialize};
-use tracing::{debug, error, info, warn};
+use tokio::time::sleep;
+use tracing::{debug, error, info};

 use crate::config::Config;
 use crate::hardware::HardwareInfo;

-/// Heartbeat-specific error that distinguishes "node not found" (needs
-/// re-registration) from transient / other failures.
-#[derive(Debug)]
-pub enum HeartbeatError {
-    /// HTTP 404 – the node_id is no longer known to Aether.
-    NodeNotFound(String),
-    /// Any other failure (network, 5xx, etc.).
-    Other(anyhow::Error),
-}
-
-impl std::fmt::Display for HeartbeatError {
-    fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
-        match self {
-            Self::NodeNotFound(msg) => write!(f, "node not found: {}", msg),
-            Self::Other(e) => write!(f, "{}", e),
-        }
-    }
-}
-
 #[derive(Debug, Serialize)]
 struct RegisterRequest {
    name: String,
@@ -32,14 +16,13 @@ struct RegisterRequest {
    #[serde(skip_serializing_if = "Option::is_none")]
    region: Option<String>,
    heartbeat_interval: u64,
-    #[serde(skip_serializing_if = "std::ops::Not::not")]
-    tls_enabled: bool,
-    #[serde(skip_serializing_if = "Option::is_none")]
-    tls_cert_fingerprint: Option<String>,
    #[serde(skip_serializing_if = "Option::is_none")]
    hardware_info: Option<serde_json::Value>,
    #[serde(skip_serializing_if = "Option::is_none")]
    estimated_max_concurrency: Option<u64>,
+    #[serde(skip_serializing_if = "Option::is_none")]
+    proxy_metadata: Option<serde_json::Value>,
+    tunnel_mode: bool,
 }

 #[derive(Debug, Deserialize)]
@@ -47,17 +30,6 @@ pub struct RegisterResponse {
    pub node_id: String,
 }

-#[derive(Debug, Serialize)]
-struct HeartbeatRequest {
-    node_id: String,
-    #[serde(skip_serializing_if = "Option::is_none")]
-    active_connections: Option<i64>,
-    #[serde(skip_serializing_if = "Option::is_none")]
-    total_requests: Option<i64>,
-    #[serde(skip_serializing_if = "Option::is_none")]
-    avg_latency_ms: Option<f64>,
-}
-
 /// Remote configuration pushed by the Aether management backend.
 #[derive(Debug, Clone, Deserialize)]
 pub struct RemoteConfig {
@@ -65,29 +37,6 @@ pub struct RemoteConfig {
    pub allowed_ports: Option<Vec<u16>>,
    pub log_level: Option<String>,
    pub heartbeat_interval: Option<u64>,
-    pub timestamp_tolerance: Option<u64>,
-}
-
-/// Parsed heartbeat response from Aether.
-#[derive(Debug, Deserialize)]
-struct HeartbeatResponseBody {
-    #[serde(default)]
-    node: Option<HeartbeatNodeInfo>,
-}
-
-#[derive(Debug, Deserialize)]
-struct HeartbeatNodeInfo {
-    #[serde(default)]
-    remote_config: Option<RemoteConfig>,
-    #[serde(default)]
-    config_version: Option<u64>,
-}
-
-/// Heartbeat result returned to the caller.
-#[derive(Debug)]
-pub struct HeartbeatResult {
-    pub remote_config: Option<RemoteConfig>,
-    pub config_version: u64,
 }

 #[derive(Debug, Serialize)]
@@ -100,19 +49,44 @@ pub struct AetherClient {
    http: Client,
    base_url: String,
    token: String,
+    retry_max_attempts: u32,
+    retry_base_delay: Duration,
+    retry_max_delay: Duration,
 }

 impl AetherClient {
-    pub fn new(config: &Config) -> Self {
-        let http = Client::builder()
-            .timeout(std::time::Duration::from_secs(10))
-            .build()
-            .expect("failed to create HTTP client");
+    pub fn new(config: &Config, aether_url: &str, management_token: &str) -> Self {
+        let mut builder = Client::builder()
+            .timeout(Duration::from_secs(config.aether_request_timeout_secs))
+            .connect_timeout(Duration::from_secs(config.aether_connect_timeout_secs))
+            .pool_max_idle_per_host(config.aether_pool_max_idle_per_host)
+            .pool_idle_timeout(Duration::from_secs(config.aether_pool_idle_timeout_secs))
+            .tcp_nodelay(config.aether_tcp_nodelay);
+
+        if config.aether_tcp_keepalive_secs > 0 {
+            builder =
+                builder.tcp_keepalive(Some(Duration::from_secs(config.aether_tcp_keepalive_secs)));
+        } else {
+            builder = builder.tcp_keepalive(None);
+        }
+
+        if config.aether_http2 {
+            builder = builder.http2_adaptive_window(true);
+        }
+
+        let http = builder.build().expect("failed to create HTTP client");
+
+        let retry_base_delay = Duration::from_millis(config.aether_retry_base_delay_ms);
+        let retry_max_delay =
+            Duration::from_millis(config.aether_retry_max_delay_ms).max(retry_base_delay);

        Self {
            http,
-            base_url: config.aether_url.trim_end_matches('/').to_string(),
-            token: config.management_token.clone(),
+            base_url: aether_url.trim_end_matches('/').to_string(),
+            token: management_token.to_string(),
+            retry_max_attempts: config.aether_retry_max_attempts.max(1),
+            retry_base_delay,
+            retry_max_delay,
        }
    }

@@ -122,38 +96,42 @@ impl AetherClient {
    pub async fn register(
        &self,
        config: &Config,
+        node_name: &str,
        public_ip: &str,
-        tls_enabled: bool,
-        tls_cert_fingerprint: Option<&str>,
        hw: Option<&HardwareInfo>,
    ) -> anyhow::Result<String> {
        let url = format!("{}/api/admin/proxy-nodes/register", self.base_url);
        let body = RegisterRequest {
-            name: config.node_name.clone(),
+            name: node_name.to_string(),
            ip: public_ip.to_string(),
-            port: config.listen_port,
+            port: 0,
            region: config.node_region.clone(),
            heartbeat_interval: config.heartbeat_interval,
-            tls_enabled,
-            tls_cert_fingerprint: tls_cert_fingerprint.map(|s| s.to_string()),
            hardware_info: hw.and_then(|h| serde_json::to_value(h).ok()),
            estimated_max_concurrency: hw.map(|h| h.estimated_max_concurrency),
+            proxy_metadata: Some(serde_json::json!({
+                "version": env!("CARGO_PKG_VERSION"),
+            })),
+            tunnel_mode: true,
        };

        info!(
            url = %url,
            name = %body.name,
            ip = %body.ip,
-            port = body.port,
            "registering with Aether"
        );

        let resp = self
-            .http
-            .post(&url)
-            .header("Authorization", format!("Bearer {}", self.token))
-            .json(&body)
-            .send()
+            .send_with_retry(
+                || {
+                    self.http
+                        .post(&url)
+                        .header("Authorization", format!("Bearer {}", self.token))
+                        .json(&body)
+                },
+                "register",
+            )
            .await?;

        let status = resp.status();
@@ -167,76 +145,6 @@ impl AetherClient {
        Ok(data.node_id)
    }

-    /// Send heartbeat to Aether.
-    ///
-    /// On success, returns any remote config included in the response.
-    /// Returns [`HeartbeatError::NodeNotFound`] on HTTP 404 so the caller
-    /// can trigger re-registration.
-    pub async fn heartbeat(
-        &self,
-        node_id: &str,
-        active_connections: Option<i64>,
-        total_requests: Option<i64>,
-        avg_latency_ms: Option<f64>,
-    ) -> Result<HeartbeatResult, HeartbeatError> {
-        let url = format!("{}/api/admin/proxy-nodes/heartbeat", self.base_url);
-        let body = HeartbeatRequest {
-            node_id: node_id.to_string(),
-            active_connections,
-            total_requests,
-            avg_latency_ms,
-        };
-
-        debug!(node_id = %node_id, "sending heartbeat");
-
-        let resp = self
-            .http
-            .post(&url)
-            .header("Authorization", format!("Bearer {}", self.token))
-            .json(&body)
-            .send()
-            .await
-            .map_err(|e| HeartbeatError::Other(e.into()))?;
-
-        let status = resp.status();
-        if !status.is_success() {
-            let text = resp.text().await.unwrap_or_default();
-            warn!(status = %status, body = %text, "heartbeat failed");
-            if status == StatusCode::NOT_FOUND {
-                return Err(HeartbeatError::NodeNotFound(text));
-            }
-            return Err(HeartbeatError::Other(anyhow::anyhow!(
-                "heartbeat failed (HTTP {}): {}",
-                status,
-                text
-            )));
-        }
-
-        // Parse remote config from response (best-effort)
-        let result = match resp.json::<HeartbeatResponseBody>().await {
-            Ok(body) => {
-                let (remote_config, config_version) = match body.node {
-                    Some(node) => (node.remote_config, node.config_version.unwrap_or(0)),
-                    None => (None, 0),
-                };
-                HeartbeatResult {
-                    remote_config,
-                    config_version,
-                }
-            }
-            Err(e) => {
-                debug!(error = %e, "failed to parse heartbeat response body");
-                HeartbeatResult {
-                    remote_config: None,
-                    config_version: 0,
-                }
-            }
-        };
-
-        debug!(node_id = %node_id, config_version = result.config_version, "heartbeat ok");
-        Ok(result)
-    }
-
    /// Unregister this node from Aether (graceful shutdown).
    pub async fn unregister(&self, node_id: &str) -> anyhow::Result<()> {
        let url = format!("{}/api/admin/proxy-nodes/unregister", self.base_url);
@@ -247,11 +155,15 @@ impl AetherClient {
        info!(node_id = %node_id, "unregistering from Aether");

        let resp = self
-            .http
-            .post(&url)
-            .header("Authorization", format!("Bearer {}", self.token))
-            .json(&body)
-            .send()
+            .send_with_retry(
+                || {
+                    self.http
+                        .post(&url)
+                        .header("Authorization", format!("Bearer {}", self.token))
+                        .json(&body)
+                },
+                "unregister",
+            )
            .await;

        match resp {
@@ -271,4 +183,75 @@ impl AetherClient {
            }
        }
    }
+
+    async fn send_with_retry<F>(
+        &self,
+        mut make_req: F,
+        label: &str,
+    ) -> Result<reqwest::Response, reqwest::Error>
+    where
+        F: FnMut() -> reqwest::RequestBuilder,
+    {
+        let mut attempt: u32 = 0;
+        let mut delay = self.retry_base_delay;
+
+        loop {
+            attempt = attempt.saturating_add(1);
+            let resp = make_req().send().await;
+            match resp {
+                Ok(resp) => {
+                    if should_retry_status(resp.status()) && attempt < self.retry_max_attempts {
+                        let sleep_for = jitter_delay(delay);
+                        debug!(
+                            attempt,
+                            status = %resp.status(),
+                            sleep_ms = sleep_for.as_millis(),
+                            label,
+                            "Aether request retrying"
+                        );
+                        sleep(sleep_for).await;
+                        let next_delay = delay.checked_mul(2).unwrap_or(self.retry_max_delay);
+                        delay = std::cmp::min(next_delay, self.retry_max_delay);
+                        continue;
+                    }
+                    return Ok(resp);
+                }
+                Err(e) => {
+                    if attempt < self.retry_max_attempts {
+                        let sleep_for = jitter_delay(delay);
+                        debug!(
+                            attempt,
+                            error = %e,
+                            sleep_ms = sleep_for.as_millis(),
+                            label,
+                            "Aether request retrying"
+                        );
+                        sleep(sleep_for).await;
+                        let next_delay = delay.checked_mul(2).unwrap_or(self.retry_max_delay);
+                        delay = std::cmp::min(next_delay, self.retry_max_delay);
+                        continue;
+                    }
+                    return Err(e);
+                }
+            }
+        }
+    }
+}
+
+fn should_retry_status(status: StatusCode) -> bool {
+    status.is_server_error()
+        || status == StatusCode::TOO_MANY_REQUESTS
+        || status == StatusCode::REQUEST_TIMEOUT
+}
+
+fn jitter_delay(base: Duration) -> Duration {
+    if base.is_zero() {
+        return base;
+    }
+    let nanos = SystemTime::now()
+        .duration_since(UNIX_EPOCH)
+        .map(|d| d.subsec_nanos() as u64)
+        .unwrap_or(0);
+    let jitter_ms = nanos % 100;
+    base + Duration::from_millis(jitter_ms)
 }
@@ -1,112 +0,0 @@
-use std::sync::atomic::Ordering;
-use std::sync::Arc;
-
-use tokio::sync::watch;
-use tracing::{debug, error, info, warn};
-
-use crate::registration::client::HeartbeatError;
-use crate::runtime;
-use crate::state::AppState;
-
-/// Run periodic heartbeat task until shutdown signal.
-///
-/// When Aether responds with 404 (node not found), this task automatically
-/// re-registers the node and updates the shared `node_id` so the proxy
-/// server and future heartbeats use the new identity.
-///
-/// When the heartbeat response includes a `remote_config`, it is applied
-/// to the [`DynamicConfig`](crate::runtime::DynamicConfig) so the proxy
-/// picks up changes without a restart.
-pub async fn run(state: &Arc<AppState>, mut shutdown_rx: watch::Receiver<bool>) {
-    let mut consecutive_failures: u32 = 0;
-
-    // Skip the first tick (registration already acts as initial heartbeat)
-    let initial_interval = state.dynamic.read().unwrap().heartbeat_interval;
-    tokio::select! {
-        _ = tokio::time::sleep(std::time::Duration::from_secs(initial_interval)) => {}
-        _ = shutdown_rx.changed() => {
-            debug!("heartbeat task stopping (during initial wait)");
-            return;
-        }
-    }
-
-    loop {
-        let current_node_id = state.node_id.read().unwrap().clone();
-        let active_conns = state.active_connections.load(Ordering::Relaxed) as i64;
-
-        match state
-            .aether_client
-            .heartbeat(&current_node_id, Some(active_conns), None, None)
-            .await
-        {
-            Ok(result) => {
-                if consecutive_failures > 0 {
-                    info!(
-                        previous_failures = consecutive_failures,
-                        "heartbeat recovered"
-                    );
-                }
-                consecutive_failures = 0;
-
-                // Apply remote config if present and version changed
-                if let Some(ref remote) = result.remote_config {
-                    runtime::apply_remote_config(&state.dynamic, remote, result.config_version);
-                }
-            }
-            Err(HeartbeatError::NodeNotFound(_)) => {
-                warn!(
-                    old_node_id = %current_node_id,
-                    "node not found, re-registering"
-                );
-                match state
-                    .aether_client
-                    .register(
-                        &state.config,
-                        &state.public_ip,
-                        state.config.enable_tls,
-                        state.tls_fingerprint.as_deref(),
-                        Some(&state.hardware_info),
-                    )
-                    .await
-                {
-                    Ok(new_id) => {
-                        info!(
-                            old_node_id = %current_node_id,
-                            new_node_id = %new_id,
-                            "re-registered successfully"
-                        );
-                        *state.node_id.write().unwrap() = new_id;
-                        consecutive_failures = 0;
-                    }
-                    Err(e) => {
-                        consecutive_failures += 1;
-                        error!(
-                            error = %e,
-                            consecutive_failures,
-                            "re-registration failed"
-                        );
-                    }
-                }
-            }
-            Err(HeartbeatError::Other(e)) => {
-                consecutive_failures += 1;
-                warn!(
-                    error = %e,
-                    consecutive_failures,
-                    "heartbeat failed"
-                );
-            }
-        }
-
-        // Read interval from dynamic config (may have been updated remotely)
-        let interval_secs = state.dynamic.read().unwrap().heartbeat_interval;
-
-        tokio::select! {
-            _ = tokio::time::sleep(std::time::Duration::from_secs(interval_secs)) => {}
-            _ = shutdown_rx.changed() => {
-                debug!("heartbeat task stopping");
-                break;
-            }
-        }
-    }
-}
@@ -1,2 +1 @@
 pub mod client;
-pub mod heartbeat;
@@ -5,18 +5,18 @@
 //! management backend through the heartbeat response.

 use std::collections::HashSet;
-use std::sync::{Arc, OnceLock, RwLock};
+use std::sync::{Arc, OnceLock};

+use arc_swap::ArcSwap;
 use tracing::info;

 use crate::config::Config;

 /// Configuration that can be changed at runtime without restart.
-#[derive(Debug)]
+#[derive(Debug, Clone)]
 pub struct DynamicConfig {
    pub node_name: String,
-    pub allowed_ports: HashSet<u16>,
-    pub timestamp_tolerance: u64,
+    pub allowed_ports: Arc<HashSet<u16>>,
    pub log_level: String,
    pub heartbeat_interval: u64,
    /// Monotonically increasing version from the backend.
@@ -29,8 +29,7 @@ impl DynamicConfig {
    pub fn from_config(config: &Config) -> Self {
        Self {
            node_name: config.node_name.clone(),
-            allowed_ports: config.allowed_ports.iter().copied().collect(),
-            timestamp_tolerance: config.timestamp_tolerance,
+            allowed_ports: Arc::new(config.allowed_ports.iter().copied().collect()),
            log_level: config.log_level.clone(),
            heartbeat_interval: config.heartbeat_interval,
            config_version: 0,
@@ -38,10 +37,10 @@ impl DynamicConfig {
    }
 }

-/// Shared dynamic config handle.
-pub type SharedDynamicConfig = Arc<RwLock<DynamicConfig>>;
+/// Shared dynamic config handle (lock-free reads via ArcSwap).
+pub type SharedDynamicConfig = Arc<ArcSwap<DynamicConfig>>;

-// ── Log-level hot-reload ─────────────────────────────────────────────────────
+// -- Log-level hot-reload -----

 /// Global log-level reloader function, set during tracing init.
 type LogReloader = Box<dyn Fn(&str) + Send + Sync>;
@@ -55,53 +54,50 @@ pub fn set_log_reloader(f: LogReloader) {

 /// Apply a remote config update to the dynamic config.
 ///
+/// Uses copy-on-write: loads the current snapshot, clones it, applies changes,
+/// and stores the new Arc. Reads are always lock-free.
+///
 /// Returns `true` if the config was actually changed.
 pub fn apply_remote_config(
    dynamic: &SharedDynamicConfig,
    remote: &crate::registration::client::RemoteConfig,
    version: u64,
 ) -> bool {
-    let mut cfg = dynamic.write().unwrap();
+    let current = dynamic.load();

-    if version <= cfg.config_version {
+    if version <= current.config_version {
        return false;
    }

+    let mut new_cfg = (**current).clone();
    let mut changed = Vec::new();

    if let Some(ref name) = remote.node_name {
-        if *name != cfg.node_name {
-            changed.push(format!("node_name → {}", name));
-            cfg.node_name = name.clone();
+        if *name != new_cfg.node_name {
+            changed.push(format!("node_name -> {}", name));
+            new_cfg.node_name = name.clone();
        }
    }

    if let Some(ref ports) = remote.allowed_ports {
        let new_set: HashSet<u16> = ports.iter().copied().collect();
-        if new_set != cfg.allowed_ports {
-            changed.push(format!("allowed_ports → {:?}", ports));
-            cfg.allowed_ports = new_set;
-        }
-    }
-
-    if let Some(tol) = remote.timestamp_tolerance {
-        if tol != cfg.timestamp_tolerance {
-            changed.push(format!("timestamp_tolerance → {}", tol));
-            cfg.timestamp_tolerance = tol;
+        if new_set != *new_cfg.allowed_ports {
+            changed.push(format!("allowed_ports -> {:?}", ports));
+            new_cfg.allowed_ports = Arc::new(new_set);
        }
    }

    if let Some(interval) = remote.heartbeat_interval {
-        if interval != cfg.heartbeat_interval {
-            changed.push(format!("heartbeat_interval → {}s", interval));
-            cfg.heartbeat_interval = interval;
+        if interval != new_cfg.heartbeat_interval {
+            changed.push(format!("heartbeat_interval -> {}s", interval));
+            new_cfg.heartbeat_interval = interval;
        }
    }

    if let Some(ref level) = remote.log_level {
-        if *level != cfg.log_level {
-            changed.push(format!("log_level → {}", level));
-            cfg.log_level = level.clone();
+        if *level != new_cfg.log_level {
+            changed.push(format!("log_level -> {}", level));
+            new_cfg.log_level = level.clone();
            // Hot-reload tracing filter
            if let Some(reloader) = LOG_RELOADER.get() {
                reloader(level);
@@ -109,15 +105,17 @@ pub fn apply_remote_config(
        }
    }

-    cfg.config_version = version;
+    let has_changes = !changed.is_empty();

-    if !changed.is_empty() {
+    if has_changes {
+        new_cfg.config_version = version;
        info!(
            version,
            changes = %changed.join(", "),
            "remote config applied"
        );
+        dynamic.store(Arc::new(new_cfg));
    }

-    !changed.is_empty()
+    has_changes
 }
@@ -0,0 +1,66 @@
+//! Safe DNS resolver for reqwest that reuses validated addresses from DnsCache.
+//!
+//! This resolver ensures reqwest connects only to addresses that have been
+//! previously validated by `target_filter::validate_target()`, eliminating
+//! the TOCTTOU gap where DNS rebinding could redirect traffic to private IPs.
+
+use std::net::SocketAddr;
+use std::sync::Arc;
+
+use reqwest::dns::{Addrs, Name, Resolve, Resolving};
+
+use crate::target_filter::{self, DnsCache};
+
+/// A DNS resolver that serves validated public addresses from the shared DnsCache.
+///
+/// When reqwest needs to resolve a hostname, this resolver returns addresses
+/// from the cache (populated by `validate_target()` during request validation).
+/// If the hostname is not in cache (shouldn't happen in normal flow), it
+/// performs a fresh resolution with private-IP filtering.
+pub struct SafeDnsResolver {
+    dns_cache: Arc<DnsCache>,
+}
+
+impl SafeDnsResolver {
+    pub fn new(dns_cache: Arc<DnsCache>) -> Self {
+        Self { dns_cache }
+    }
+}
+
+impl Resolve for SafeDnsResolver {
+    fn resolve(&self, name: Name) -> Resolving {
+        let dns_cache = Arc::clone(&self.dns_cache);
+        Box::pin(async move {
+            let host = name.as_str();
+
+            // Try cache first (should be populated by validate_target).
+            // reqwest resolves by hostname only (no port), so use host-only lookup.
+            if let Some(addrs) = dns_cache.get_by_host(host).await {
+                let socket_addrs: Vec<SocketAddr> = (*addrs).clone();
+                return Ok(Box::new(socket_addrs.into_iter()) as Addrs);
+            }
+
+            // Fallback: resolve with private-IP filtering (defensive).
+            // This path should rarely be hit since validate_target() runs first.
+            // We don't know the real port here (reqwest Resolve only gives hostname),
+            // so resolve directly without caching to avoid polluting the cache with
+            // an incorrect port-based key.
+            let addr_str = format!("{}:0", host);
+            let resolved: Vec<SocketAddr> = tokio::net::lookup_host(&addr_str)
+                .await
+                .map_err(|e| -> Box<dyn std::error::Error + Send + Sync> { Box::new(e) })?
+                .filter(|addr| !target_filter::is_private_ip(&addr.ip()))
+                .collect();
+
+            if resolved.is_empty() {
+                return Err(Box::new(std::io::Error::other(format!(
+                    "all resolved addresses for {} are private/reserved",
+                    host
+                )))
+                    as Box<dyn std::error::Error + Send + Sync>);
+            }
+
+            Ok(Box::new(resolved.into_iter()) as Addrs)
+        })
+    }
+}
@@ -21,7 +21,7 @@ pub fn install_service(config_path: &Path) -> anyhow::Result<()> {
        anyhow::bail!("systemd not available");
    }
    if !is_root() {
-        anyhow::bail!("root required, use: sudo aether-proxy setup");
+        anyhow::bail!("root required, use: sudo ./aether-proxy setup");
    }

    let exe_path = std::env::current_exe()?.canonicalize()?;
@@ -67,6 +67,7 @@ pub fn install_service(config_path: &Path) -> anyhow::Result<()> {
         Restart=on-failure\n\
         RestartSec=5\n\
         LimitNOFILE=65535\n\
+         UMask=0077\n\
         \n\
         [Install]\n\
         WantedBy=multi-user.target\n",
@@ -93,11 +94,11 @@ pub fn install_service(config_path: &Path) -> anyhow::Result<()> {

    eprintln!();
    eprintln!("  Commands:");
-    eprintln!("    aether-proxy status          # service status");
-    eprintln!("    aether-proxy logs            # tail logs");
-    eprintln!("    sudo aether-proxy restart    # restart");
-    eprintln!("    sudo aether-proxy stop       # stop");
-    eprintln!("    sudo aether-proxy uninstall  # remove service");
+    eprintln!("    ./aether-proxy status          # service status");
+    eprintln!("    ./aether-proxy logs            # tail logs");
+    eprintln!("    sudo ./aether-proxy restart    # restart");
+    eprintln!("    sudo ./aether-proxy stop       # stop");
+    eprintln!("    sudo ./aether-proxy uninstall  # remove service");
    eprintln!();

    Ok(())
@@ -165,7 +166,7 @@ pub fn is_service_active() -> bool {

 fn ensure_service_installed() -> anyhow::Result<()> {
    if !std::path::Path::new(UNIT_PATH).exists() {
-        anyhow::bail!("service not installed, run `sudo aether-proxy setup` first");
+        anyhow::bail!("service not installed, run `sudo ./aether-proxy setup` first");
    }
    Ok(())
 }
@@ -173,7 +174,7 @@ fn ensure_service_installed() -> anyhow::Result<()> {
 fn ensure_root_and_service() -> anyhow::Result<()> {
    ensure_service_installed()?;
    if !is_root() {
-        anyhow::bail!("root required, use: sudo aether-proxy <command>");
+        anyhow::bail!("root required, use: sudo ./aether-proxy <command>");
    }
    Ok(())
 }
@@ -2,7 +2,8 @@
 //!
 //! Launched via `aether-proxy setup [path]`.  Presents a full-screen form
 //! backed by ratatui where the user can navigate fields, edit values, and
-//! save to a TOML config file.
+//! save to a TOML config file.  Supports multi-server configuration via
+//! a tabbed interface.

 use std::io;
 use std::path::PathBuf;
@@ -19,13 +20,13 @@ use ratatui::widgets::{Block, Borders, Paragraph};
 use ratatui::Frame;
 use ratatui::Terminal;

-use crate::config::ConfigFile;
+use crate::config::{ConfigFile, ServerEntry};

 /// Outcome of the setup wizard, returned to the caller.
 pub enum SetupOutcome {
    /// Config saved; systemd service installed and started.
    ServiceInstalled,
-    /// Config saved; no service — caller should start the proxy directly.
+    /// Config saved; no service -- caller should start the proxy directly.
    ReadyToRun(PathBuf),
    /// User quit without saving.
    Cancelled,
@@ -34,13 +35,12 @@ pub enum SetupOutcome {
 /// Column width reserved for the field label (chars).
 const LABEL_WIDTH: usize = 22;

-// ── Field types ──────────────────────────────────────────────────────────────
+// -- Field types --------------------------------------------------------------

 #[derive(Clone, Copy, PartialEq)]
 enum FieldKind {
    Text,
    Secret,
-    Number,
    Bool,
    LogLevel,
 }
@@ -53,31 +53,15 @@ struct Field {
    required: bool,
    help: &'static str,
 }
+// -- Server tab ---------------------------------------------------------------

-// ── App state ────────────────────────────────────────────────────────────────
-
-#[derive(PartialEq)]
-enum Mode {
-    Normal,
-    Editing,
-}
-
-struct App {
+/// A single server tab's editable fields.
+struct ServerTab {
    fields: Vec<Field>,
-    selected: usize,
-    mode: Mode,
-    edit_buffer: String,
-    edit_cursor: usize, // char index
-    config_path: PathBuf,
-    modified: bool,
-    message: Option<(String, Instant, bool)>, // (text, when, is_error)
-    scroll_offset: usize,
-    saved_once: bool,
-    pending_quit: bool, // true after first q/Esc with unsaved changes
 }

-impl App {
-    fn new(config_path: PathBuf) -> Self {
+impl ServerTab {
+    fn new() -> Self {
        Self {
            fields: vec![
                Field {
@@ -86,7 +70,7 @@ impl App {
                    value: String::new(),
                    kind: FieldKind::Text,
                    required: true,
-                    help: "Aether 服务器 URL (如 https://aether.example.com)",
+                    help: "Aether URL (e.g. https://aether.example.com)",
                },
                Field {
                    label: "Management Token",
@@ -94,23 +78,7 @@ impl App {
                    value: String::new(),
                    kind: FieldKind::Secret,
                    required: true,
-                    help: "Aether 管理 API Token (ae_xxx)",
-                },
-                Field {
-                    label: "HMAC Key",
-                    key: "hmac_key",
-                    value: String::new(),
-                    kind: FieldKind::Secret,
-                    required: true,
-                    help: "HMAC-SHA256 签名密钥，用于代理请求认证",
-                },
-                Field {
-                    label: "Listen Port",
-                    key: "listen_port",
-                    value: "18080".into(),
-                    kind: FieldKind::Number,
-                    required: true,
-                    help: "代理服务监听端口",
+                    help: "Aether Management Token (ae_xxx)",
                },
                Field {
                    label: "Node Name",
@@ -118,15 +86,60 @@ impl App {
                    value: "proxy-01".into(),
                    kind: FieldKind::Text,
                    required: true,
-                    help: "节点名称，用于在 Aether 后台识别",
+                    help: "Node name for identification in Aether dashboard",
                },
+            ],
+        }
+    }
+
+    fn from_entry(entry: &ServerEntry) -> Self {
+        let mut tab = Self::new();
+        tab.fields[0].value = entry.aether_url.clone();
+        tab.fields[1].value = entry.management_token.clone();
+        if let Some(ref name) = entry.node_name {
+            tab.fields[2].value = name.clone();
+        }
+        tab
+    }
+}
+
+// -- App state ----------------------------------------------------------------
+
+#[derive(PartialEq)]
+enum Mode {
+    Normal,
+    Editing,
+}
+
+struct App {
+    server_tabs: Vec<ServerTab>,
+    active_tab: usize,
+    global_fields: Vec<Field>,
+    selected: usize,
+    mode: Mode,
+    edit_buffer: String,
+    edit_cursor: usize,
+    config_path: PathBuf,
+    modified: bool,
+    message: Option<(String, Instant, bool)>,
+    scroll_offset: usize,
+    saved_once: bool,
+    pending_quit: bool,
+    confirm_delete: bool,
+}
+impl App {
+    fn new(config_path: PathBuf) -> Self {
+        Self {
+            server_tabs: vec![ServerTab::new()],
+            active_tab: 0,
+            global_fields: vec![
                Field {
                    label: "Log Level",
                    key: "log_level",
                    value: "info".into(),
                    kind: FieldKind::LogLevel,
                    required: true,
-                    help: "日志级别 -- Enter 切换: trace / debug / info / warn / error",
+                    help: "Log level -- Enter to cycle: trace / debug / info / warn / error",
                },
                Field {
                    label: "Log JSON",
@@ -134,7 +147,7 @@ impl App {
                    value: "false".into(),
                    kind: FieldKind::Bool,
                    required: true,
-                    help: "是否以 JSON 格式输出日志 -- Enter 切换",
+                    help: "Output logs as JSON -- Enter to toggle",
                },
                Field {
                    label: "Install Service",
@@ -147,7 +160,7 @@ impl App {
                    .into(),
                    kind: FieldKind::Bool,
                    required: true,
-                    help: "注册为 systemd 开机启动服务 (需要 root 权限) -- Enter 切换",
+                    help: "Install as systemd service (requires root) -- Enter to toggle",
                },
            ],
            selected: 0,
@@ -160,10 +173,47 @@ impl App {
            scroll_offset: 0,
            saved_once: false,
            pending_quit: false,
+            confirm_delete: false,
        }
    }

-    // ── Config ↔ fields ──────────────────────────────────────────────────
+    // -- Field accessors (unified index across server + global) ---------------
+
+    fn server_field_count(&self) -> usize {
+        self.server_tabs[self.active_tab].fields.len()
+    }
+
+    fn total_field_count(&self) -> usize {
+        self.server_field_count() + self.global_fields.len()
+    }
+
+    fn selected_field(&self) -> &Field {
+        let sc = self.server_field_count();
+        if self.selected < sc {
+            &self.server_tabs[self.active_tab].fields[self.selected]
+        } else {
+            &self.global_fields[self.selected - sc]
+        }
+    }
+
+    fn selected_field_mut(&mut self) -> &mut Field {
+        let sc = self.server_field_count();
+        if self.selected < sc {
+            &mut self.server_tabs[self.active_tab].fields[self.selected]
+        } else {
+            &mut self.global_fields[self.selected - sc]
+        }
+    }
+
+    fn clamp_selection(&mut self) {
+        let max = self.total_field_count();
+        if self.selected >= max {
+            self.selected = max.saturating_sub(1);
+        }
+        self.scroll_offset = 0;
+        self.confirm_delete = false;
+    }
+    // -- Config <-> fields -----------------------------------------------------

    fn load_from_file(&mut self) {
        if let Ok(cfg) = ConfigFile::load(&self.config_path) {
@@ -172,13 +222,9 @@ impl App {
    }

    fn apply_config(&mut self, cfg: &ConfigFile) {
-        for field in &mut self.fields {
+        // Global fields
+        for field in &mut self.global_fields {
            let val: Option<String> = match field.key {
-                "aether_url" => cfg.aether_url.clone(),
-                "management_token" => cfg.management_token.clone(),
-                "hmac_key" => cfg.hmac_key.clone(),
-                "listen_port" => cfg.listen_port.map(|v| v.to_string()),
-                "node_name" => cfg.node_name.clone(),
                "log_level" => cfg.log_level.clone(),
                "log_json" => cfg.log_json.map(|v| v.to_string()),
                _ => None,
@@ -187,39 +233,76 @@ impl App {
                field.value = v;
            }
        }
+
+        // Server tabs
+        let servers = cfg.effective_servers();
+        if servers.is_empty() {
+            let mut tab = ServerTab::new();
+            // Single-server fallback: use top-level node_name
+            if let Some(ref name) = cfg.node_name {
+                tab.fields[2].value = name.clone();
+            }
+            self.server_tabs = vec![tab];
+        } else {
+            self.server_tabs = servers.iter().map(ServerTab::from_entry).collect();
+            // For single-server mode, node_name might be in top-level only
+            if self.server_tabs.len() == 1 && self.server_tabs[0].fields[2].value.is_empty() {
+                if let Some(ref name) = cfg.node_name {
+                    self.server_tabs[0].fields[2].value = name.clone();
+                }
+            }
+        }
+        self.active_tab = 0;
+        self.selected = 0;
+        self.scroll_offset = 0;
    }

    fn to_config(&self) -> ConfigFile {
-        let get = |key: &str| -> Option<String> {
-            self.fields
+        let get_global = |key: &str| -> Option<String> {
+            self.global_fields
                .iter()
                .find(|f| f.key == key)
                .map(|f| f.value.clone())
                .filter(|v| !v.is_empty())
        };

-        ConfigFile {
-            aether_url: get("aether_url"),
-            management_token: get("management_token"),
-            hmac_key: get("hmac_key"),
-            listen_port: get("listen_port").and_then(|v| v.parse().ok()),
-            public_ip: None,
-            node_name: get("node_name"),
-            node_region: None,
-            heartbeat_interval: None,
-            allowed_ports: None,
-            timestamp_tolerance: None,
-            log_level: get("log_level"),
-            log_json: get("log_json").and_then(|v| v.parse().ok()),
-            enable_tls: None,
-            tls_cert: None,
-            tls_key: None,
-        }
+        let get_tab = |tab: &ServerTab, key: &str| -> Option<String> {
+            tab.fields
+                .iter()
+                .find(|f| f.key == key)
+                .map(|f| f.value.clone())
+                .filter(|v| !v.is_empty())
+        };
+
+        let mut cfg = ConfigFile {
+            log_level: get_global("log_level"),
+            log_json: get_global("log_json").and_then(|v| v.parse().ok()),
+            ..ConfigFile::default()
+        };
+
+        // Always write [[servers]] format; old top-level fields are read-only compat
+        cfg.servers = self
+            .server_tabs
+            .iter()
+            .map(|tab| ServerEntry {
+                aether_url: get_tab(tab, "aether_url").unwrap_or_default(),
+                management_token: get_tab(tab, "management_token").unwrap_or_default(),
+                node_name: get_tab(tab, "node_name"),
+            })
+            .collect();
+        cfg
    }

    fn save(&mut self) -> anyhow::Result<()> {
        let cfg = self.to_config();
        cfg.save(&self.config_path)?;
+        // Restrict config file permissions to owner-only (contains management token).
+        #[cfg(unix)]
+        {
+            use std::os::unix::fs::PermissionsExt;
+            let _ =
+                std::fs::set_permissions(&self.config_path, std::fs::Permissions::from_mode(0o600));
+        }
        self.modified = false;
        self.saved_once = true;
        self.message = Some((
@@ -229,27 +312,33 @@ impl App {
        ));
        Ok(())
    }
-
-    // ── Scrolling ────────────────────────────────────────────────────────
+    // -- Scrolling ---------------------------------------------------------------

    fn ensure_visible(&mut self, visible_rows: usize) {
        if visible_rows == 0 {
            return;
        }
-        if self.selected < self.scroll_offset {
-            self.scroll_offset = self.selected;
-        } else if self.selected >= self.scroll_offset + visible_rows {
-            self.scroll_offset = self.selected - visible_rows + 1;
+        // Account for separator line between server and global fields
+        let display_row = if self.selected >= self.server_field_count() {
+            self.selected + 1
+        } else {
+            self.selected
+        };
+        if display_row < self.scroll_offset {
+            self.scroll_offset = display_row;
+        } else if display_row >= self.scroll_offset + visible_rows {
+            self.scroll_offset = display_row - visible_rows + 1;
        }
    }

-    // ── Key handling ─────────────────────────────────────────────────────
+    // -- Key handling -------------------------------------------------------------

    /// Returns `true` when the app should exit.
    fn handle_key(&mut self, key: KeyEvent) -> bool {
        // Expire old messages (but keep quit-confirmation messages alive)
        if let Some((_, when, _)) = &self.message {
-            if !self.pending_quit && when.elapsed() > Duration::from_secs(4) {
+            if !self.pending_quit && !self.confirm_delete && when.elapsed() > Duration::from_secs(4)
+            {
                self.message = None;
            }
        }
@@ -264,7 +353,7 @@ impl App {
    }

    fn handle_normal(&mut self, key: KeyEvent) -> bool {
-        // ── Quit handling (with unsaved-changes confirmation) ─────────
+        // -- Quit handling (with unsaved-changes confirmation) -----------------
        let is_quit_key = matches!(key.code, KeyCode::Char('q') | KeyCode::Esc);

        if is_quit_key {
@@ -272,6 +361,7 @@ impl App {
                return true;
            }
            self.pending_quit = true;
+            self.confirm_delete = false;
            self.message = Some((
                "unsaved changes! q again to discard, ^S to save".into(),
                Instant::now(),
@@ -280,14 +370,21 @@ impl App {
            return false;
        }

-        // Any other key cancels the pending quit
+        // Any other key cancels pending quit / pending delete
        if self.pending_quit {
            self.pending_quit = false;
            self.message = None;
        }
+        if self.confirm_delete && !matches!(key.code, KeyCode::Delete | KeyCode::Char('x')) {
+            self.confirm_delete = false;
+            self.message = None;
+        }

        match key.code {
-            KeyCode::Char('s') if key.modifiers.contains(KeyModifiers::CONTROL) => {
+            KeyCode::Char('s')
+                if key.modifiers.contains(KeyModifiers::CONTROL)
+                    || key.modifiers.contains(KeyModifiers::SUPER) =>
+            {
                if let Err(e) = self.save() {
                    self.message = Some((format!("error: {}", e), Instant::now(), true));
                }
@@ -296,23 +393,20 @@ impl App {
                self.selected = self.selected.saturating_sub(1);
            }
            KeyCode::Down | KeyCode::Char('j') => {
-                if self.selected + 1 < self.fields.len() {
+                if self.selected + 1 < self.total_field_count() {
                    self.selected += 1;
                }
            }
            KeyCode::Home => self.selected = 0,
-            KeyCode::End => self.selected = self.fields.len() - 1,
+            KeyCode::End => self.selected = self.total_field_count() - 1,
            KeyCode::Enter | KeyCode::Char(' ') => {
-                let field = &self.fields[self.selected];
-                match field.kind {
+                let kind = self.selected_field().kind;
+                let key_str = self.selected_field().key;
+                let value = self.selected_field().value.clone();
+                match kind {
                    FieldKind::Bool => {
-                        let toggled = if field.value == "true" {
-                            "false"
-                        } else {
-                            "true"
-                        };
-                        // Block enabling service install without root/systemd
-                        if field.key == "install_service"
+                        let toggled = if value == "true" { "false" } else { "true" };
+                        if key_str == "install_service"
                            && toggled == "true"
                            && !super::service::is_available()
                        {
@@ -322,27 +416,79 @@ impl App {
                                true,
                            ));
                        } else {
-                            self.fields[self.selected].value = toggled.into();
+                            self.selected_field_mut().value = toggled.into();
                            self.modified = true;
                        }
                    }
                    FieldKind::LogLevel => {
                        const LEVELS: &[&str] = &["trace", "debug", "info", "warn", "error"];
-                        let idx = LEVELS.iter().position(|l| *l == field.value).unwrap_or(2);
-                        self.fields[self.selected].value = LEVELS[(idx + 1) % LEVELS.len()].into();
+                        let idx = LEVELS.iter().position(|l| *l == value).unwrap_or(2);
+                        self.selected_field_mut().value = LEVELS[(idx + 1) % LEVELS.len()].into();
                        self.modified = true;
                    }
                    _ => {
-                        self.edit_buffer = field.value.clone();
+                        self.edit_buffer = value;
                        self.edit_cursor = self.edit_buffer.chars().count();
                        self.mode = Mode::Editing;
                    }
                }
            }
+            // -- Tab navigation --
            KeyCode::Tab => {
-                // Quick save shortcut
-                if let Err(e) = self.save() {
-                    self.message = Some((format!("error: {}", e), Instant::now(), true));
+                if self.server_tabs.len() > 1 {
+                    self.active_tab = (self.active_tab + 1) % self.server_tabs.len();
+                    self.clamp_selection();
+                }
+            }
+            KeyCode::BackTab => {
+                if self.server_tabs.len() > 1 {
+                    self.active_tab = if self.active_tab == 0 {
+                        self.server_tabs.len() - 1
+                    } else {
+                        self.active_tab - 1
+                    };
+                    self.clamp_selection();
+                }
+            }
+            KeyCode::Char(c @ '1'..='9') if !key.modifiers.contains(KeyModifiers::CONTROL) => {
+                let idx = (c as usize) - ('1' as usize);
+                if idx < self.server_tabs.len() && idx != self.active_tab {
+                    self.active_tab = idx;
+                    self.clamp_selection();
+                }
+            }
+            // -- Add / remove server --
+            KeyCode::Char('+') | KeyCode::Char('a') => {
+                self.server_tabs.push(ServerTab::new());
+                self.active_tab = self.server_tabs.len() - 1;
+                self.selected = 0;
+                self.scroll_offset = 0;
+                self.modified = true;
+                self.message = Some((
+                    format!("added server {}", self.server_tabs.len()),
+                    Instant::now(),
+                    false,
+                ));
+            }
+            KeyCode::Delete | KeyCode::Char('x') => {
+                if self.server_tabs.len() <= 1 {
+                    self.message =
+                        Some(("cannot remove the last server".into(), Instant::now(), true));
+                } else if self.confirm_delete {
+                    let removed = self.active_tab + 1;
+                    self.server_tabs.remove(self.active_tab);
+                    self.active_tab = self.active_tab.min(self.server_tabs.len() - 1);
+                    self.clamp_selection();
+                    self.modified = true;
+                    self.message =
+                        Some((format!("server {} removed", removed), Instant::now(), false));
+                } else {
+                    self.confirm_delete = true;
+                    self.message = Some((
+                        "press Delete/x again to remove this server".into(),
+                        Instant::now(),
+                        true,
+                    ));
                }
            }
            _ => {}
@@ -353,12 +499,11 @@ impl App {
    fn handle_edit(&mut self, key: KeyEvent) {
        match key.code {
            KeyCode::Esc => {
-                // Cancel -- discard changes to this field
                self.mode = Mode::Normal;
            }
            KeyCode::Enter => {
                if self.validate_edit() {
-                    self.fields[self.selected].value = self.edit_buffer.clone();
+                    self.selected_field_mut().value = self.edit_buffer.clone();
                    self.modified = true;
                    self.mode = Mode::Normal;
                } else {
@@ -399,12 +544,7 @@ impl App {
    }

    fn validate_edit(&self) -> bool {
-        let kind = self.fields[self.selected].kind;
-        let buf = &self.edit_buffer;
-        match kind {
-            FieldKind::Number => buf.is_empty() || buf.parse::<u64>().is_ok(),
-            _ => true,
-        }
+        true
    }

    /// Byte offset of the char at `char_idx`.
@@ -416,13 +556,11 @@ impl App {
            .unwrap_or(self.edit_buffer.len())
    }
 }
-
-// ── Rendering ────────────────────────────────────────────────────────────────
+// -- Rendering ----------------------------------------------------------------

 fn ui(f: &mut Frame, app: &mut App) {
    let area = f.area();

-    // Outer block
    let title = if app.modified {
        " Aether Proxy Setup [*] "
    } else {
@@ -438,53 +576,52 @@ fn ui(f: &mut Frame, app: &mut App) {
    let inner = outer.inner(area);
    f.render_widget(outer, area);

-    // Split: fields | footer
-    let chunks = Layout::vertical([Constraint::Min(1), Constraint::Length(4)]).split(inner);
+    // Split: fields | tab bar | footer
+    let chunks = Layout::vertical([
+        Constraint::Min(1),
+        Constraint::Length(1),
+        Constraint::Length(4),
+    ])
+    .split(inner);

-    let fields_area = chunks[0];
-    let footer_area = chunks[1];
-
-    render_fields(f, app, fields_area);
-    render_footer(f, app, footer_area);
+    render_fields(f, app, chunks[0]);
+    render_tab_bar(f, app, chunks[1]);
+    render_footer(f, app, chunks[2]);
 }

 fn render_fields(f: &mut Frame, app: &mut App, area: Rect) {
    let visible = area.height as usize;
    app.ensure_visible(visible);

+    let server_count = app.server_field_count();
    let mut lines: Vec<Line> = Vec::new();
+    // display_row tracks the actual row index (including separator)
+    let mut display_row: usize = 0;

-    for (i, field) in app.fields.iter().enumerate() {
-        if i < app.scroll_offset || i >= app.scroll_offset + visible {
-            continue;
+    // Server fields
+    for i in 0..server_count {
+        if display_row >= app.scroll_offset && display_row < app.scroll_offset + visible {
+            lines.push(build_field_line(app, i, display_row));
        }
+        display_row += 1;
+    }

-        let selected = i == app.selected;
-        let indicator = if selected { " > " } else { "   " };
+    // Separator line
+    if display_row >= app.scroll_offset && display_row < app.scroll_offset + visible {
+        lines.push(Line::from(Span::styled(
+            "   ----------------------------------------",
+            Style::default().fg(Color::DarkGray),
+        )));
+    }
+    display_row += 1;

-        let label_style = if selected {
-            Style::default()
-                .fg(Color::Cyan)
-                .add_modifier(Modifier::BOLD)
-        } else {
-            Style::default().fg(Color::DarkGray)
-        };
-
-        let padded_label = format!("{:<width$}", field.label, width = LABEL_WIDTH);
-
-        // Value display
-        let (value_text, value_style) = if app.mode == Mode::Editing && selected {
-            (app.edit_buffer.clone(), Style::default().fg(Color::Yellow))
-        } else {
-            field_display(field)
-        };
-
-        lines.push(Line::from(vec![
-            Span::styled(indicator, label_style),
-            Span::styled(padded_label, label_style),
-            Span::raw("  "),
-            Span::styled(value_text, value_style),
-        ]));
+    // Global fields
+    for i in 0..app.global_fields.len() {
+        let field_idx = server_count + i;
+        if display_row >= app.scroll_offset && display_row < app.scroll_offset + visible {
+            lines.push(build_field_line(app, field_idx, display_row));
+        }
+        display_row += 1;
    }

    let paragraph = Paragraph::new(lines);
@@ -492,8 +629,12 @@ fn render_fields(f: &mut Frame, app: &mut App, area: Rect) {

    // Cursor position while editing
    if app.mode == Mode::Editing {
-        let row_in_view = app.selected - app.scroll_offset;
-        // prefix: 3 (indicator) + LABEL_WIDTH + 2 (gap) = 27
+        let sel_display_row = if app.selected >= server_count {
+            app.selected + 1
+        } else {
+            app.selected
+        };
+        let row_in_view = sel_display_row.saturating_sub(app.scroll_offset);
        let prefix: u16 = 3 + LABEL_WIDTH as u16 + 2;
        let cx = area.x + prefix + app.edit_cursor as u16;
        let cy = area.y + row_in_view as u16;
@@ -502,6 +643,40 @@ fn render_fields(f: &mut Frame, app: &mut App, area: Rect) {
        }
    }
 }
+fn build_field_line(app: &App, field_idx: usize, _display_row: usize) -> Line<'static> {
+    let sc = app.server_field_count();
+    let field = if field_idx < sc {
+        &app.server_tabs[app.active_tab].fields[field_idx]
+    } else {
+        &app.global_fields[field_idx - sc]
+    };
+
+    let selected = field_idx == app.selected;
+    let indicator = if selected { " > " } else { "   " };
+
+    let label_style = if selected {
+        Style::default()
+            .fg(Color::Cyan)
+            .add_modifier(Modifier::BOLD)
+    } else {
+        Style::default().fg(Color::DarkGray)
+    };
+
+    let padded_label = format!("{:<width$}", field.label, width = LABEL_WIDTH);
+
+    let (value_text, value_style) = if app.mode == Mode::Editing && selected {
+        (app.edit_buffer.clone(), Style::default().fg(Color::Yellow))
+    } else {
+        field_display(field)
+    };
+
+    Line::from(vec![
+        Span::styled(indicator.to_string(), label_style),
+        Span::styled(padded_label, label_style),
+        Span::raw("  "),
+        Span::styled(value_text, value_style),
+    ])
+}

 /// Returns (display_text, style) for a field in normal mode.
 fn field_display(field: &Field) -> (String, Style) {
@@ -545,18 +720,54 @@ fn field_display(field: &Field) -> (String, Style) {
        _ => (field.value.clone(), Style::default().fg(Color::White)),
    }
 }
+fn render_tab_bar(f: &mut Frame, app: &App, area: Rect) {
+    let mut spans: Vec<Span> = Vec::new();
+    spans.push(Span::raw(" "));
+
+    for (i, tab) in app.server_tabs.iter().enumerate() {
+        let num = i + 1;
+        let name = tab
+            .fields
+            .iter()
+            .find(|f| f.key == "node_name")
+            .filter(|f| !f.value.is_empty())
+            .map(|f| f.value.clone())
+            .unwrap_or_else(|| format!("Server {}", num));
+
+        let label = format!(" {} {} ", num, name);
+
+        if i == app.active_tab {
+            spans.push(Span::styled(
+                label,
+                Style::default()
+                    .fg(Color::Black)
+                    .bg(Color::Cyan)
+                    .add_modifier(Modifier::BOLD),
+            ));
+        } else {
+            spans.push(Span::styled(label, Style::default().fg(Color::DarkGray)));
+        }
+        spans.push(Span::raw(" "));
+    }
+
+    spans.push(Span::styled(" + Add ", Style::default().fg(Color::Green)));
+
+    f.render_widget(Paragraph::new(Line::from(spans)), area);
+}

 fn render_footer(f: &mut Frame, app: &App, area: Rect) {
-    let help = app.fields[app.selected].help;
+    let help = app.selected_field().help;

    let keybindings = if app.mode == Mode::Editing {
        "Enter confirm  Esc cancel"
+    } else if app.server_tabs.len() > 1 {
+        "j/k select  Enter edit  Tab switch  + add  x remove  ^S save  q quit"
    } else {
-        "Up/Down select  Enter edit  ^S save  q quit"
+        "j/k select  Enter edit  + add server  ^S save  q quit"
    };

    let mut status_spans: Vec<Span> = vec![Span::styled(
-        keybindings,
+        format!(" {}", keybindings),
        Style::default().fg(Color::DarkGray),
    )];

@@ -572,18 +783,7 @@ fn render_footer(f: &mut Frame, app: &App, area: Rect) {
            format!(" {}", help),
            Style::default().fg(Color::DarkGray),
        )),
-        Line::from(
-            status_spans
-                .into_iter()
-                .map(|mut s| {
-                    // add left padding to first span
-                    if s.content.as_ref() == keybindings {
-                        s.content = format!(" {}", s.content).into();
-                    }
-                    s
-                })
-                .collect::<Vec<_>>(),
-        ),
+        Line::from(status_spans),
    ];

    let footer = Paragraph::new(footer_text).block(
@@ -594,11 +794,9 @@ fn render_footer(f: &mut Frame, app: &App, area: Rect) {

    f.render_widget(footer, area);
 }
-
-// ── Entry point ──────────────────────────────────────────────────────────────
+// -- Entry point --------------------------------------------------------------

 pub fn run(config_path: PathBuf) -> anyhow::Result<SetupOutcome> {
-    // Setup terminal
    terminal::enable_raw_mode()?;
    let mut stdout = io::stdout();
    execute!(stdout, EnterAlternateScreen)?;
@@ -610,14 +808,13 @@ pub fn run(config_path: PathBuf) -> anyhow::Result<SetupOutcome> {

    let result = event_loop(&mut terminal, &mut app);

-    // Restore terminal
    terminal::disable_raw_mode()?;
    execute!(terminal.backend_mut(), LeaveAlternateScreen)?;
    terminal.show_cursor()?;

    result?;

-    // ── Post-TUI: decide outcome ─────────────────────────────────────
+    // -- Post-TUI: decide outcome ---------------------------------------------

    if !app.saved_once {
        return Ok(SetupOutcome::Cancelled);
@@ -628,7 +825,7 @@ pub fn run(config_path: PathBuf) -> anyhow::Result<SetupOutcome> {
    eprintln!();

    let wants_service = app
-        .fields
+        .global_fields
        .iter()
        .find(|f| f.key == "install_service")
        .map(|f| f.value == "true")
@@ -642,13 +839,10 @@ pub fn run(config_path: PathBuf) -> anyhow::Result<SetupOutcome> {
                eprintln!("  Starting proxy directly instead.\n");
            }
        }
-    } else {
-        // Uninstall service if it was previously installed but toggled off
-        if super::service::is_installed() {
-            if let Err(e) = super::service::uninstall_service() {
-                eprintln!("  Service uninstall failed: {}", e);
-                eprintln!();
-            }
+    } else if super::service::is_installed() {
+        if let Err(e) = super::service::uninstall_service() {
+            eprintln!("  Service uninstall failed: {}", e);
+            eprintln!();
        }
    }

@@ -664,7 +858,6 @@ fn event_loop(

        if event::poll(Duration::from_millis(200))? {
            if let Event::Key(key) = event::read()? {
-                // Only handle Press events (ignore Release on Windows)
                if key.kind == KeyEventKind::Press && app.handle_key(key) {
                    break;
                }
@@ -17,13 +17,6 @@ const CURRENT_VERSION: &str = env!("CARGO_PKG_VERSION");
 struct GithubRelease {
    tag_name: String,
    name: String,
-    assets: Vec<GithubAsset>,
-}
-
-#[derive(serde::Deserialize)]
-struct GithubAsset {
-    name: String,
-    id: u64,
 }

 // ── Platform detection ───────────────────────────────────────────────────────
@@ -49,17 +42,15 @@ fn detect_platform() -> &'static str {
 // ── GitHub HTTP client ───────────────────────────────────────────────────────

 fn build_github_client() -> anyhow::Result<reqwest::Client> {
-    let token = std::env::var("GITHUB_TOKEN").map_err(|_| {
-        anyhow::anyhow!(
-            "GITHUB_TOKEN is required (private repo).\n  Set it via: export GITHUB_TOKEN=ghp_xxx"
-        )
-    })?;
-
    let mut headers = reqwest::header::HeaderMap::new();
-    headers.insert(
-        reqwest::header::AUTHORIZATION,
-        reqwest::header::HeaderValue::from_str(&format!("Bearer {}", token))?,
-    );
+
+    if let Ok(token) = std::env::var("GITHUB_TOKEN") {
+        headers.insert(
+            reqwest::header::AUTHORIZATION,
+            reqwest::header::HeaderValue::from_str(&format!("Bearer {}", token))?,
+        );
+    }
+
    headers.insert(
        reqwest::header::ACCEPT,
        reqwest::header::HeaderValue::from_static("application/vnd.github+json"),
@@ -119,16 +110,18 @@ async fn fetch_release(
    }
 }

-// ── Download & verify ────────────────────────────────────────────────────────
+// ── Download via GitHub release direct links ─────────────────────────────────

-async fn download_asset_bytes(
+/// Download a release asset via the public direct download URL:
+/// `https://github.com/{repo}/releases/download/{tag}/{filename}`
+async fn download_release_file(
    client: &reqwest::Client,
-    asset: &GithubAsset,
+    tag: &str,
+    filename: &str,
 ) -> anyhow::Result<Vec<u8>> {
-    // Use GitHub API asset endpoint for reliable private repo downloads
    let url = format!(
-        "{}/repos/{}/releases/assets/{}",
-        GITHUB_API_BASE, GITHUB_REPO, asset.id
+        "https://github.com/{}/releases/download/{}/{}",
+        GITHUB_REPO, tag, filename
    );
    let resp = client
        .get(&url)
@@ -138,7 +131,7 @@ async fn download_asset_bytes(
    if !resp.status().is_success() {
        anyhow::bail!(
            "download failed for '{}' (HTTP {})",
-            asset.name,
+            filename,
            resp.status(),
        );
    }
@@ -161,28 +154,16 @@ fn parse_checksum(sums_text: &str, filename: &str) -> anyhow::Result<String> {

 async fn download_and_verify(
    client: &reqwest::Client,
-    release: &GithubRelease,
+    tag: &str,
    platform: &str,
    dest: &Path,
 ) -> anyhow::Result<()> {
    let archive_name = format!("aether-proxy-{}.tar.gz", platform);

-    let archive_asset = release
-        .assets
-        .iter()
-        .find(|a| a.name == archive_name)
-        .ok_or_else(|| anyhow::anyhow!("asset '{}' not found in release", archive_name))?;
-
-    let checksum_asset = release
-        .assets
-        .iter()
-        .find(|a| a.name == "SHA256SUMS.txt")
-        .ok_or_else(|| anyhow::anyhow!("SHA256SUMS.txt not found in release"))?;
-
    eprintln!("  Downloading {}...", archive_name);
    let (archive_bytes, checksum_bytes) = tokio::try_join!(
-        download_asset_bytes(client, archive_asset),
-        download_asset_bytes(client, checksum_asset),
+        download_release_file(client, tag, &archive_name),
+        download_release_file(client, tag, "SHA256SUMS.txt"),
    )?;
    let checksum_text = String::from_utf8(checksum_bytes)?;

@@ -300,8 +281,17 @@ fn atomic_replace(new_binary: &Path) -> anyhow::Result<PathBuf> {

 // ── Public entry point ───────────────────────────────────────────────────────

-/// `aether-proxy upgrade [version]` -- self-upgrade from GitHub releases.
-pub async fn cmd_upgrade(version: Option<String>) -> anyhow::Result<()> {
+#[derive(Clone, Copy)]
+enum RestartMode {
+    BestEffort,
+    Required,
+}
+
+async fn execute_upgrade(
+    version: Option<&str>,
+    require_root: bool,
+    restart_mode: RestartMode,
+) -> anyhow::Result<()> {
    // Resolve exe path once; reuse throughout the function
    let current_exe = std::env::current_exe()?.canonicalize()?;
    let exe_dir = current_exe
@@ -309,8 +299,12 @@ pub async fn cmd_upgrade(version: Option<String>) -> anyhow::Result<()> {
        .ok_or_else(|| anyhow::anyhow!("cannot determine binary directory"))?;
    let temp_path = exe_dir.join(".aether-proxy.upgrade.tmp");

-    // Check write permission to binary directory
-    if !super::service::is_root() {
+    if require_root {
+        if !super::service::is_root() {
+            anyhow::bail!("automatic upgrade requires root privileges");
+        }
+    } else if !super::service::is_root() {
+        // Check write permission to binary directory for manual upgrade mode.
        let test_path = exe_dir.join(".aether-proxy.write-test");
        match std::fs::File::create(&test_path) {
            Ok(_) => {
@@ -330,7 +324,7 @@ pub async fn cmd_upgrade(version: Option<String>) -> anyhow::Result<()> {
    eprintln!("  Current version: {}", CURRENT_VERSION);

    let client = build_github_client()?;
-    let release = fetch_release(&client, version.as_deref()).await?;
+    let release = fetch_release(&client, version).await?;
    let target_tag = &release.tag_name;
    let target_semver = target_tag.strip_prefix("proxy-v").unwrap_or(target_tag);

@@ -348,7 +342,7 @@ pub async fn cmd_upgrade(version: Option<String>) -> anyhow::Result<()> {
    eprintln!("  Upgrading: {} -> {}", CURRENT_VERSION, target_semver);
    eprintln!();

-    if let Err(e) = download_and_verify(&client, &release, platform, &temp_path).await {
+    if let Err(e) = download_and_verify(&client, target_tag, platform, &temp_path).await {
        let _ = std::fs::remove_file(&temp_path);
        return Err(e);
    }
@@ -360,13 +354,38 @@ pub async fn cmd_upgrade(version: Option<String>) -> anyhow::Result<()> {
        }
    };

-    // Restart systemd service if running
-    if super::service::is_service_active() {
-        eprintln!("  Restarting systemd service...");
-        super::service::run_cmd("systemctl", &["restart", "aether-proxy"])?;
-        eprintln!("  Service restarted.");
-    } else {
-        eprintln!("  No active systemd service detected, skipping restart.");
+    match restart_mode {
+        RestartMode::BestEffort => {
+            // Restart systemd service if running.
+            // Use best-effort: binary is already replaced, so a restart failure should
+            // not abort the whole upgrade -- the user can restart manually.
+            if super::service::is_service_active() {
+                if super::service::is_root() {
+                    eprintln!("  Restarting systemd service...");
+                    match super::service::run_cmd("systemctl", &["restart", "aether-proxy"]) {
+                        Ok(()) => eprintln!("  Service restarted."),
+                        Err(e) => {
+                            eprintln!("  WARNING: failed to restart service: {}", e);
+                            eprintln!("  Run manually: sudo systemctl restart aether-proxy");
+                        }
+                    }
+                } else {
+                    eprintln!("  Systemd service is active, but restart requires root.");
+                    eprintln!("  Run: sudo systemctl restart aether-proxy");
+                    eprintln!("  Skipping restart.");
+                }
+            } else {
+                eprintln!("  No active systemd service detected, skipping restart.");
+            }
+        }
+        RestartMode::Required => {
+            if !super::service::is_root() {
+                anyhow::bail!("automatic upgrade requires root privileges");
+            }
+            eprintln!("  Restarting systemd service...");
+            super::service::run_cmd("systemctl", &["restart", "aether-proxy"])?;
+            eprintln!("  Service restarted.");
+        }
    }

    eprintln!();
@@ -377,3 +396,16 @@ pub async fn cmd_upgrade(version: Option<String>) -> anyhow::Result<()> {
    );
    Ok(())
 }
+
+/// `aether-proxy upgrade [version]` -- self-upgrade from GitHub releases.
+pub async fn cmd_upgrade(version: Option<String>) -> anyhow::Result<()> {
+    execute_upgrade(version.as_deref(), false, RestartMode::BestEffort).await
+}
+
+/// Perform automatic upgrade to a specific version.
+///
+/// This path is designed for server-pushed upgrades in systemd/root scenarios:
+/// it requires root and requires a successful `systemctl restart aether-proxy`.
+pub async fn perform_upgrade(version: &str) -> anyhow::Result<()> {
+    execute_upgrade(Some(version), true, RestartMode::Required).await
+}
@@ -1,30 +1,77 @@
 //! Shared application state passed to all subsystems.
-//!
-//! Consolidates the multiple `Arc<...>` parameters that were previously
-//! threaded individually through proxy server, heartbeat, and handlers.

-use std::sync::atomic::AtomicU64;
+use std::sync::atomic::{AtomicU64, Ordering};
 use std::sync::{Arc, RwLock};
-
-use tokio_rustls::TlsAcceptor;
+use std::time::Duration;

 use crate::config::Config;
-use crate::hardware::HardwareInfo;
 use crate::registration::client::AetherClient;
 use crate::runtime::SharedDynamicConfig;
+use crate::target_filter::DnsCache;
+use crate::upstream_client::UpstreamClient;

-/// Central application state shared across all tasks.
+/// Central application state shared across all servers/tunnels.
 pub struct AppState {
    pub config: Arc<Config>,
+    /// DNS cache for upstream target resolution (shared).
+    pub dns_cache: Arc<DnsCache>,
+    /// Hyper client for tunnel upstream requests with validated DNS and connection timing.
+    pub upstream_client: UpstreamClient,
+    /// Shared TLS config for tunnel WebSocket connections (avoids re-parsing root CAs on each reconnect).
+    pub tunnel_tls_config: Arc<rustls::ClientConfig>,
+}
+
+/// Per-server state: one instance per Aether server connection.
+pub struct ServerContext {
+    /// Human-readable label for logging (e.g. "server-0").
+    pub server_label: String,
+    /// Aether server URL for this connection.
+    pub aether_url: String,
+    /// Management token for this server.
+    pub management_token: String,
+    /// Resolved node name at registration time (per-server override or global fallback).
+    /// After startup, the active node_name is read from `dynamic` (may be updated remotely).
+    #[allow(dead_code)]
+    pub node_name: String,
+    /// Node ID assigned by this Aether server.
    pub node_id: Arc<RwLock<String>>,
+    /// API client for this server.
+    pub aether_client: Arc<AetherClient>,
+    /// Dynamic config from this server's heartbeat ACKs.
    pub dynamic: SharedDynamicConfig,
-    pub aether_client: Arc<AetherClient>,
-    pub hardware_info: Arc<HardwareInfo>,
-    pub public_ip: String,
-    pub tls_fingerprint: Option<String>,
-    pub tls_acceptor: Option<TlsAcceptor>,
-    /// Shared reqwest client for delegate mode (proxy issues upstream requests directly).
-    pub delegate_client: reqwest::Client,
-    /// Active connection count for metrics reporting.
+    /// Per-server active connection count.
    pub active_connections: Arc<AtomicU64>,
+    /// Per-server request/latency metrics.
+    pub metrics: Arc<ProxyMetrics>,
+}
+
+/// Aggregate metrics for reporting to Aether.
+pub struct ProxyMetrics {
+    pub total_requests: AtomicU64,
+    /// Cumulative connection-establishment latency in nanoseconds
+    /// (DNS + TCP/TLS + TTFB, excludes response body streaming).
+    pub total_latency_ns: AtomicU64,
+    pub failed_requests: AtomicU64,
+    pub dns_failures: AtomicU64,
+    pub stream_errors: AtomicU64,
+}
+
+impl ProxyMetrics {
+    pub fn new() -> Self {
+        Self {
+            total_requests: AtomicU64::new(0),
+            total_latency_ns: AtomicU64::new(0),
+            failed_requests: AtomicU64::new(0),
+            dns_failures: AtomicU64::new(0),
+            stream_errors: AtomicU64::new(0),
+        }
+    }
+
+    /// Record a completed request with its connection-establishment latency
+    /// (DNS + TCP/TLS + TTFB, excludes response body streaming).
+    pub fn record_request(&self, connect_elapsed: Duration) {
+        let nanos = u64::try_from(connect_elapsed.as_nanos()).unwrap_or(u64::MAX);
+        self.total_requests.fetch_add(1, Ordering::Release);
+        self.total_latency_ns.fetch_add(nanos, Ordering::Release);
+    }
 }
@@ -0,0 +1,381 @@
+use std::collections::{HashMap, HashSet};
+use std::net::{IpAddr, Ipv4Addr, Ipv6Addr, SocketAddr};
+use std::sync::Arc;
+use std::time::{Duration, Instant};
+
+use tokio::sync::RwLock;
+
+/// Check if an IP address belongs to a private/reserved network.
+pub fn is_private_ip(ip: &IpAddr) -> bool {
+    match ip {
+        IpAddr::V4(v4) => is_private_ipv4(v4),
+        IpAddr::V6(v6) => is_private_ipv6(v6),
+    }
+}
+
+fn is_private_ipv4(ip: &Ipv4Addr) -> bool {
+    let octets = ip.octets();
+    // 10.0.0.0/8
+    if octets[0] == 10 {
+        return true;
+    }
+    // 172.16.0.0/12
+    if octets[0] == 172 && (16..=31).contains(&octets[1]) {
+        return true;
+    }
+    // 192.168.0.0/16
+    if octets[0] == 192 && octets[1] == 168 {
+        return true;
+    }
+    // 127.0.0.0/8
+    if octets[0] == 127 {
+        return true;
+    }
+    // 169.254.0.0/16 (link-local)
+    if octets[0] == 169 && octets[1] == 254 {
+        return true;
+    }
+    // 0.0.0.0/8
+    if octets[0] == 0 {
+        return true;
+    }
+    // 100.64.0.0/10 (CGNAT / shared address space)
+    if octets[0] == 100 && (64..=127).contains(&octets[1]) {
+        return true;
+    }
+    // 192.0.0.0/24 (IETF protocol assignments)
+    if octets[0] == 192 && octets[1] == 0 && octets[2] == 0 {
+        return true;
+    }
+    // 198.18.0.0/15 (benchmark testing)
+    if octets[0] == 198 && (18..=19).contains(&octets[1]) {
+        return true;
+    }
+    // 240.0.0.0/4 (reserved for future use)
+    if octets[0] >= 240 {
+        return true;
+    }
+    false
+}
+
+fn is_private_ipv6(ip: &Ipv6Addr) -> bool {
+    // ::1 loopback
+    if ip.is_loopback() {
+        return true;
+    }
+    // :: unspecified
+    if ip.is_unspecified() {
+        return true;
+    }
+    let segments = ip.segments();
+    // fc00::/7 (ULA) - first byte is 0xfc or 0xfd
+    if segments[0] & 0xfe00 == 0xfc00 {
+        return true;
+    }
+    // fe80::/10 (link-local)
+    if segments[0] & 0xffc0 == 0xfe80 {
+        return true;
+    }
+    // IPv4-mapped IPv6 (::ffff:x.x.x.x) - check the embedded IPv4
+    if let Some(v4) = ip.to_ipv4_mapped() {
+        return is_private_ipv4(&v4);
+    }
+    false
+}
+
+#[derive(Debug)]
+pub enum FilterError {
+    PrivateIp(IpAddr),
+    PortNotAllowed(u16),
+    DnsResolutionFailed(String),
+    NoPublicAddrs(String),
+}
+
+impl std::fmt::Display for FilterError {
+    fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
+        match self {
+            Self::PrivateIp(ip) => write!(f, "target IP {} is in private/reserved range", ip),
+            Self::PortNotAllowed(port) => write!(f, "port {} not in allowed list", port),
+            Self::DnsResolutionFailed(host) => write!(f, "DNS resolution failed for {}", host),
+            Self::NoPublicAddrs(host) => {
+                write!(
+                    f,
+                    "all resolved addresses for {} are private/reserved",
+                    host
+                )
+            }
+        }
+    }
+}
+
+struct DnsCacheEntry {
+    addrs: Arc<Vec<SocketAddr>>,
+    expires_at: Instant,
+    inserted_at: Instant,
+}
+
+/// Lightweight DNS cache with TTL + capacity bounds.
+/// Stores all public resolved addresses per host (used by SafeDnsResolver
+/// to ensure reqwest connects to the same validated addresses).
+pub struct DnsCache {
+    ttl: Duration,
+    capacity: usize,
+    entries: RwLock<HashMap<String, DnsCacheEntry>>,
+}
+
+impl DnsCache {
+    pub fn new(ttl: Duration, capacity: usize) -> Self {
+        Self {
+            ttl,
+            capacity,
+            entries: RwLock::new(HashMap::new()),
+        }
+    }
+
+    /// Look up cached public addresses for a host (any port).
+    ///
+    /// Used by `SafeDnsResolver` which only knows the hostname — returns the
+    /// first unexpired entry whose key starts with `host:`.
+    pub async fn get_by_host(&self, host: &str) -> Option<Arc<Vec<SocketAddr>>> {
+        if self.capacity == 0 || self.ttl.is_zero() {
+            return None;
+        }
+        let prefix = format!("{}:", host.to_ascii_lowercase());
+        let now = Instant::now();
+        let entries = self.entries.read().await;
+        for (key, entry) in entries.iter() {
+            if key.starts_with(&prefix) && entry.expires_at > now {
+                return Some(Arc::clone(&entry.addrs));
+            }
+        }
+        None
+    }
+
+    /// Look up cached public addresses for a host + port.
+    pub async fn get(&self, host: &str, port: u16) -> Option<Arc<Vec<SocketAddr>>> {
+        if self.capacity == 0 || self.ttl.is_zero() {
+            return None;
+        }
+        let key = Self::key(host, port);
+        let now = Instant::now();
+
+        // Fast path: read lock for cache hit
+        {
+            let entries = self.entries.read().await;
+            match entries.get(&key) {
+                Some(entry) if entry.expires_at > now => return Some(Arc::clone(&entry.addrs)),
+                None => return None,
+                Some(_) => {} // expired, fall through to evict
+            }
+        }
+
+        // Slow path: write lock to remove expired entry
+        let mut entries = self.entries.write().await;
+        entries.remove(&key);
+        None
+    }
+
+    /// Insert resolved public addresses into cache.
+    pub async fn insert(&self, host: &str, port: u16, addrs: Arc<Vec<SocketAddr>>) {
+        if self.capacity == 0 || self.ttl.is_zero() || addrs.is_empty() {
+            return;
+        }
+        let key = Self::key(host, port);
+        let now = Instant::now();
+        let mut entries = self.entries.write().await;
+        entries.retain(|_, entry| entry.expires_at > now);
+        while entries.len() >= self.capacity {
+            let oldest_key = entries
+                .iter()
+                .min_by_key(|(_, entry)| entry.inserted_at)
+                .map(|(key, _)| key.clone());
+            if let Some(key) = oldest_key {
+                entries.remove(&key);
+            } else {
+                break;
+            }
+        }
+        entries.insert(
+            key,
+            DnsCacheEntry {
+                addrs,
+                expires_at: now + self.ttl,
+                inserted_at: now,
+            },
+        );
+    }
+
+    fn key(host: &str, port: u16) -> String {
+        format!("{}:{}", host.to_ascii_lowercase(), port)
+    }
+}
+
+/// Resolve a hostname to public (non-private) socket addresses.
+///
+/// Results are cached in `dns_cache`. Private/reserved IPs are filtered out.
+/// Returns an error if no public addresses remain after filtering.
+pub async fn resolve_public_addrs(
+    host: &str,
+    port: u16,
+    dns_cache: &DnsCache,
+) -> Result<Vec<SocketAddr>, FilterError> {
+    // Cache hit
+    if let Some(addrs) = dns_cache.get(host, port).await {
+        return Ok((*addrs).clone());
+    }
+
+    // Async DNS resolution
+    let addr_str = format!("{}:{}", host, port);
+    let resolved: Vec<SocketAddr> = tokio::net::lookup_host(&addr_str)
+        .await
+        .map_err(|_| FilterError::DnsResolutionFailed(host.to_string()))?
+        .collect();
+
+    if resolved.is_empty() {
+        return Err(FilterError::DnsResolutionFailed(host.to_string()));
+    }
+
+    // Filter out private/reserved addresses
+    let public: Vec<SocketAddr> = resolved
+        .into_iter()
+        .filter(|addr| !is_private_ip(&addr.ip()))
+        .collect();
+
+    if public.is_empty() {
+        return Err(FilterError::NoPublicAddrs(host.to_string()));
+    }
+
+    // Cache the validated public addresses
+    let arc_addrs = Arc::new(public);
+    dns_cache.insert(host, port, Arc::clone(&arc_addrs)).await;
+    Ok((*arc_addrs).clone())
+}
+
+/// Validate that the target host:port is allowed.
+///
+/// Performs port whitelist check, private IP filtering, and DNS resolution
+/// with caching. The resolved addresses are stored in the shared DnsCache
+/// so that the SafeDnsResolver can reuse them, eliminating the TOCTTOU gap.
+pub async fn validate_target(
+    host: &str,
+    port: u16,
+    allowed_ports: &HashSet<u16>,
+    dns_cache: &DnsCache,
+) -> Result<Vec<SocketAddr>, FilterError> {
+    // Port whitelist check
+    if !allowed_ports.contains(&port) {
+        return Err(FilterError::PortNotAllowed(port));
+    }
+
+    // Try parsing as IP directly (no DNS needed)
+    if let Ok(ip) = host.parse::<IpAddr>() {
+        if is_private_ip(&ip) {
+            return Err(FilterError::PrivateIp(ip));
+        }
+        return Ok(vec![SocketAddr::new(ip, port)]);
+    }
+
+    // Resolve and validate DNS (populates cache for SafeDnsResolver)
+    resolve_public_addrs(host, port, dns_cache).await
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    fn ports() -> HashSet<u16> {
+        [80, 443, 8080, 8443].into_iter().collect()
+    }
+
+    fn cache() -> DnsCache {
+        DnsCache::new(Duration::from_secs(60), 128)
+    }
+
+    #[test]
+    fn test_private_ipv4() {
+        assert!(is_private_ip(&IpAddr::V4(Ipv4Addr::new(10, 0, 0, 1))));
+        assert!(is_private_ip(&IpAddr::V4(Ipv4Addr::new(172, 16, 0, 1))));
+        assert!(is_private_ip(&IpAddr::V4(Ipv4Addr::new(192, 168, 1, 1))));
+        assert!(is_private_ip(&IpAddr::V4(Ipv4Addr::new(127, 0, 0, 1))));
+        assert!(is_private_ip(&IpAddr::V4(Ipv4Addr::new(169, 254, 1, 1))));
+        assert!(is_private_ip(&IpAddr::V4(Ipv4Addr::new(0, 0, 0, 0))));
+        // CGNAT
+        assert!(is_private_ip(&IpAddr::V4(Ipv4Addr::new(100, 64, 0, 1))));
+        assert!(is_private_ip(&IpAddr::V4(Ipv4Addr::new(
+            100, 127, 255, 254
+        ))));
+        assert!(!is_private_ip(&IpAddr::V4(Ipv4Addr::new(
+            100, 63, 255, 254
+        ))));
+        // Benchmark testing
+        assert!(is_private_ip(&IpAddr::V4(Ipv4Addr::new(198, 18, 0, 1))));
+        // Reserved
+        assert!(is_private_ip(&IpAddr::V4(Ipv4Addr::new(240, 0, 0, 1))));
+        // Public
+        assert!(!is_private_ip(&IpAddr::V4(Ipv4Addr::new(8, 8, 8, 8))));
+        assert!(!is_private_ip(&IpAddr::V4(Ipv4Addr::new(203, 0, 113, 1))));
+    }
+
+    #[test]
+    fn test_private_ipv6() {
+        assert!(is_private_ip(&IpAddr::V6(Ipv6Addr::LOCALHOST)));
+        assert!(is_private_ip(&IpAddr::V6(Ipv6Addr::UNSPECIFIED)));
+        // fc00::1 (ULA)
+        assert!(is_private_ip(&IpAddr::V6(Ipv6Addr::new(
+            0xfc00, 0, 0, 0, 0, 0, 0, 1
+        ))));
+        // fe80::1 (link-local)
+        assert!(is_private_ip(&IpAddr::V6(Ipv6Addr::new(
+            0xfe80, 0, 0, 0, 0, 0, 0, 1
+        ))));
+    }
+
+    #[tokio::test]
+    async fn test_port_not_allowed() {
+        let cache = cache();
+        let result = validate_target("8.8.8.8", 22, &ports(), &cache).await;
+        assert!(matches!(result, Err(FilterError::PortNotAllowed(22))));
+    }
+
+    #[tokio::test]
+    async fn test_private_ip_blocked() {
+        let cache = cache();
+        let result = validate_target("127.0.0.1", 80, &ports(), &cache).await;
+        assert!(matches!(result, Err(FilterError::PrivateIp(_))));
+    }
+
+    #[tokio::test]
+    async fn test_public_ip_allowed() {
+        let cache = cache();
+        let result = validate_target("8.8.8.8", 443, &ports(), &cache).await;
+        assert!(result.is_ok());
+        let addrs = result.unwrap();
+        assert_eq!(addrs.len(), 1);
+        assert_eq!(addrs[0].ip(), IpAddr::V4(Ipv4Addr::new(8, 8, 8, 8)));
+    }
+
+    #[tokio::test]
+    async fn test_cache_stores_multiple_addrs() {
+        let cache = cache();
+        let addrs = vec![
+            SocketAddr::new(IpAddr::V4(Ipv4Addr::new(1, 1, 1, 1)), 443),
+            SocketAddr::new(IpAddr::V4(Ipv4Addr::new(1, 0, 0, 1)), 443),
+        ];
+        cache
+            .insert("example.com", 443, Arc::new(addrs.clone()))
+            .await;
+        let cached = cache.get("example.com", 443).await.unwrap();
+        assert_eq!(*cached, addrs);
+    }
+
+    #[tokio::test]
+    async fn test_cache_key_case_insensitive() {
+        let cache = cache();
+        let addrs = vec![SocketAddr::new(IpAddr::V4(Ipv4Addr::new(1, 1, 1, 1)), 443)];
+        cache
+            .insert("Example.COM", 443, Arc::new(addrs.clone()))
+            .await;
+        let cached = cache.get("example.com", 443).await.unwrap();
+        assert_eq!(*cached, addrs);
+    }
+}
@@ -0,0 +1,238 @@
+//! WebSocket tunnel client: connect, authenticate, and run the tunnel.
+
+use std::sync::Arc;
+use std::time::Duration;
+
+use tokio::net::TcpStream;
+use tokio::sync::watch;
+use tokio_tungstenite::tungstenite::client::IntoClientRequest;
+use tokio_tungstenite::tungstenite::http;
+use tokio_tungstenite::tungstenite::protocol::WebSocketConfig;
+use tracing::{debug, info, warn};
+
+use crate::state::{AppState, ServerContext};
+
+use super::{dispatcher, heartbeat, writer};
+
+/// Outcome of a tunnel session.
+pub enum TunnelOutcome {
+    /// Graceful shutdown requested by the local process.
+    Shutdown,
+    /// Remote side disconnected or connection lost — should reconnect.
+    Disconnected,
+}
+
+/// Connect to Aether's WebSocket tunnel endpoint and run until disconnected.
+///
+/// `conn_idx` identifies which connection in the pool this is (0-based).
+/// Only connection 0 sends heartbeats to avoid resetting shared metrics.
+pub async fn connect_and_run(
+    state: &Arc<AppState>,
+    server: &Arc<ServerContext>,
+    conn_idx: usize,
+    shutdown: &mut watch::Receiver<bool>,
+) -> Result<TunnelOutcome, anyhow::Error> {
+    let ws_url = build_tunnel_url(server);
+    info!(url = %ws_url, conn = conn_idx, "connecting tunnel");
+
+    // Build WebSocket request with auth headers
+    let mut request = ws_url.clone().into_client_request()?;
+    let headers = request.headers_mut();
+    headers.insert(
+        "Authorization",
+        http::HeaderValue::from_str(&format!("Bearer {}", server.management_token))?,
+    );
+    let node_id = server.node_id.read().unwrap().clone();
+    headers.insert("X-Node-Id", http::HeaderValue::from_str(&node_id)?);
+    // Use dynamic node_name (may be updated by remote config) instead of
+    // the static server.node_name, so that remote name changes take effect
+    // on the next reconnect.
+    let dynamic_node_name = server.dynamic.load().node_name.clone();
+    headers.insert(
+        "X-Node-Name",
+        http::HeaderValue::from_str(&dynamic_node_name)?,
+    );
+    // Advertise per-connection max concurrent streams so the backend can
+    // respect the proxy's capacity limit (backward-compatible: old backends
+    // ignore this header).
+    let max_streams = state.config.tunnel_max_streams.unwrap_or(128);
+    headers.insert("X-Tunnel-Max-Streams", http::HeaderValue::from(max_streams));
+
+    // Parse host:port from URL
+    let uri: http::Uri = ws_url.parse()?;
+    let host = uri
+        .host()
+        .ok_or_else(|| anyhow::anyhow!("missing host in tunnel URL"))?;
+    let is_tls = uri.scheme_str() == Some("wss");
+    let port = uri.port_u16().unwrap_or(if is_tls { 443 } else { 80 });
+
+    // TCP connect with timeout
+    let connect_timeout = Duration::from_secs(state.config.tunnel_connect_timeout_secs);
+    let tcp_stream = tokio::time::timeout(connect_timeout, TcpStream::connect((host, port)))
+        .await
+        .map_err(|_| {
+            anyhow::anyhow!(
+                "tunnel TCP connect timeout ({}s)",
+                connect_timeout.as_secs()
+            )
+        })??;
+
+    // Configure TCP parameters via socket2
+    configure_tcp_socket(&tcp_stream, state);
+
+    // WebSocket upgrade (with TLS if wss://)
+    let connector = if is_tls {
+        Some(tokio_tungstenite::Connector::Rustls(Arc::clone(
+            &state.tunnel_tls_config,
+        )))
+    } else {
+        None
+    };
+    // Match Python-side _MAX_FRAME_SIZE (64 MiB) to prevent tungstenite's
+    // default 16 MiB limit from rejecting large AI API payloads (multi-image
+    // base64 requests can exceed 16 MiB).
+    let ws_config = WebSocketConfig {
+        max_frame_size: Some(64 << 20),
+        max_message_size: Some(64 << 20),
+        ..Default::default()
+    };
+    let handshake_timeout = Duration::from_secs(state.config.tunnel_connect_timeout_secs);
+    let (ws_stream, _response) = tokio::time::timeout(
+        handshake_timeout,
+        tokio_tungstenite::client_async_tls_with_config(
+            request,
+            tcp_stream,
+            Some(ws_config),
+            connector,
+        ),
+    )
+    .await
+    .map_err(|_| {
+        anyhow::anyhow!(
+            "tunnel WebSocket handshake timeout ({}s)",
+            handshake_timeout.as_secs()
+        )
+    })??;
+    info!(
+        conn = conn_idx,
+        tcp_keepalive_secs = state.config.tunnel_tcp_keepalive_secs,
+        tcp_nodelay = state.config.tunnel_tcp_nodelay,
+        connect_timeout_secs = state.config.tunnel_connect_timeout_secs,
+        stale_timeout_secs = state.config.tunnel_stale_timeout_secs,
+        "tunnel connected"
+    );
+
+    // NOTE: reconnect_attempts reset is handled by the caller (mod.rs)
+    // based on how long the connection stayed alive.
+
+    // Split into read/write halves
+    let (ws_sink, ws_read) = futures_util::StreamExt::split(ws_stream);
+
+    // Spawn writer task (with WebSocket ping keepalive)
+    let ping_interval = Duration::from_secs(state.config.tunnel_ping_interval_secs);
+    let (frame_tx, mut writer_handle) = writer::spawn_writer(ws_sink, ping_interval);
+
+    // Spawn heartbeat task (only for primary connection to avoid
+    // resetting shared atomic metrics via swap(0))
+    let hb_handle = if conn_idx == 0 {
+        heartbeat::spawn(
+            Arc::clone(&state.config),
+            Arc::clone(server),
+            frame_tx.clone(),
+            shutdown.clone(),
+        )
+    } else {
+        heartbeat::spawn_noop()
+    };
+
+    // Run dispatcher (blocks until disconnect or shutdown).
+    // Also watch for writer exit — if the write half dies (e.g. the peer
+    // closed the connection) but the read half stays open, dispatcher would
+    // block forever on `ws_stream.next()`.  Monitoring `writer_handle`
+    // ensures we detect this and trigger a reconnect promptly.
+    let state_clone = Arc::clone(state);
+    let server_clone = Arc::clone(server);
+    let outcome = tokio::select! {
+        result = dispatcher::run(state_clone, server_clone, ws_read, frame_tx.clone(), hb_handle) => {
+            match result {
+                Ok(()) => TunnelOutcome::Disconnected,
+                Err(e) => return Err(e),
+            }
+        }
+        writer_result = &mut writer_handle => {
+            match writer_result {
+                Ok(()) => warn!("writer task exited normally, triggering reconnect"),
+                Err(e) => {
+                    if e.is_panic() {
+                        tracing::error!(error = %e, "writer task panicked, triggering reconnect");
+                    } else {
+                        warn!(error = %e, "writer task cancelled, triggering reconnect");
+                    }
+                }
+            }
+            TunnelOutcome::Disconnected
+        }
+        _ = shutdown.changed() => {
+            debug!("shutdown during tunnel dispatch");
+            TunnelOutcome::Shutdown
+        }
+    };
+
+    // Drop our sender; the writer will exit once all stream handler clones
+    // are also dropped (i.e. after they finish their in-flight work).
+    drop(frame_tx);
+
+    // Wait for the writer task to finish with a generous timeout — the
+    // dispatcher already waits up to 30s for stream handlers, so 35s here
+    // covers that plus a small margin.
+    // Skip if the writer already exited (the select branch that fired).
+    if !writer_handle.is_finished() {
+        let _ = tokio::time::timeout(Duration::from_secs(35), writer_handle).await;
+    }
+
+    info!("tunnel disconnected");
+    Ok(outcome)
+}
+
+/// Configure TCP keepalive and NODELAY on an established socket.
+fn configure_tcp_socket(stream: &TcpStream, state: &Arc<AppState>) {
+    let sock_ref = socket2::SockRef::from(stream);
+
+    if state.config.tunnel_tcp_keepalive_secs > 0 {
+        let keepalive = socket2::TcpKeepalive::new()
+            .with_time(Duration::from_secs(state.config.tunnel_tcp_keepalive_secs))
+            .with_interval(Duration::from_secs(5));
+        #[cfg(not(target_os = "windows"))]
+        let keepalive = keepalive.with_retries(3);
+        if let Err(e) = sock_ref.set_tcp_keepalive(&keepalive) {
+            warn!(error = %e, "failed to set TCP keepalive on tunnel socket");
+        }
+    }
+
+    if state.config.tunnel_tcp_nodelay {
+        if let Err(e) = sock_ref.set_nodelay(true) {
+            warn!(error = %e, "failed to set TCP_NODELAY on tunnel socket");
+        }
+    }
+}
+
+/// Build rustls ClientConfig with system root certificates.
+pub fn build_tls_config() -> rustls::ClientConfig {
+    let root_store =
+        rustls::RootCertStore::from_iter(webpki_roots::TLS_SERVER_ROOTS.iter().cloned());
+    rustls::ClientConfig::builder()
+        .with_root_certificates(root_store)
+        .with_no_client_auth()
+}
+
+fn build_tunnel_url(server: &ServerContext) -> String {
+    let base = server.aether_url.trim_end_matches('/');
+    let ws_base = if base.starts_with("https://") {
+        base.replacen("https://", "wss://", 1)
+    } else if base.starts_with("http://") {
+        base.replacen("http://", "ws://", 1)
+    } else {
+        format!("wss://{}", base)
+    };
+    format!("{}/api/internal/proxy-tunnel", ws_base)
+}
@@ -0,0 +1,249 @@
+//! Frame dispatcher: reads incoming WebSocket frames and routes them.
+
+use std::collections::HashMap;
+use std::sync::Arc;
+use std::time::Duration;
+
+use bytes::Bytes;
+use futures_util::StreamExt;
+use tokio::sync::mpsc;
+use tokio::task::JoinHandle;
+use tokio_tungstenite::tungstenite::Message;
+use tracing::{debug, error, info, warn};
+
+use crate::state::{AppState, ServerContext};
+
+use super::heartbeat::HeartbeatHandle;
+use super::protocol::{decompress_if_gzip, Frame, MsgType, RequestMeta};
+use super::stream_handler;
+use super::writer::FrameSender;
+
+/// Run the dispatcher loop, reading from the WebSocket stream.
+pub async fn run<S>(
+    state: Arc<AppState>,
+    server: Arc<ServerContext>,
+    mut ws_stream: S,
+    frame_tx: FrameSender,
+    heartbeat: HeartbeatHandle,
+) -> Result<(), anyhow::Error>
+where
+    S: StreamExt<Item = Result<Message, tokio_tungstenite::tungstenite::Error>>
+        + Unpin
+        + Send
+        + 'static,
+{
+    // Active streams: stream_id -> body sender
+    let mut streams: HashMap<u32, mpsc::Sender<Frame>> = HashMap::new();
+    // Track spawned stream handlers so we can wait for them on shutdown
+    let mut handler_handles: Vec<JoinHandle<()>> = Vec::new();
+    let max_streams = state.config.tunnel_max_streams.unwrap_or(128) as usize;
+    let mut frames_since_cleanup: u32 = 0;
+    let stale_timeout = Duration::from_secs(state.config.tunnel_stale_timeout_secs);
+
+    // Track last time we received any data to detect stale connections
+    let mut last_data_at = tokio::time::Instant::now();
+
+    let read_err = loop {
+        let msg_result = tokio::select! {
+            msg = ws_stream.next() => {
+                match msg {
+                    Some(r) => r,
+                    None => break None,
+                }
+            }
+            _ = tokio::time::sleep_until(last_data_at + stale_timeout) => {
+                warn!(
+                    stale_secs = stale_timeout.as_secs(),
+                    "tunnel connection stale, no data received"
+                );
+                break None;
+            }
+        };
+
+        let msg = match msg_result {
+            Ok(m) => m,
+            Err(e) => {
+                error!(error = %e, "WebSocket read error");
+                break Some(e);
+            }
+        };
+
+        // Any successfully received message proves the connection is alive
+        last_data_at = tokio::time::Instant::now();
+
+        let data = match msg {
+            Message::Binary(data) => Bytes::from(data),
+            Message::Ping(_) => continue,
+            Message::Pong(_) => continue,
+            Message::Close(_) => {
+                info!("received WebSocket close");
+                break None;
+            }
+            _ => continue,
+        };
+
+        let frame = match Frame::decode(data) {
+            Ok(f) => f,
+            Err(e) => {
+                warn!(error = %e, "failed to decode frame");
+                continue;
+            }
+        };
+
+        match frame.msg_type {
+            MsgType::RequestHeaders => {
+                // Decompress if the frame is gzip-compressed, then parse metadata
+                let payload = match decompress_if_gzip(&frame) {
+                    Ok(p) => p,
+                    Err(e) => {
+                        warn!(stream_id = frame.stream_id, error = %e, "frame decompress failed");
+                        continue;
+                    }
+                };
+                let meta: RequestMeta = match serde_json::from_slice(&payload) {
+                    Ok(m) => m,
+                    Err(e) => {
+                        warn!(stream_id = frame.stream_id, error = %e, "invalid request metadata");
+                        // Use try_send to avoid blocking the read loop
+                        if frame_tx
+                            .try_send(Frame::new(
+                                frame.stream_id,
+                                MsgType::StreamError,
+                                0,
+                                Bytes::from(format!("invalid request metadata: {e}")),
+                            ))
+                            .is_err()
+                        {
+                            warn!(
+                                stream_id = frame.stream_id,
+                                "writer channel full, StreamError dropped"
+                            );
+                        }
+                        continue;
+                    }
+                };
+
+                if streams.len() >= max_streams {
+                    warn!(
+                        stream_id = frame.stream_id,
+                        "max concurrent streams reached"
+                    );
+                    if frame_tx
+                        .try_send(Frame::new(
+                            frame.stream_id,
+                            MsgType::StreamError,
+                            0,
+                            Bytes::from("max concurrent streams reached"),
+                        ))
+                        .is_err()
+                    {
+                        warn!(
+                            stream_id = frame.stream_id,
+                            "writer channel full, StreamError dropped"
+                        );
+                    }
+                    continue;
+                }
+
+                // Create body channel and spawn handler
+                let (body_tx, body_rx) = mpsc::channel::<Frame>(64);
+                streams.insert(frame.stream_id, body_tx);
+
+                let state_clone = Arc::clone(&state);
+                let server_clone = Arc::clone(&server);
+                let tx_clone = frame_tx.clone();
+                let sid = frame.stream_id;
+                let handle = tokio::spawn(async move {
+                    stream_handler::handle_stream(
+                        state_clone,
+                        server_clone,
+                        sid,
+                        meta,
+                        body_rx,
+                        tx_clone,
+                    )
+                    .await;
+                });
+                handler_handles.push(handle);
+
+                debug!(stream_id = frame.stream_id, "new stream started");
+            }
+
+            MsgType::RequestBody => {
+                if let Some(tx) = streams.get(&frame.stream_id) {
+                    let is_end = frame.is_end_stream();
+                    let sid = frame.stream_id;
+                    let _ = tx.send(frame).await;
+                    if is_end {
+                        streams.remove(&sid);
+                    }
+                }
+            }
+
+            MsgType::StreamEnd | MsgType::StreamError => {
+                // Client-side cancellation or end
+                if let Some(tx) = streams.remove(&frame.stream_id) {
+                    let _ = tx.send(frame).await;
+                }
+            }
+
+            MsgType::Ping => {
+                // Use try_send to avoid blocking the read loop when writer is congested
+                if frame_tx
+                    .try_send(Frame::control(MsgType::Pong, frame.payload))
+                    .is_err()
+                {
+                    warn!("writer channel full, Pong dropped");
+                }
+            }
+
+            MsgType::HeartbeatAck => {
+                heartbeat.on_ack(frame.payload).await;
+            }
+
+            MsgType::GoAway => {
+                info!("received GOAWAY");
+                break None;
+            }
+
+            _ => {
+                debug!(msg_type = ?frame.msg_type, "ignoring unexpected frame type");
+            }
+        }
+
+        // Periodically clean up finished handles to avoid unbounded growth.
+        // Trigger every 64 frames OR when the count exceeds max_streams.
+        frames_since_cleanup += 1;
+        if frames_since_cleanup >= 64 || handler_handles.len() > max_streams {
+            handler_handles.retain(|h| !h.is_finished());
+            frames_since_cleanup = 0;
+        }
+    };
+
+    // Drop body senders so stream handlers waiting on body_rx will unblock
+    streams.clear();
+
+    // Wait for active stream handlers to finish so their frame_tx clones
+    // are dropped before the writer closes the sink.
+    drain_handlers(handler_handles).await;
+
+    match read_err {
+        Some(e) => Err(e.into()),
+        None => Ok(()),
+    }
+}
+
+/// Wait for all active stream handlers to finish (with a timeout).
+async fn drain_handlers(handles: Vec<JoinHandle<()>>) {
+    if handles.is_empty() {
+        return;
+    }
+    let count = handles.len();
+    debug!(count, "waiting for active stream handlers to finish");
+    let _ = tokio::time::timeout(Duration::from_secs(30), async {
+        for h in handles {
+            let _ = h.await;
+        }
+    })
+    .await;
+}
@@ -0,0 +1,340 @@
+//! Tunnel heartbeat: sends metrics over the tunnel, processes ACKs.
+
+use std::sync::atomic::{AtomicBool, Ordering};
+use std::sync::Arc;
+use std::time::Duration;
+use std::time::SystemTime;
+use std::time::UNIX_EPOCH;
+
+use bytes::Bytes;
+use tokio::sync::watch;
+use tracing::{debug, info, warn};
+
+use crate::config::Config;
+use crate::registration::client::RemoteConfig;
+use crate::runtime;
+use crate::state::ServerContext;
+
+use super::protocol::{Frame, MsgType};
+use super::writer::FrameSender;
+
+const CURRENT_VERSION: &str = env!("CARGO_PKG_VERSION");
+static UPGRADE_IN_PROGRESS: AtomicBool = AtomicBool::new(false);
+static NON_ROOT_UPGRADE_WARNED: AtomicBool = AtomicBool::new(false);
+
+enum AckDecision {
+    Accept {
+        heartbeat_id: Option<u64>,
+        upgrade_to: Option<String>,
+    },
+    Ignore,
+}
+
+/// Handle for the dispatcher to forward HeartbeatAck frames.
+#[derive(Clone)]
+pub struct HeartbeatHandle {
+    ack_tx: tokio::sync::mpsc::Sender<Bytes>,
+}
+
+impl HeartbeatHandle {
+    pub async fn on_ack(&self, payload: Bytes) {
+        let _ = self.ack_tx.send(payload).await;
+    }
+}
+
+/// Create a no-op heartbeat handle that silently discards ACKs.
+/// Used for non-primary tunnel connections (conn_idx > 0) to avoid
+/// resetting shared atomic metrics via `swap(0)`.
+pub fn spawn_noop() -> HeartbeatHandle {
+    let (ack_tx, _) = tokio::sync::mpsc::channel::<Bytes>(1);
+    // receiver is immediately dropped; on_ack() calls will silently fail
+    HeartbeatHandle { ack_tx }
+}
+
+#[derive(Debug, Clone, Copy, Default)]
+struct HeartbeatSnapshot {
+    requests: u64,
+    latency_ns: u64,
+    failed: u64,
+    dns_failures: u64,
+    stream_errors: u64,
+}
+
+/// Spawn the heartbeat task. Returns a handle for forwarding ACKs.
+pub fn spawn(
+    _config: Arc<Config>,
+    server: Arc<ServerContext>,
+    frame_tx: FrameSender,
+    mut shutdown: watch::Receiver<bool>,
+) -> HeartbeatHandle {
+    let (ack_tx, mut ack_rx) = tokio::sync::mpsc::channel::<Bytes>(4);
+
+    tokio::spawn(async move {
+        // Read initial interval from dynamic config (may be updated by remote config).
+        let initial_interval = Duration::from_secs(server.dynamic.load().heartbeat_interval);
+        let mut current_interval = initial_interval;
+        // At most one in-flight heartbeat snapshot is tracked at a time.
+        // Snapshot is only cleared after receiving an ACK, which avoids losing
+        // interval counters when ACK/frame delivery is temporarily unstable.
+        let mut pending: Option<(u64, HeartbeatSnapshot)> = None;
+        let mut next_heartbeat_id: u64 = 1;
+        let heartbeat_session_id = format!(
+            "{}-{}",
+            std::process::id(),
+            SystemTime::now()
+                .duration_since(UNIX_EPOCH)
+                .unwrap_or_default()
+                .as_nanos()
+        );
+
+        // Skip first immediate tick by sleeping first.
+        tokio::time::sleep(current_interval).await;
+
+        loop {
+            tokio::select! {
+                _ = tokio::time::sleep(current_interval) => {
+                    let (heartbeat_id, snapshot) = if let Some((id, snap)) = pending {
+                        (id, snap)
+                    } else {
+                        let snap = collect_snapshot(&server);
+                        let id = next_heartbeat_id;
+                        next_heartbeat_id = next_heartbeat_id.wrapping_add(1);
+                        if next_heartbeat_id == 0 {
+                            next_heartbeat_id = 1;
+                        }
+                        pending = Some((id, snap));
+                        (id, snap)
+                    };
+
+                    let payload = build_heartbeat_payload(
+                        &server,
+                        &heartbeat_session_id,
+                        heartbeat_id,
+                        snapshot
+                    );
+                    let frame = Frame::control(MsgType::HeartbeatData, payload);
+                    if frame_tx.send(frame).await.is_err() {
+                        if let Some((_, snap)) = pending.take() {
+                            restore_snapshot(&server, snap);
+                        }
+                        break; // Writer closed
+                    }
+                    debug!("sent heartbeat data");
+
+                    // Re-read interval from dynamic config (remote config may have
+                    // updated it since the last heartbeat).
+                    let new_interval = Duration::from_secs(
+                        server.dynamic.load().heartbeat_interval
+                    );
+                    if new_interval != current_interval {
+                        debug!(
+                            old_secs = current_interval.as_secs(),
+                            new_secs = new_interval.as_secs(),
+                            "heartbeat interval updated from dynamic config"
+                        );
+                        current_interval = new_interval;
+                    }
+                }
+                Some(ack_payload) = ack_rx.recv() => {
+                    match handle_ack(&server, &ack_payload) {
+                        AckDecision::Accept {
+                            heartbeat_id: ack_id,
+                            upgrade_to,
+                        } => {
+                            if let Some((pending_id, _)) = pending {
+                                match ack_id {
+                                    Some(id) if id == pending_id => {
+                                        pending = None;
+                                    }
+                                    None => {
+                                        // Backward-compatible with servers that don't echo
+                                        // heartbeat_id in ACK payload yet.
+                                        pending = None;
+                                    }
+                                    _ => {}
+                                }
+                            }
+                            maybe_trigger_upgrade(upgrade_to);
+                        }
+                        AckDecision::Ignore => {}
+                    }
+                }
+                _ = shutdown.changed() => {
+                    debug!("heartbeat task shutting down");
+                    if let Some((_, snap)) = pending.take() {
+                        restore_snapshot(&server, snap);
+                    }
+                    break;
+                }
+            }
+        }
+    });
+
+    HeartbeatHandle { ack_tx }
+}
+
+fn collect_snapshot(server: &ServerContext) -> HeartbeatSnapshot {
+    HeartbeatSnapshot {
+        requests: server.metrics.total_requests.swap(0, Ordering::AcqRel),
+        latency_ns: server.metrics.total_latency_ns.swap(0, Ordering::AcqRel),
+        failed: server.metrics.failed_requests.swap(0, Ordering::AcqRel),
+        dns_failures: server.metrics.dns_failures.swap(0, Ordering::AcqRel),
+        stream_errors: server.metrics.stream_errors.swap(0, Ordering::AcqRel),
+    }
+}
+
+fn restore_snapshot(server: &ServerContext, snap: HeartbeatSnapshot) {
+    if snap.requests > 0 {
+        server
+            .metrics
+            .total_requests
+            .fetch_add(snap.requests, Ordering::Release);
+    }
+    if snap.latency_ns > 0 {
+        server
+            .metrics
+            .total_latency_ns
+            .fetch_add(snap.latency_ns, Ordering::Release);
+    }
+    if snap.failed > 0 {
+        server
+            .metrics
+            .failed_requests
+            .fetch_add(snap.failed, Ordering::Release);
+    }
+    if snap.dns_failures > 0 {
+        server
+            .metrics
+            .dns_failures
+            .fetch_add(snap.dns_failures, Ordering::Release);
+    }
+    if snap.stream_errors > 0 {
+        server
+            .metrics
+            .stream_errors
+            .fetch_add(snap.stream_errors, Ordering::Release);
+    }
+}
+
+fn build_heartbeat_payload(
+    server: &ServerContext,
+    heartbeat_session_id: &str,
+    heartbeat_id: u64,
+    snapshot: HeartbeatSnapshot,
+) -> Bytes {
+    let node_id = server.node_id.read().unwrap().clone();
+
+    let avg_latency_ms = if snapshot.requests > 0 {
+        Some(snapshot.latency_ns as f64 / snapshot.requests as f64 / 1_000_000.0)
+    } else {
+        None
+    };
+
+    let payload = serde_json::json!({
+        "node_id": node_id,
+        "heartbeat_session_id": heartbeat_session_id,
+        "heartbeat_id": heartbeat_id,
+        "active_connections": server.active_connections.load(Ordering::Acquire),
+        "total_requests": snapshot.requests,
+        "avg_latency_ms": avg_latency_ms,
+        "failed_requests": snapshot.failed,
+        "dns_failures": snapshot.dns_failures,
+        "stream_errors": snapshot.stream_errors,
+        "proxy_metadata": {
+            "version": CURRENT_VERSION,
+        },
+    });
+
+    Bytes::from(serde_json::to_vec(&payload).unwrap_or_default())
+}
+
+fn handle_ack(server: &ServerContext, payload: &[u8]) -> AckDecision {
+    if payload.is_empty() {
+        return AckDecision::Accept {
+            heartbeat_id: None,
+            upgrade_to: None,
+        };
+    }
+
+    #[derive(serde::Deserialize)]
+    struct AckPayload {
+        #[serde(default)]
+        remote_config: Option<RemoteConfig>,
+        #[serde(default)]
+        config_version: u64,
+        #[serde(default)]
+        heartbeat_id: Option<u64>,
+        #[serde(default)]
+        upgrade_to: Option<String>,
+    }
+
+    match serde_json::from_slice::<AckPayload>(payload) {
+        Ok(ack) => {
+            if let Some(ref rc) = ack.remote_config {
+                runtime::apply_remote_config(&server.dynamic, rc, ack.config_version);
+            }
+            AckDecision::Accept {
+                heartbeat_id: ack.heartbeat_id,
+                upgrade_to: ack.upgrade_to.and_then(normalize_upgrade_target),
+            }
+        }
+        Err(e) => {
+            warn!(error = %e, "failed to parse heartbeat ACK");
+            AckDecision::Ignore
+        }
+    }
+}
+
+fn normalize_upgrade_target(raw: String) -> Option<String> {
+    let trimmed = raw.trim();
+    if trimmed.is_empty() {
+        return None;
+    }
+    let normalized = trimmed.strip_prefix("proxy-v").unwrap_or(trimmed);
+    if normalized == CURRENT_VERSION {
+        return None;
+    }
+    Some(normalized.to_string())
+}
+
+fn maybe_trigger_upgrade(version: Option<String>) {
+    let Some(target_version) = version else {
+        return;
+    };
+    if !crate::setup::service::is_root() {
+        if NON_ROOT_UPGRADE_WARNED
+            .compare_exchange(false, true, Ordering::AcqRel, Ordering::Acquire)
+            .is_ok()
+        {
+            warn!(
+                target_version = %target_version,
+                "remote upgrade skipped: root privileges are required"
+            );
+        }
+        return;
+    }
+    if UPGRADE_IN_PROGRESS
+        .compare_exchange(false, true, Ordering::AcqRel, Ordering::Acquire)
+        .is_err()
+    {
+        debug!(target_version = %target_version, "upgrade already in progress, ignoring");
+        return;
+    }
+
+    tokio::spawn(async move {
+        info!(target_version = %target_version, "received remote upgrade instruction");
+        match crate::setup::upgrade::perform_upgrade(&target_version).await {
+            Ok(()) => {
+                info!(target_version = %target_version, "remote upgrade finished");
+            }
+            Err(e) => {
+                warn!(
+                    target_version = %target_version,
+                    error = %e,
+                    "remote upgrade failed"
+                );
+                UPGRADE_IN_PROGRESS.store(false, Ordering::Release);
+            }
+        }
+    });
+}
@@ -0,0 +1,236 @@
+pub mod client;
+pub mod dispatcher;
+pub mod heartbeat;
+pub mod protocol;
+pub mod stream_handler;
+pub mod writer;
+
+use std::sync::Arc;
+use std::time::{Duration, Instant, SystemTime, UNIX_EPOCH};
+
+use tokio::sync::watch;
+use tracing::{error, info};
+
+use crate::state::{AppState, ServerContext};
+
+/// If a tunnel stays connected at least this long, treat the next disconnect
+/// as a non-failure and reset reconnect backoff.
+const STABLE_SESSION_RESET_AFTER: Duration = Duration::from_secs(30);
+/// Startup staggering step per secondary connection, used to avoid
+/// simultaneous bursts when a pool of tunnels starts together.
+const STARTUP_STAGGER_STEP_MS: u64 = 150;
+/// Upper bound for startup staggering.
+const MAX_STARTUP_STAGGER_MS: u64 = 1_500;
+/// Keep a tiny floor for repeated reconnects; first retry is still immediate.
+const MIN_RECONNECT_DELAY_MS: u64 = 50;
+/// Even under sustained failures, keep probing frequently so recovery is fast
+/// once cross-border network quality improves.
+const RECONNECT_PROBE_MAX_DELAY_MS: u64 = 3_000;
+
+/// Run the tunnel mode main loop (connect, dispatch, reconnect).
+///
+/// `conn_idx` identifies which connection in the pool this is (0-based).
+/// Only connection 0 sends heartbeats to avoid resetting shared metrics.
+pub async fn run(
+    state: &Arc<AppState>,
+    server: &Arc<ServerContext>,
+    conn_idx: usize,
+    mut shutdown: watch::Receiver<bool>,
+) {
+    info!(server = %server.server_label, conn = conn_idx, "starting tunnel");
+    let reconnect_salt = compute_connection_salt(server, conn_idx);
+
+    let startup_delay = compute_startup_stagger(conn_idx, reconnect_salt);
+    if !startup_delay.is_zero() {
+        info!(
+            server = %server.server_label,
+            conn = conn_idx,
+            delay_ms = startup_delay.as_millis(),
+            "startup stagger before first connect"
+        );
+        tokio::select! {
+            _ = tokio::time::sleep(startup_delay) => {}
+            _ = shutdown.changed() => {
+                info!(server = %server.server_label, conn = conn_idx, "shutdown requested during startup stagger");
+                return;
+            }
+        }
+    }
+
+    let mut consecutive_failures: u32 = 0;
+
+    loop {
+        let started_at = Instant::now();
+        match client::connect_and_run(state, server, conn_idx, &mut shutdown).await {
+            Ok(client::TunnelOutcome::Shutdown) => {
+                info!(server = %server.server_label, conn = conn_idx, "tunnel shut down gracefully");
+                return;
+            }
+            Ok(client::TunnelOutcome::Disconnected) => {
+                info!(server = %server.server_label, conn = conn_idx, "tunnel disconnected, reconnecting");
+            }
+            Err(e) => {
+                error!(server = %server.server_label, conn = conn_idx, error = %e, "tunnel connection error, reconnecting");
+            }
+        }
+
+        if *shutdown.borrow() {
+            info!(server = %server.server_label, conn = conn_idx, "shutdown requested, not reconnecting");
+            return;
+        }
+
+        // Reset backoff after a stable session to keep recovery snappy when
+        // failures are only occasional.
+        let connected_for = started_at.elapsed();
+        if connected_for >= STABLE_SESSION_RESET_AFTER {
+            consecutive_failures = 0;
+        } else {
+            consecutive_failures = consecutive_failures.saturating_add(1);
+        }
+
+        let reconnect_delay = compute_reconnect_delay(
+            state.config.tunnel_reconnect_base_ms,
+            state.config.tunnel_reconnect_max_ms,
+            consecutive_failures,
+            reconnect_salt,
+        );
+        info!(
+            server = %server.server_label,
+            conn = conn_idx,
+            failures = consecutive_failures,
+            delay_ms = reconnect_delay.as_millis(),
+            "waiting before reconnect"
+        );
+
+        tokio::select! {
+            _ = tokio::time::sleep(reconnect_delay) => {}
+            _ = shutdown.changed() => {
+                info!(server = %server.server_label, conn = conn_idx, "shutdown requested during reconnect wait");
+                return;
+            }
+        }
+    }
+}
+
+fn compute_connection_salt(server: &ServerContext, conn_idx: usize) -> u64 {
+    // FNV-1a style hash over server label + connection index.
+    let mut h: u64 = 0xcbf29ce484222325;
+    for &b in server.server_label.as_bytes() {
+        h ^= b as u64;
+        h = h.wrapping_mul(0x100000001b3);
+    }
+    h ^= conn_idx as u64;
+    mix_u64(h)
+}
+
+fn compute_startup_stagger(conn_idx: usize, salt: u64) -> Duration {
+    if conn_idx == 0 {
+        return Duration::ZERO;
+    }
+    let base = (conn_idx as u64).saturating_mul(STARTUP_STAGGER_STEP_MS);
+    let jitter = mix_u64(salt) % 301; // 0..=300ms
+    Duration::from_millis((base + jitter).min(MAX_STARTUP_STAGGER_MS))
+}
+
+fn compute_reconnect_delay(
+    base_ms: u64,
+    max_ms: u64,
+    consecutive_failures: u32,
+    salt: u64,
+) -> Duration {
+    // First retry should be immediate to maximize recovery speed on transient
+    // blips (the user's primary expectation in poor networks).
+    if consecutive_failures <= 1 {
+        return Duration::ZERO;
+    }
+
+    // Keep a sane minimum for repeated failures.
+    let base_ms = base_ms.max(MIN_RECONNECT_DELAY_MS);
+    let max_ms = max_ms.max(base_ms);
+    let cap_ms = compute_reconnect_cap_ms(base_ms, max_ms, consecutive_failures)
+        .min(RECONNECT_PROBE_MAX_DELAY_MS.max(base_ms));
+
+    // Equal-jitter: randomize in [cap/2, cap], preventing synchronized reconnect
+    // storms while keeping reconnect latency bounded.
+    if cap_ms <= 1 {
+        return Duration::from_millis(cap_ms);
+    }
+
+    let half = cap_ms / 2;
+    let span = cap_ms - half;
+    let now_nanos = SystemTime::now()
+        .duration_since(UNIX_EPOCH)
+        .map(|d| d.subsec_nanos() as u64)
+        .unwrap_or(0);
+    let mixed = mix_u64(now_nanos ^ salt);
+    let jitter = if span == 0 { 0 } else { mixed % (span + 1) };
+    Duration::from_millis(half + jitter)
+}
+
+fn compute_reconnect_cap_ms(base_ms: u64, max_ms: u64, consecutive_failures: u32) -> u64 {
+    if consecutive_failures <= 1 {
+        return base_ms.min(max_ms);
+    }
+
+    let shift = (consecutive_failures - 1).min(31);
+    let factor = 1u64 << shift;
+    base_ms.saturating_mul(factor).min(max_ms)
+}
+
+fn mix_u64(mut x: u64) -> u64 {
+    // SplitMix64 finalizer - cheap bit mixing for pseudo-random jitter.
+    x ^= x >> 30;
+    x = x.wrapping_mul(0xbf58476d1ce4e5b9);
+    x ^= x >> 27;
+    x = x.wrapping_mul(0x94d049bb133111eb);
+    x ^ (x >> 31)
+}
+
+#[cfg(test)]
+mod tests {
+    use std::time::Duration;
+
+    use super::{
+        compute_reconnect_cap_ms, compute_reconnect_delay, compute_startup_stagger,
+        MAX_STARTUP_STAGGER_MS, RECONNECT_PROBE_MAX_DELAY_MS, STARTUP_STAGGER_STEP_MS,
+    };
+
+    #[test]
+    fn reconnect_cap_grows_exponentially_and_caps() {
+        let base = 500;
+        let max = 30_000;
+        assert_eq!(compute_reconnect_cap_ms(base, max, 0), 500);
+        assert_eq!(compute_reconnect_cap_ms(base, max, 1), 500);
+        assert_eq!(compute_reconnect_cap_ms(base, max, 2), 1_000);
+        assert_eq!(compute_reconnect_cap_ms(base, max, 3), 2_000);
+        assert_eq!(compute_reconnect_cap_ms(base, max, 4), 4_000);
+        assert_eq!(compute_reconnect_cap_ms(base, max, 5), 8_000);
+        assert_eq!(compute_reconnect_cap_ms(base, max, 6), 16_000);
+        assert_eq!(compute_reconnect_cap_ms(base, max, 7), 30_000);
+        assert_eq!(compute_reconnect_cap_ms(base, max, 20), 30_000);
+    }
+
+    #[test]
+    fn startup_stagger_is_zero_for_primary_and_bounded_for_secondary() {
+        assert_eq!(compute_startup_stagger(0, 42), Duration::ZERO);
+
+        let d1 = compute_startup_stagger(1, 42);
+        let d2 = compute_startup_stagger(2, 42);
+
+        assert!(d1 >= Duration::from_millis(STARTUP_STAGGER_STEP_MS));
+        assert!(d1 <= Duration::from_millis(MAX_STARTUP_STAGGER_MS));
+        assert!(d2 >= Duration::from_millis(STARTUP_STAGGER_STEP_MS * 2));
+        assert!(d2 <= Duration::from_millis(MAX_STARTUP_STAGGER_MS));
+    }
+
+    #[test]
+    fn reconnect_delay_is_immediate_on_first_failure() {
+        assert_eq!(compute_reconnect_delay(700, 45_000, 1, 123), Duration::ZERO);
+    }
+
+    #[test]
+    fn reconnect_delay_stays_within_probe_ceiling_after_many_failures() {
+        let d = compute_reconnect_delay(500, 45_000, 100, 12345);
+        assert!(d <= Duration::from_millis(RECONNECT_PROBE_MAX_DELAY_MS));
+    }
+}
@@ -0,0 +1,260 @@
+//! Binary frame protocol for WebSocket tunnel multiplexing.
+//!
+//! Frame layout (10-byte header + variable payload):
+//! ```text
+//! | stream_id (4B) | msg_type (1B) | flags (1B) | payload_len (4B) | payload (NB) |
+//! ```
+
+use bytes::{Buf, BufMut, Bytes, BytesMut};
+
+pub const HEADER_SIZE: usize = 10;
+
+/// Frame flags.
+pub mod flags {
+    pub const END_STREAM: u8 = 0x01;
+    pub const GZIP_COMPRESSED: u8 = 0x02;
+}
+
+/// Message types for the tunnel protocol.
+#[derive(Debug, Clone, Copy, PartialEq, Eq)]
+#[repr(u8)]
+pub enum MsgType {
+    RequestHeaders = 0x01,
+    RequestBody = 0x02,
+    ResponseHeaders = 0x03,
+    ResponseBody = 0x04,
+    StreamEnd = 0x05,
+    StreamError = 0x06,
+    Ping = 0x10,
+    Pong = 0x11,
+    GoAway = 0x12,
+    HeartbeatData = 0x13,
+    HeartbeatAck = 0x14,
+}
+
+impl MsgType {
+    pub fn from_u8(v: u8) -> Option<Self> {
+        match v {
+            0x01 => Some(Self::RequestHeaders),
+            0x02 => Some(Self::RequestBody),
+            0x03 => Some(Self::ResponseHeaders),
+            0x04 => Some(Self::ResponseBody),
+            0x05 => Some(Self::StreamEnd),
+            0x06 => Some(Self::StreamError),
+            0x10 => Some(Self::Ping),
+            0x11 => Some(Self::Pong),
+            0x12 => Some(Self::GoAway),
+            0x13 => Some(Self::HeartbeatData),
+            0x14 => Some(Self::HeartbeatAck),
+            _ => None,
+        }
+    }
+}
+
+/// A single multiplexed frame.
+#[derive(Debug, Clone)]
+pub struct Frame {
+    pub stream_id: u32,
+    pub msg_type: MsgType,
+    pub flags: u8,
+    pub payload: Bytes,
+}
+
+impl Frame {
+    pub fn new(stream_id: u32, msg_type: MsgType, flags: u8, payload: impl Into<Bytes>) -> Self {
+        Self {
+            stream_id,
+            msg_type,
+            flags,
+            payload: payload.into(),
+        }
+    }
+
+    /// Control frame (stream_id = 0).
+    pub fn control(msg_type: MsgType, payload: impl Into<Bytes>) -> Self {
+        Self::new(0, msg_type, 0, payload)
+    }
+
+    pub fn is_end_stream(&self) -> bool {
+        self.flags & flags::END_STREAM != 0
+    }
+
+    pub fn is_gzip(&self) -> bool {
+        self.flags & flags::GZIP_COMPRESSED != 0
+    }
+
+    /// Encode into a binary buffer.
+    pub fn encode(&self) -> Bytes {
+        let mut buf = BytesMut::with_capacity(HEADER_SIZE + self.payload.len());
+        buf.put_u32(self.stream_id);
+        buf.put_u8(self.msg_type as u8);
+        buf.put_u8(self.flags);
+        buf.put_u32(self.payload.len() as u32);
+        buf.put(self.payload.clone());
+        buf.freeze()
+    }
+
+    /// Decode from a binary buffer.
+    pub fn decode(mut data: Bytes) -> Result<Self, ProtocolError> {
+        if data.len() < HEADER_SIZE {
+            return Err(ProtocolError::TooShort {
+                expected: HEADER_SIZE,
+                actual: data.len(),
+            });
+        }
+        let stream_id = data.get_u32();
+        let msg_type_raw = data.get_u8();
+        let frame_flags = data.get_u8();
+        let payload_len = data.get_u32() as usize;
+
+        if data.remaining() < payload_len {
+            return Err(ProtocolError::Incomplete {
+                expected: HEADER_SIZE + payload_len,
+                actual: HEADER_SIZE + data.remaining(),
+            });
+        }
+
+        let msg_type =
+            MsgType::from_u8(msg_type_raw).ok_or(ProtocolError::UnknownMsgType(msg_type_raw))?;
+        let payload = data.split_to(payload_len);
+
+        Ok(Self {
+            stream_id,
+            msg_type,
+            flags: frame_flags,
+            payload,
+        })
+    }
+}
+
+/// Protocol errors.
+#[derive(Debug, thiserror::Error)]
+pub enum ProtocolError {
+    #[error("frame too short: expected {expected} bytes, got {actual}")]
+    TooShort { expected: usize, actual: usize },
+    #[error("frame incomplete: expected {expected} bytes, got {actual}")]
+    Incomplete { expected: usize, actual: usize },
+    #[error("unknown message type: 0x{0:02x}")]
+    UnknownMsgType(u8),
+}
+
+/// JSON payload for REQUEST_HEADERS frames.
+#[derive(Debug, 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)
+        }
+    }
+}
+
+/// JSON payload for RESPONSE_HEADERS frames.
+#[derive(Debug, serde::Serialize)]
+pub struct ResponseMeta {
+    pub status: u16,
+    /// Header list preserving duplicates (e.g. multiple Set-Cookie).
+    pub headers: Vec<(String, String)>,
+}
+
+// ---------------------------------------------------------------------------
+// Tunnel frame compression helpers
+// ---------------------------------------------------------------------------
+
+/// Minimum payload size to attempt gzip compression (bytes).
+const COMPRESS_MIN_SIZE: usize = 512;
+
+/// If the frame has the GZIP_COMPRESSED flag, decompress the payload; otherwise
+/// return a clone of the raw payload bytes.
+pub fn decompress_if_gzip(frame: &Frame) -> Result<Bytes, std::io::Error> {
+    if frame.is_gzip() {
+        decompress_gzip(&frame.payload)
+    } else {
+        Ok(frame.payload.clone())
+    }
+}
+
+/// Gzip-compress `data` if it is large enough and compression actually shrinks
+/// the payload. Returns `(payload, extra_flags)` where `extra_flags` contains
+/// `GZIP_COMPRESSED` when compression was applied.
+pub fn compress_payload(data: Bytes) -> (Bytes, u8) {
+    if data.len() >= COMPRESS_MIN_SIZE {
+        if let Ok(compressed) = compress_gzip(&data) {
+            if compressed.len() < data.len() {
+                return (compressed, flags::GZIP_COMPRESSED);
+            }
+        }
+    }
+    (data, 0)
+}
+
+fn decompress_gzip(data: &[u8]) -> Result<Bytes, std::io::Error> {
+    use flate2::read::GzDecoder;
+    use std::io::Read;
+    let mut decoder = GzDecoder::new(data);
+    let mut buf = Vec::new();
+    decoder.read_to_end(&mut buf)?;
+    Ok(Bytes::from(buf))
+}
+
+fn compress_gzip(data: &[u8]) -> Result<Bytes, std::io::Error> {
+    use flate2::write::GzEncoder;
+    use flate2::Compression;
+    use std::io::Write;
+    let mut encoder = GzEncoder::new(Vec::new(), Compression::fast());
+    encoder.write_all(data)?;
+    let compressed = encoder.finish()?;
+    Ok(Bytes::from(compressed))
+}
+
+#[cfg(test)]
+mod tests {
+    use super::RequestMeta;
+
+    #[test]
+    fn request_meta_accepts_integer_timeout() {
+        let raw = br#"{"method":"GET","url":"https://example.com","headers":{},"timeout":15}"#;
+        let meta: RequestMeta = serde_json::from_slice(raw).expect("parse request meta");
+        assert_eq!(meta.timeout, 15);
+    }
+
+    #[test]
+    fn request_meta_accepts_integer_like_float_timeout() {
+        let raw = br#"{"method":"GET","url":"https://example.com","headers":{},"timeout":15.0}"#;
+        let meta: RequestMeta = serde_json::from_slice(raw).expect("parse request meta");
+        assert_eq!(meta.timeout, 15);
+    }
+}
@@ -0,0 +1,506 @@
+//! Per-stream request handler.
+//!
+//! 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;
+
+use super::protocol::{
+    compress_payload, decompress_if_gzip, flags, Frame as TunnelFrame, MsgType, RequestMeta,
+    ResponseMeta,
+};
+use super::writer::FrameSender;
+
+/// Maximum response body chunk size per frame (32 KB).
+const MAX_CHUNK_SIZE: usize = 32 * 1024;
+
+/// Timeout for sending a single frame to the writer channel.
+/// If the writer is congested (TCP backpressure), we abandon the stream
+/// rather than blocking indefinitely and exhausting the stream pool.
+const FRAME_SEND_TIMEOUT: Duration = Duration::from_secs(30);
+
+/// Minimum allowed upstream request timeout (seconds).
+const MIN_TIMEOUT_SECS: u64 = 5;
+/// Maximum allowed upstream request timeout (seconds).
+const MAX_TIMEOUT_SECS: u64 = 300;
+
+/// Headers that must not be forwarded to upstream (hop-by-hop or security-sensitive).
+///
+/// `host` and `content-length` are managed by the HTTP client (reqwest/hyper):
+/// - `host` → translated to `:authority` pseudo-header in HTTP/2; forwarding
+///   the original `host` alongside `:authority` triggers PROTOCOL_ERROR on
+///   strict H2 implementations (e.g. Google APIs).
+/// - `content-length` → recalculated by hyper from the actual body; a stale
+///   value from the tunnel (body may have been re-compressed) causes H2
+///   PROTOCOL_ERROR when it mismatches the real frame length.
+const BLOCKED_HEADERS: &[&str] = &[
+    "connection",
+    "content-length",
+    "host",
+    "keep-alive",
+    "proxy-authenticate",
+    "proxy-authorization",
+    "proxy-connection",
+    "te",
+    "trailer",
+    "transfer-encoding",
+    "upgrade",
+];
+
+/// Handle a single stream: receive body, execute upstream, send response.
+pub async fn handle_stream(
+    state: Arc<AppState>,
+    server: Arc<ServerContext>,
+    stream_id: u32,
+    meta: RequestMeta,
+    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, body_rx, &frame_tx).await;
+
+    server.active_connections.fetch_sub(1, Ordering::Release);
+    if let Some(d) = connect_elapsed {
+        server.metrics.record_request(d);
+    }
+}
+
+/// Send a frame to the writer with a timeout. Returns false if send failed.
+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(_)) => {
+            // Channel closed (writer exited)
+            false
+        }
+        Err(_) => {
+            // Timeout — writer is congested
+            warn!("frame send timeout (writer congested), abandoning stream");
+            false
+        }
+    }
+}
+
+/// Returns the connection-establishment duration (DNS + TCP/TLS + TTFB) if the
+/// upstream request succeeded, or `None` if the request never reached the
+/// response-headers stage.
+async fn handle_stream_inner(
+    state: &AppState,
+    server: &ServerContext,
+    stream_id: u32,
+    meta: RequestMeta,
+    body_rx: mpsc::Receiver<TunnelFrame>,
+    frame_tx: &FrameSender,
+) -> Option<Duration> {
+    // Validate target
+    let target_url = match url::Url::parse(&meta.url) {
+        Ok(u) => u,
+        Err(e) => {
+            send_error(frame_tx, stream_id, &format!("invalid URL: {e}")).await;
+            return None;
+        }
+    };
+
+    // Only allow http/https schemes (block file://, data://, etc.)
+    match target_url.scheme() {
+        "http" | "https" => {}
+        other => {
+            send_error(
+                frame_tx,
+                stream_id,
+                &format!("unsupported URL scheme: {other}"),
+            )
+            .await;
+            return None;
+        }
+    }
+
+    let host = match target_url.host_str() {
+        Some(h) => h.to_string(),
+        None => {
+            send_error(frame_tx, stream_id, "missing host in URL").await;
+            return None;
+        }
+    };
+    let port = target_url.port_or_known_default().unwrap_or(443);
+
+    // DNS + target validation (populates dns_cache for SafeDnsResolver)
+    let connect_start = Instant::now();
+    {
+        let allowed_ports = Arc::clone(&server.dynamic.load().allowed_ports);
+        if let Err(e) =
+            target_filter::validate_target(&host, port, &allowed_ports, &state.dns_cache).await
+        {
+            server.metrics.dns_failures.fetch_add(1, Ordering::Release);
+            send_error(frame_tx, stream_id, &format!("target blocked: {e}")).await;
+            return None;
+        }
+    }
+    let dns_ms = connect_start.elapsed().as_millis() as u64;
+
+    // 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(request_body)
+    {
+        Ok(request) => request,
+        Err(e) => {
+            send_error(
+                frame_tx,
+                stream_id,
+                &format!("invalid upstream request: {e}"),
+            )
+            .await;
+            return None;
+        }
+    };
+
+    let headers = request.headers_mut();
+    for (k, v) in &meta.headers {
+        let k_lower = k.to_ascii_lowercase();
+        if BLOCKED_HEADERS.contains(&k_lower.as_str()) {
+            continue;
+        }
+        if let (Ok(name), Ok(value)) = (
+            hyper::header::HeaderName::from_bytes(k.as_bytes()),
+            hyper::header::HeaderValue::from_str(v),
+        ) {
+            headers.insert(name, value);
+        }
+    }
+
+    let mut captured_connection = upstream_client::capture_connection(&mut request);
+    let connection_start = Instant::now();
+    let connection_capture = tokio::spawn(async move {
+        let connected = captured_connection.wait_for_connection_metadata().await;
+        connected
+            .as_ref()
+            .map(|_| connection_start.elapsed().as_millis() as u64)
+    });
+
+    let upstream_start = Instant::now();
+    let response = match tokio::time::timeout(timeout, client.request(request)).await {
+        Ok(Ok(response)) => response,
+        Ok(Err(e)) => {
+            connection_capture.abort();
+            server
+                .metrics
+                .failed_requests
+                .fetch_add(1, Ordering::Release);
+            let msg = if e.is_connect() {
+                format!("upstream connect error: {e}")
+            } else {
+                format!("upstream error: {e}")
+            };
+            send_error(frame_tx, stream_id, &msg).await;
+            return None;
+        }
+        Err(_) => {
+            connection_capture.abort();
+            server
+                .metrics
+                .failed_requests
+                .fetch_add(1, Ordering::Release);
+            send_error(frame_tx, stream_id, "upstream timeout").await;
+            return None;
+        }
+    };
+
+    // Capture connection-establishment duration (DNS + TCP/TLS + TTFB)
+    // before proceeding to stream the response body.
+    let connect_elapsed = connect_start.elapsed();
+
+    // Send RESPONSE_HEADERS
+    let status = response.status().as_u16();
+    let ttfb_ms = upstream_start.elapsed().as_millis() as u64;
+    // Short timeout: on connection reuse hyper may never fire the connect
+    // callback, so avoid blocking indefinitely.
+    let connection_acquire_ms =
+        match tokio::time::timeout(Duration::from_millis(100), connection_capture).await {
+            Ok(Ok(ms)) => ms,
+            Ok(Err(_)) => None, // JoinError (task panicked / cancelled)
+            Err(_) => None,     // timeout -- task is detached but lightweight
+        };
+    let request_timing =
+        upstream_client::resolve_request_timing(&response, connection_acquire_ms, ttfb_ms);
+    let mut resp_headers: Vec<(String, String)> = Vec::with_capacity(response.headers().len() + 1);
+    for (k, v) in response.headers() {
+        if let Ok(vs) = v.to_str() {
+            resp_headers.push((k.as_str().to_string(), vs.to_string()));
+        }
+    }
+    let timing = serde_json::json!({
+        "dns_ms": dns_ms,
+        "connection_acquire_ms": request_timing.connection_acquire_ms,
+        "connection_reused": request_timing.connection_reused,
+        "connect_ms": request_timing.connect_ms,
+        "tls_ms": request_timing.tls_ms,
+        "ttfb_ms": ttfb_ms,
+        "upstream_ms": ttfb_ms,
+        "response_wait_ms": request_timing.response_wait_ms,
+        "upstream_processing_ms": request_timing.response_wait_ms,
+        "timing_source": "instrumented_connector",
+        "total_ms": connect_elapsed.as_millis() as u64,
+        "body_size": request_body_size.load(Ordering::Relaxed),
+        "mode": "tunnel",
+    });
+    resp_headers.push(("x-proxy-timing".to_string(), timing.to_string()));
+    let resp_meta = ResponseMeta {
+        status,
+        headers: resp_headers,
+    };
+    let meta_json: Bytes = serde_json::to_vec(&resp_meta).unwrap_or_default().into();
+    let (meta_payload, meta_flags) = compress_payload(meta_json);
+    if !send_frame(
+        frame_tx,
+        TunnelFrame::new(
+            stream_id,
+            MsgType::ResponseHeaders,
+            meta_flags,
+            meta_payload,
+        ),
+    )
+    .await
+    {
+        return Some(connect_elapsed);
+    }
+
+    // Stream response body — relay upstream bytes through the tunnel.
+    // Apply tunnel-level frame compression for chunks that benefit from it
+    // (e.g. uncompressed SSE text). Already-compressed data (gzip/br from
+    // upstream Content-Encoding) won't shrink further and will be sent as-is
+    // thanks to the size check in compress_payload().
+    let mut stream = response.into_body().into_data_stream();
+    while let Some(chunk_result) = stream.next().await {
+        match chunk_result {
+            Ok(chunk) => {
+                if chunk.len() <= MAX_CHUNK_SIZE {
+                    let (payload, extra_flags) = compress_payload(chunk);
+                    if !send_frame(
+                        frame_tx,
+                        TunnelFrame::new(stream_id, MsgType::ResponseBody, extra_flags, payload),
+                    )
+                    .await
+                    {
+                        return Some(connect_elapsed);
+                    }
+                } else {
+                    // Split oversized chunks, compress each slice
+                    let mut offset = 0;
+                    while offset < chunk.len() {
+                        let end = (offset + MAX_CHUNK_SIZE).min(chunk.len());
+                        let slice = chunk.slice(offset..end);
+                        let (payload, extra_flags) = compress_payload(slice);
+                        if !send_frame(
+                            frame_tx,
+                            TunnelFrame::new(
+                                stream_id,
+                                MsgType::ResponseBody,
+                                extra_flags,
+                                payload,
+                            ),
+                        )
+                        .await
+                        {
+                            return Some(connect_elapsed);
+                        }
+                        offset = end;
+                    }
+                }
+            }
+            Err(e) => {
+                server.metrics.stream_errors.fetch_add(1, Ordering::Release);
+                warn!(stream_id, error = %e, "upstream body read error");
+                send_error(frame_tx, stream_id, &format!("body read error: {e}")).await;
+                return Some(connect_elapsed);
+            }
+        }
+    }
+
+    // Send STREAM_END
+    let _ = send_frame(
+        frame_tx,
+        TunnelFrame::new(
+            stream_id,
+            MsgType::StreamEnd,
+            flags::END_STREAM,
+            Bytes::new(),
+        ),
+    )
+    .await;
+
+    debug!(stream_id, status, "stream completed");
+    Some(connect_elapsed)
+}
+
+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,
+        TunnelFrame::new(
+            stream_id,
+            MsgType::StreamError,
+            0,
+            Bytes::from(msg.to_string()),
+        ),
+    )
+    .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);
+    }
+}
@@ -0,0 +1,63 @@
+//! Dedicated WebSocket writer task.
+//!
+//! All frame writes go through an mpsc channel to a single writer task,
+//! avoiding contention on the WebSocket sink.  The writer also sends
+//! periodic WebSocket Ping frames to keep the connection alive through
+//! intermediary proxies (Nginx, Cloudflare, etc.).
+
+use std::time::Duration;
+
+use futures_util::SinkExt;
+use tokio::sync::mpsc;
+use tokio::task::JoinHandle;
+use tokio_tungstenite::tungstenite::Message;
+use tracing::{debug, error, trace};
+
+use super::protocol::Frame;
+
+/// Sender half — cloned by stream handlers and heartbeat.
+pub type FrameSender = mpsc::Sender<Frame>;
+
+/// Spawn the writer task. Returns the sender and a JoinHandle for cleanup.
+///
+/// `ping_interval` controls WebSocket-level Ping frequency (typically 15s).
+/// This keeps the connection alive through intermediary proxies/load-balancers.
+pub fn spawn_writer<S>(mut sink: S, ping_interval: Duration) -> (FrameSender, JoinHandle<()>)
+where
+    S: SinkExt<Message, Error = tokio_tungstenite::tungstenite::Error> + Unpin + Send + 'static,
+{
+    let (tx, mut rx) = mpsc::channel::<Frame>(256);
+
+    let handle = tokio::spawn(async move {
+        let mut ping_ticker = tokio::time::interval(ping_interval);
+        ping_ticker.tick().await; // skip first immediate tick
+
+        loop {
+            tokio::select! {
+                frame = rx.recv() => {
+                    match frame {
+                        Some(frame) => {
+                            let data = frame.encode();
+                            if let Err(e) = sink.send(Message::Binary(data.into())).await {
+                                error!(error = %e, "failed to write frame to WebSocket");
+                                break;
+                            }
+                        }
+                        None => break, // all senders dropped
+                    }
+                }
+                _ = ping_ticker.tick() => {
+                    if let Err(e) = sink.send(Message::Ping(vec![])).await {
+                        error!(error = %e, "failed to send WebSocket ping");
+                        break;
+                    }
+                    trace!("sent WebSocket ping");
+                }
+            }
+        }
+        debug!("writer task exiting");
+        let _ = sink.close().await;
+    });
+
+    (tx, handle)
+}
@@ -0,0 +1,444 @@
+use std::future::Future;
+use std::io;
+use std::net::IpAddr;
+use std::pin::Pin;
+use std::sync::Arc;
+use std::task::{Context, Poll};
+use std::time::Duration;
+
+use bytes::Bytes;
+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;
+pub use hyper_util::client::legacy::connect::capture_connection;
+use hyper_util::client::legacy::connect::dns::Name;
+use hyper_util::client::legacy::connect::{Connected, Connection, HttpConnector};
+use hyper_util::client::legacy::Client;
+use hyper_util::rt::{TokioExecutor, TokioIo, TokioTimer};
+use rustls::pki_types::ServerName;
+use rustls::ClientConfig;
+use tokio::net::TcpStream;
+use tokio_rustls::TlsConnector;
+use tower_service::Service;
+
+use crate::config::Config;
+use crate::target_filter::{self, DnsCache};
+
+type BoxError = Box<dyn std::error::Error + Send + Sync>;
+
+type PlainStream = TokioIo<TcpStream>;
+type TlsStream = TokioIo<tokio_rustls::client::TlsStream<TcpStream>>;
+
+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,
+    pub tls_ms: u64,
+}
+
+#[derive(Clone, Copy, Debug, Default)]
+pub struct RequestTiming {
+    pub connection_acquire_ms: u64,
+    pub connect_ms: u64,
+    pub tls_ms: u64,
+    pub response_wait_ms: u64,
+    pub connection_reused: bool,
+}
+
+#[derive(Clone)]
+pub struct ValidatedResolver {
+    dns_cache: Arc<DnsCache>,
+}
+
+impl ValidatedResolver {
+    pub fn new(dns_cache: Arc<DnsCache>) -> Self {
+        Self { dns_cache }
+    }
+}
+
+pub struct ValidatedAddrs {
+    inner: std::vec::IntoIter<std::net::SocketAddr>,
+}
+
+impl Iterator for ValidatedAddrs {
+    type Item = std::net::SocketAddr;
+
+    fn next(&mut self) -> Option<Self::Item> {
+        self.inner.next()
+    }
+}
+
+impl Service<Name> for ValidatedResolver {
+    type Response = ValidatedAddrs;
+    type Error = io::Error;
+    type Future = Pin<Box<dyn Future<Output = Result<Self::Response, Self::Error>> + Send>>;
+
+    fn poll_ready(&mut self, _cx: &mut Context<'_>) -> Poll<Result<(), Self::Error>> {
+        Poll::Ready(Ok(()))
+    }
+
+    fn call(&mut self, name: Name) -> Self::Future {
+        let dns_cache = Arc::clone(&self.dns_cache);
+        let host = name.as_str().to_string();
+        Box::pin(async move {
+            if let Some(addrs) = dns_cache.get_by_host(&host).await {
+                return Ok(ValidatedAddrs {
+                    inner: (*addrs).clone().into_iter(),
+                });
+            }
+
+            let resolved = target_filter::resolve_public_addrs(&host, 0, dns_cache.as_ref())
+                .await
+                .map_err(|err| io::Error::other(err.to_string()))?;
+            Ok(ValidatedAddrs {
+                inner: resolved.into_iter(),
+            })
+        })
+    }
+}
+
+#[derive(Clone)]
+pub struct InstrumentedConnector {
+    http: HttpConnector<ValidatedResolver>,
+    tls_config: Arc<ClientConfig>,
+}
+
+impl Service<Uri> for InstrumentedConnector {
+    type Response = TimedConn;
+    type Error = BoxError;
+    type Future = Pin<Box<dyn Future<Output = Result<Self::Response, Self::Error>> + Send>>;
+
+    fn poll_ready(&mut self, cx: &mut Context<'_>) -> Poll<Result<(), Self::Error>> {
+        self.http.poll_ready(cx).map_err(Into::into)
+    }
+
+    fn call(&mut self, dst: Uri) -> Self::Future {
+        let scheme = dst.scheme_str().map(|value| value.to_ascii_lowercase());
+        let tls_config = Arc::clone(&self.tls_config);
+        let connecting = self.http.call(dst.clone());
+        let connect_start = std::time::Instant::now();
+
+        Box::pin(async move {
+            match scheme.as_deref() {
+                Some("http") => {
+                    let tcp = connecting.await.map_err(|err| Box::new(err) as BoxError)?;
+                    let connect_ms = connect_start.elapsed().as_millis() as u64;
+                    Ok(TimedConn::new(
+                        MaybeHttpsStream::Http(tcp),
+                        ConnectTiming {
+                            connect_ms,
+                            tls_ms: 0,
+                        },
+                    ))
+                }
+                Some("https") => {
+                    let server_name = resolve_server_name(&dst)?;
+                    let tcp = connecting.await.map_err(|err| Box::new(err) as BoxError)?;
+                    let connect_ms = connect_start.elapsed().as_millis() as u64;
+
+                    let tls_start = std::time::Instant::now();
+                    let tls_stream = TlsConnector::from(tls_config)
+                        .connect(server_name, tcp.into_inner())
+                        .await
+                        .map_err(io::Error::other)?;
+                    let tls_ms = tls_start.elapsed().as_millis() as u64;
+
+                    Ok(TimedConn::new(
+                        MaybeHttpsStream::Https(TokioIo::new(tls_stream)),
+                        ConnectTiming { connect_ms, tls_ms },
+                    ))
+                }
+                Some(other) => Err(io::Error::other(format!("unsupported scheme {other}")).into()),
+                None => Err(io::Error::other("missing scheme").into()),
+            }
+        })
+    }
+}
+
+pub fn build_upstream_client(config: &Config, dns_cache: Arc<DnsCache>) -> UpstreamClient {
+    let mut http = HttpConnector::new_with_resolver(ValidatedResolver::new(dns_cache));
+    http.enforce_http(false);
+    http.set_connect_timeout(Some(Duration::from_secs(
+        config.upstream_connect_timeout_secs,
+    )));
+    http.set_nodelay(config.upstream_tcp_nodelay);
+    if config.upstream_tcp_keepalive_secs > 0 {
+        http.set_keepalive(Some(Duration::from_secs(
+            config.upstream_tcp_keepalive_secs,
+        )));
+    } else {
+        http.set_keepalive(None);
+    }
+
+    let connector = InstrumentedConnector {
+        http,
+        tls_config: build_tls_config(),
+    };
+
+    let mut builder = Client::builder(TokioExecutor::new());
+    builder.pool_max_idle_per_host(config.upstream_pool_max_idle_per_host);
+    builder.pool_idle_timeout(Duration::from_secs(config.upstream_pool_idle_timeout_secs));
+    builder.pool_timer(TokioTimer::new());
+    builder.build(connector)
+}
+
+pub fn resolve_request_timing<B>(
+    response: &Response<B>,
+    connection_acquire_ms: Option<u64>,
+    ttfb_ms: u64,
+) -> RequestTiming {
+    let raw = response
+        .extensions()
+        .get::<ConnectTiming>()
+        .copied()
+        .unwrap_or_default();
+
+    let raw_connection_ms = raw.connect_ms.saturating_add(raw.tls_ms);
+    let measured_acquire_ms = connection_acquire_ms.unwrap_or(raw_connection_ms.min(ttfb_ms));
+    let likely_reused = measured_acquire_ms <= 5 && raw_connection_ms > 0;
+    let connector_matches_request = raw_connection_ms <= measured_acquire_ms.saturating_add(25);
+
+    let (connect_ms, tls_ms) = if likely_reused || !connector_matches_request {
+        (0, 0)
+    } else {
+        (raw.connect_ms, raw.tls_ms)
+    };
+
+    RequestTiming {
+        connection_acquire_ms: measured_acquire_ms,
+        connect_ms,
+        tls_ms,
+        response_wait_ms: ttfb_ms.saturating_sub(measured_acquire_ms),
+        connection_reused: likely_reused,
+    }
+}
+
+fn build_tls_config() -> Arc<ClientConfig> {
+    let root_store =
+        rustls::RootCertStore::from_iter(webpki_roots::TLS_SERVER_ROOTS.iter().cloned());
+    let mut config = ClientConfig::builder()
+        .with_root_certificates(root_store)
+        .with_no_client_auth();
+    config.alpn_protocols = vec![b"h2".to_vec(), b"http/1.1".to_vec()];
+    Arc::new(config)
+}
+
+fn resolve_server_name(uri: &Uri) -> Result<ServerName<'static>, BoxError> {
+    let host = uri.host().ok_or_else(|| io::Error::other("missing host"))?;
+    let host = host.trim_start_matches('[').trim_end_matches(']');
+
+    if let Ok(ip) = host.parse::<IpAddr>() {
+        return Ok(ServerName::from(ip));
+    }
+
+    Ok(ServerName::try_from(host.to_string())?)
+}
+
+pub struct TimedConn {
+    inner: MaybeHttpsStream,
+    timing: ConnectTiming,
+}
+
+impl TimedConn {
+    fn new(inner: MaybeHttpsStream, timing: ConnectTiming) -> Self {
+        Self { inner, timing }
+    }
+}
+
+impl Connection for TimedConn {
+    fn connected(&self) -> Connected {
+        self.inner.connected().extra(self.timing)
+    }
+}
+
+impl rt::Read for TimedConn {
+    fn poll_read(
+        mut self: Pin<&mut Self>,
+        cx: &mut Context<'_>,
+        buf: rt::ReadBufCursor<'_>,
+    ) -> Poll<Result<(), io::Error>> {
+        Pin::new(&mut self.inner).poll_read(cx, buf)
+    }
+}
+
+impl rt::Write for TimedConn {
+    fn poll_write(
+        mut self: Pin<&mut Self>,
+        cx: &mut Context<'_>,
+        buf: &[u8],
+    ) -> Poll<Result<usize, io::Error>> {
+        Pin::new(&mut self.inner).poll_write(cx, buf)
+    }
+
+    fn poll_flush(mut self: Pin<&mut Self>, cx: &mut Context<'_>) -> Poll<Result<(), io::Error>> {
+        Pin::new(&mut self.inner).poll_flush(cx)
+    }
+
+    fn poll_shutdown(
+        mut self: Pin<&mut Self>,
+        cx: &mut Context<'_>,
+    ) -> Poll<Result<(), io::Error>> {
+        Pin::new(&mut self.inner).poll_shutdown(cx)
+    }
+
+    fn is_write_vectored(&self) -> bool {
+        self.inner.is_write_vectored()
+    }
+
+    fn poll_write_vectored(
+        mut self: Pin<&mut Self>,
+        cx: &mut Context<'_>,
+        bufs: &[std::io::IoSlice<'_>],
+    ) -> Poll<Result<usize, io::Error>> {
+        Pin::new(&mut self.inner).poll_write_vectored(cx, bufs)
+    }
+}
+
+pub enum MaybeHttpsStream {
+    Http(PlainStream),
+    Https(TlsStream),
+}
+
+impl Connection for MaybeHttpsStream {
+    fn connected(&self) -> Connected {
+        match self {
+            Self::Http(stream) => stream.connected(),
+            Self::Https(stream) => {
+                let (tcp, tls) = stream.inner().get_ref();
+                if tls.alpn_protocol() == Some(b"h2") {
+                    tcp.connected().negotiated_h2()
+                } else {
+                    tcp.connected()
+                }
+            }
+        }
+    }
+}
+
+impl rt::Read for MaybeHttpsStream {
+    fn poll_read(
+        self: Pin<&mut Self>,
+        cx: &mut Context<'_>,
+        buf: rt::ReadBufCursor<'_>,
+    ) -> Poll<Result<(), io::Error>> {
+        match Pin::get_mut(self) {
+            Self::Http(stream) => Pin::new(stream).poll_read(cx, buf),
+            Self::Https(stream) => Pin::new(stream).poll_read(cx, buf),
+        }
+    }
+}
+
+impl rt::Write for MaybeHttpsStream {
+    fn poll_write(
+        self: Pin<&mut Self>,
+        cx: &mut Context<'_>,
+        buf: &[u8],
+    ) -> Poll<Result<usize, io::Error>> {
+        match Pin::get_mut(self) {
+            Self::Http(stream) => Pin::new(stream).poll_write(cx, buf),
+            Self::Https(stream) => Pin::new(stream).poll_write(cx, buf),
+        }
+    }
+
+    fn poll_flush(self: Pin<&mut Self>, cx: &mut Context<'_>) -> Poll<Result<(), io::Error>> {
+        match Pin::get_mut(self) {
+            Self::Http(stream) => Pin::new(stream).poll_flush(cx),
+            Self::Https(stream) => Pin::new(stream).poll_flush(cx),
+        }
+    }
+
+    fn poll_shutdown(self: Pin<&mut Self>, cx: &mut Context<'_>) -> Poll<Result<(), io::Error>> {
+        match Pin::get_mut(self) {
+            Self::Http(stream) => Pin::new(stream).poll_shutdown(cx),
+            Self::Https(stream) => Pin::new(stream).poll_shutdown(cx),
+        }
+    }
+
+    fn is_write_vectored(&self) -> bool {
+        match self {
+            Self::Http(stream) => stream.is_write_vectored(),
+            Self::Https(stream) => stream.is_write_vectored(),
+        }
+    }
+
+    fn poll_write_vectored(
+        self: Pin<&mut Self>,
+        cx: &mut Context<'_>,
+        bufs: &[std::io::IoSlice<'_>],
+    ) -> Poll<Result<usize, io::Error>> {
+        match Pin::get_mut(self) {
+            Self::Http(stream) => Pin::new(stream).poll_write_vectored(cx, bufs),
+            Self::Https(stream) => Pin::new(stream).poll_write_vectored(cx, bufs),
+        }
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+    use hyper::Response;
+
+    #[test]
+    fn fresh_connection_uses_connector_breakdown() {
+        let mut response = Response::new(());
+        response.extensions_mut().insert(ConnectTiming {
+            connect_ms: 80,
+            tls_ms: 40,
+        });
+
+        let timing = resolve_request_timing(&response, Some(125), 600);
+
+        assert_eq!(timing.connection_acquire_ms, 125);
+        assert_eq!(timing.connect_ms, 80);
+        assert_eq!(timing.tls_ms, 40);
+        assert_eq!(timing.response_wait_ms, 475);
+        assert!(!timing.connection_reused);
+    }
+
+    #[test]
+    fn reused_connection_zeroes_stale_connect_timings() {
+        let mut response = Response::new(());
+        response.extensions_mut().insert(ConnectTiming {
+            connect_ms: 70,
+            tls_ms: 30,
+        });
+
+        let timing = resolve_request_timing(&response, Some(0), 310);
+
+        assert_eq!(timing.connection_acquire_ms, 0);
+        assert_eq!(timing.connect_ms, 0);
+        assert_eq!(timing.tls_ms, 0);
+        assert_eq!(timing.response_wait_ms, 310);
+        assert!(timing.connection_reused);
+    }
+
+    #[test]
+    fn falls_back_to_connector_timings_when_capture_missing() {
+        let mut response = Response::new(());
+        response.extensions_mut().insert(ConnectTiming {
+            connect_ms: 55,
+            tls_ms: 25,
+        });
+
+        let timing = resolve_request_timing(&response, None, 400);
+
+        assert_eq!(timing.connection_acquire_ms, 80);
+        assert_eq!(timing.connect_ms, 55);
+        assert_eq!(timing.tls_ms, 25);
+        assert_eq!(timing.response_wait_ms, 320);
+        assert!(!timing.connection_reused);
+    }
+}
@@ -3,13 +3,15 @@ Alembic 环境配置
 用于数据库迁移的运行时环境设置
 """

-from logging.config import fileConfig
-from sqlalchemy import engine_from_config, pool
-from alembic import context
 import os
 import sys
+from logging.config import fileConfig
 from pathlib import Path

+from sqlalchemy import engine_from_config, pool, text
+
+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__)))

@@ -48,6 +50,11 @@ if config.config_file_name is not None:
 # 目标元数据（包含所有表定义）
 target_metadata = Base.metadata

+# PostgreSQL 全局迁移锁，避免多进程并发执行 Alembic 导致竞态（重复加列/索引等）
+# 使用会话级 advisory lock（pg_advisory_lock），在迁移完成后手动释放。
+# ID 由 crc32("aether-alembic-migration") 拼接生成，仅需全局唯一即可。
+MIGRATION_ADVISORY_LOCK_ID = 582694137405821
+

 def run_migrations_offline() -> None:
    """
@@ -83,15 +90,28 @@ def run_migrations_online() -> None:
    )

    with connectable.connect() as connection:
-        context.configure(
-            connection=connection,
-            target_metadata=target_metadata,
-            compare_type=True,  # 比较列类型变更
-            compare_server_default=True,  # 比较默认值变更
-        )
+        try:
+            # 使用会话级 advisory lock（非事务级），避免干扰 Alembic 的事务管理。
+            # pg_advisory_lock 在会话结束时自动释放，不受 COMMIT/ROLLBACK 影响。
+            if connection.dialect.name == "postgresql":
+                connection.execute(
+                    text("SELECT pg_advisory_lock(:lock_id)"),
+                    {"lock_id": MIGRATION_ADVISORY_LOCK_ID},
+                )
+                connection.commit()

-        with context.begin_transaction():
-            context.run_migrations()
+            context.configure(
+                connection=connection,
+                target_metadata=target_metadata,
+                compare_type=True,
+                compare_server_default=True,
+                transaction_per_migration=True,  # 每个迁移文件独立事务，完成即提交
+            )
+
+            with context.begin_transaction():
+                context.run_migrations()
+        except Exception:
+            raise


 # 根据模式选择运行方式
@@ -226,26 +226,50 @@ def upgrade() -> None:
            sa.Column("proxy", postgresql.JSONB(), nullable=True, comment="代理配置"),
        )

-    # 从端点迁移数据到 provider
-    op.execute("""
-        UPDATE providers p
-        SET
-            timeout = COALESCE(
-                p.timeout,
-                (SELECT MAX(e.timeout) FROM provider_endpoints e WHERE e.provider_id = p.id AND e.timeout IS NOT NULL),
-                300
-            ),
-            max_retries = COALESCE(
-                p.max_retries,
-                (SELECT MAX(e.max_retries) FROM provider_endpoints e WHERE e.provider_id = p.id AND e.max_retries IS NOT NULL),
-                2
-            ),
+    # 从端点迁移数据到 provider（动态构建 SQL，仅引用存在的列）
+    ep_has_timeout = _column_exists("provider_endpoints", "timeout")
+    ep_has_max_retries = _column_exists("provider_endpoints", "max_retries")
+    ep_has_proxy = _column_exists("provider_endpoints", "proxy")
+
+    set_clauses = []
+    if _column_exists("providers", "timeout"):
+        if ep_has_timeout:
+            set_clauses.append("""
+                timeout = COALESCE(
+                    p.timeout,
+                    (SELECT MAX(e.timeout) FROM provider_endpoints e WHERE e.provider_id = p.id AND e.timeout IS NOT NULL),
+                    300
+                )""")
+        else:
+            set_clauses.append("timeout = COALESCE(p.timeout, 300)")
+
+    if _column_exists("providers", "max_retries"):
+        if ep_has_max_retries:
+            set_clauses.append("""
+                max_retries = COALESCE(
+                    p.max_retries,
+                    (SELECT MAX(e.max_retries) FROM provider_endpoints e WHERE e.provider_id = p.id AND e.max_retries IS NOT NULL),
+                    2
+                )""")
+        else:
+            set_clauses.append("max_retries = COALESCE(p.max_retries, 2)")
+
+    if _column_exists("providers", "proxy") and ep_has_proxy:
+        set_clauses.append("""
            proxy = COALESCE(
                p.proxy,
                (SELECT e.proxy FROM provider_endpoints e WHERE e.provider_id = p.id AND e.proxy IS NOT NULL ORDER BY e.created_at LIMIT 1)
-            )
-        WHERE p.timeout IS NULL OR p.max_retries IS NULL
-    """)
+            )""")
+
+    if set_clauses:
+        where_parts = []
+        if _column_exists("providers", "timeout"):
+            where_parts.append("p.timeout IS NULL")
+        if _column_exists("providers", "max_retries"):
+            where_parts.append("p.max_retries IS NULL")
+        where_clause = " OR ".join(where_parts) if where_parts else "TRUE"
+        sql = "UPDATE providers p SET " + ", ".join(set_clauses) + " WHERE " + where_clause
+        op.execute(sql)

    # ========== 5. providers: display_name -> name ==========
    # 注意：这里假设 display_name 已经被重命名为 name
@@ -0,0 +1,46 @@
+"""Add provider_request_body and client_response_body columns to usage table
+
+Revision ID: 7e8f9a0b1c2d
+Revises: 6d7e8f9a0b1c
+Create Date: 2026-02-20 18:00:00.000000
+
+"""
+
+from __future__ import annotations
+
+from collections.abc import Sequence
+
+from sqlalchemy import text
+
+from alembic import op
+
+# revision identifiers, used by Alembic.
+revision: str = "7e8f9a0b1c2d"
+down_revision: str | None = "6d7e8f9a0b1c"
+branch_labels: str | Sequence[str] | None = None
+depends_on: str | Sequence[str] | None = None
+
+
+def upgrade() -> None:
+    conn = op.get_bind()
+    # Use PostgreSQL native IF NOT EXISTS to avoid duplicate-column races
+    # when migrations are triggered concurrently (e.g. startup + manual run).
+    conn.execute(text("ALTER TABLE usage ADD COLUMN IF NOT EXISTS provider_request_body JSON"))
+    conn.execute(
+        text("ALTER TABLE usage ADD COLUMN IF NOT EXISTS provider_request_body_compressed BYTEA")
+    )
+    conn.execute(text("ALTER TABLE usage ADD COLUMN IF NOT EXISTS client_response_body JSON"))
+    conn.execute(
+        text("ALTER TABLE usage ADD COLUMN IF NOT EXISTS client_response_body_compressed BYTEA")
+    )
+
+
+def downgrade() -> None:
+    conn = op.get_bind()
+    for col in (
+        "client_response_body_compressed",
+        "client_response_body",
+        "provider_request_body_compressed",
+        "provider_request_body",
+    ):
+        conn.execute(text(f"ALTER TABLE usage DROP COLUMN IF EXISTS {col}"))
@@ -0,0 +1,91 @@
+"""Add api_family and endpoint_kind columns to usage table
+
+Revision ID: 8f9a0b1c2d3e
+Revises: 7e8f9a0b1c2d
+Create Date: 2026-02-21 15:00:00.000000
+
+"""
+
+from __future__ import annotations
+
+from collections.abc import Sequence
+
+import sqlalchemy as sa
+from sqlalchemy import inspect, text
+
+from alembic import op
+
+# revision identifiers, used by Alembic.
+revision: str = "8f9a0b1c2d3e"
+down_revision: str | None = "7e8f9a0b1c2d"
+branch_labels: str | Sequence[str] | None = None
+depends_on: str | Sequence[str] | None = None
+
+# Usage 表新增列
+NEW_COLUMNS = [
+    ("api_family", sa.String(50)),
+    ("endpoint_kind", sa.String(50)),
+    ("provider_api_family", sa.String(50)),
+    ("provider_endpoint_kind", sa.String(50)),
+]
+
+# 新增索引
+NEW_INDEXES = [
+    ("idx_usage_api_family", "usage", ["api_family"]),
+    ("idx_usage_endpoint_kind", "usage", ["endpoint_kind"]),
+    ("idx_usage_family_kind", "usage", ["api_family", "endpoint_kind"]),
+]
+
+
+def upgrade() -> None:
+    conn = op.get_bind()
+
+    # 使用 PostgreSQL 原生 IF NOT EXISTS，比 inspect 更可靠（避免同一事务内缓存问题）
+    col_definitions = {
+        "api_family": "VARCHAR(50)",
+        "endpoint_kind": "VARCHAR(50)",
+        "provider_api_family": "VARCHAR(50)",
+        "provider_endpoint_kind": "VARCHAR(50)",
+    }
+    for col_name, col_type_sql in col_definitions.items():
+        conn.execute(text(f"ALTER TABLE usage ADD COLUMN IF NOT EXISTS {col_name} {col_type_sql}"))
+
+    # 数据迁移：从 api_format 解析 api_family + endpoint_kind
+    conn.execute(text("""
+            UPDATE usage SET
+                api_family = lower(split_part(api_format, ':', 1)),
+                endpoint_kind = lower(split_part(api_format, ':', 2))
+            WHERE api_format IS NOT NULL
+              AND api_format LIKE '%%:%%'
+              AND api_family IS NULL
+        """))
+    conn.execute(text("""
+            UPDATE usage SET
+                provider_api_family = lower(split_part(endpoint_api_format, ':', 1)),
+                provider_endpoint_kind = lower(split_part(endpoint_api_format, ':', 2))
+            WHERE endpoint_api_format IS NOT NULL
+              AND endpoint_api_format LIKE '%%:%%'
+              AND provider_api_family IS NULL
+        """))
+
+    # 创建索引
+    inspector = inspect(conn)
+    existing_indexes = {idx["name"] for idx in inspector.get_indexes("usage")}
+    for idx_name, table, columns in NEW_INDEXES:
+        if idx_name not in existing_indexes:
+            op.create_index(idx_name, table, columns)
+
+
+def downgrade() -> None:
+    conn = op.get_bind()
+    inspector = inspect(conn)
+
+    existing_indexes = {idx["name"] for idx in inspector.get_indexes("usage")}
+    for idx_name, _, _ in reversed(NEW_INDEXES):
+        if idx_name in existing_indexes:
+            op.drop_index(idx_name, table_name="usage")
+
+    existing_columns = {col["name"] for col in inspector.get_columns("usage")}
+    for col_name, _ in reversed(NEW_COLUMNS):
+        if col_name in existing_columns:
+            op.drop_column("usage", col_name)
@@ -0,0 +1,106 @@
+"""Add tunnel mode fields and remove IP forwarding fields
+
+Revision ID: 9a0b1c2d3e4f
+Revises: 8f9a0b1c2d3e
+Create Date: 2026-02-24 17:00:00.000000
+
+"""
+
+from __future__ import annotations
+
+from collections.abc import Sequence
+
+import sqlalchemy as sa
+from sqlalchemy import inspect
+
+from alembic import op
+
+revision: str = "9a0b1c2d3e4f"
+down_revision: str | None = "8f9a0b1c2d3e"
+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:
+    # 添加 tunnel 模式字段
+    if not column_exists("proxy_nodes", "tunnel_mode"):
+        op.add_column(
+            "proxy_nodes",
+            sa.Column(
+                "tunnel_mode",
+                sa.Boolean(),
+                nullable=False,
+                server_default=sa.text("false"),
+                comment="是否使用 WebSocket 隧道模式",
+            ),
+        )
+    if not column_exists("proxy_nodes", "tunnel_connected"):
+        op.add_column(
+            "proxy_nodes",
+            sa.Column(
+                "tunnel_connected",
+                sa.Boolean(),
+                nullable=False,
+                server_default=sa.text("false"),
+                comment="隧道是否已连接",
+            ),
+        )
+    if not column_exists("proxy_nodes", "tunnel_connected_at"):
+        op.add_column(
+            "proxy_nodes",
+            sa.Column(
+                "tunnel_connected_at",
+                sa.DateTime(timezone=True),
+                nullable=True,
+                comment="隧道最近一次建立时间",
+            ),
+        )
+
+    # tunnel 模式节点不需要 port，将其置零
+    op.execute("UPDATE proxy_nodes SET port = 0 WHERE tunnel_mode = true")
+
+    # 移除旧的 IP 转发字段
+    if column_exists("proxy_nodes", "tls_enabled"):
+        op.drop_column("proxy_nodes", "tls_enabled")
+    if column_exists("proxy_nodes", "tls_cert_fingerprint"):
+        op.drop_column("proxy_nodes", "tls_cert_fingerprint")
+
+
+def downgrade() -> None:
+    # 恢复 IP 转发字段
+    if not column_exists("proxy_nodes", "tls_cert_fingerprint"):
+        op.add_column(
+            "proxy_nodes",
+            sa.Column(
+                "tls_cert_fingerprint",
+                sa.String(128),
+                nullable=True,
+                comment="TLS 证书 SHA-256 指纹（hex）",
+            ),
+        )
+    if not column_exists("proxy_nodes", "tls_enabled"):
+        op.add_column(
+            "proxy_nodes",
+            sa.Column(
+                "tls_enabled",
+                sa.Boolean(),
+                nullable=False,
+                server_default=sa.text("false"),
+                comment="是否启用 TLS 加密",
+            ),
+        )
+
+    # 移除 tunnel 模式字段
+    if column_exists("proxy_nodes", "tunnel_connected_at"):
+        op.drop_column("proxy_nodes", "tunnel_connected_at")
+    if column_exists("proxy_nodes", "tunnel_connected"):
+        op.drop_column("proxy_nodes", "tunnel_connected")
+    if column_exists("proxy_nodes", "tunnel_mode"):
+        op.drop_column("proxy_nodes", "tunnel_mode")
@@ -0,0 +1,224 @@
+"""Add cache_creation columns, clean up capability settings, add user_model_usage_counts,
+enforce global_model_id NOT NULL
+
+1. Add cache_creation_input_tokens_5m and cache_creation_input_tokens_1h to usage table.
+2. Clean up cache_1h/context_1m/gemini_files from user-configurable settings
+   (now auto-detected via REQUEST_PARAM mode).
+3. Create user_model_usage_counts table for per-user per-model atomic usage counters.
+4. Enforce models.global_model_id NOT NULL (delete orphan models without global model).
+
+Revision ID: b2c3d4e5f6a7
+Revises: 9a0b1c2d3e4f
+Create Date: 2026-02-28 14:00:00.000000
+
+"""
+
+from __future__ import annotations
+
+import json
+import uuid
+from collections.abc import Sequence
+from datetime import datetime, timezone
+
+import sqlalchemy as sa
+from sqlalchemy import inspect
+
+from alembic import op
+
+revision: str = "b2c3d4e5f6a7"
+down_revision: str | None = "9a0b1c2d3e4f"
+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()
+    insp = inspect(bind)
+    columns = [c["name"] for c in insp.get_columns(table_name)]
+    return column_name in columns
+
+
+def table_exists(table_name: str) -> bool:
+    bind = op.get_bind()
+    insp = inspect(bind)
+    return table_name in insp.get_table_names()
+
+
+def index_exists(table_name: str, index_name: str) -> bool:
+    bind = op.get_bind()
+    insp = inspect(bind)
+    return any(idx["name"] == index_name for idx in insp.get_indexes(table_name))
+
+
+def upgrade() -> None:
+    # --- 1. Add cache_creation columns ---
+    if not column_exists("usage", "cache_creation_input_tokens_5m"):
+        op.add_column(
+            "usage",
+            sa.Column(
+                "cache_creation_input_tokens_5m",
+                sa.Integer(),
+                nullable=False,
+                server_default=sa.text("0"),
+                comment="5min TTL cache creation input tokens",
+            ),
+        )
+    if not column_exists("usage", "cache_creation_input_tokens_1h"):
+        op.add_column(
+            "usage",
+            sa.Column(
+                "cache_creation_input_tokens_1h",
+                sa.Integer(),
+                nullable=False,
+                server_default=sa.text("0"),
+                comment="1h TTL cache creation input tokens",
+            ),
+        )
+
+    # --- 2. Clean up stale capability settings (pure Python, DB-agnostic) ---
+    stale_keys = {"cache_1h", "context_1m", "gemini_files"}
+    conn = op.get_bind()
+
+    # ApiKey.force_capabilities: dict-like JSON, remove stale keys
+    rows = conn.execute(
+        sa.text("SELECT id, force_capabilities FROM api_keys WHERE force_capabilities IS NOT NULL")
+    ).fetchall()
+    for row in rows:
+        raw = row[1]
+        if raw is None:
+            continue
+        data = raw if isinstance(raw, dict) else json.loads(raw)
+        cleaned = {k: v for k, v in data.items() if k not in stale_keys}
+        new_val = json.dumps(cleaned) if cleaned else None
+        conn.execute(
+            sa.text("UPDATE api_keys SET force_capabilities = :val WHERE id = :id"),
+            {"val": new_val, "id": row[0]},
+        )
+
+    # User.model_capability_settings: nested dict {model_key: {cap: val}}, remove stale keys
+    rows = conn.execute(
+        sa.text(
+            "SELECT id, model_capability_settings FROM users"
+            " WHERE model_capability_settings IS NOT NULL"
+        )
+    ).fetchall()
+    for row in rows:
+        raw = row[1]
+        if raw is None:
+            continue
+        data = raw if isinstance(raw, dict) else json.loads(raw)
+        cleaned = {}
+        for model_key, caps in data.items():
+            cap_cleaned = {k: v for k, v in caps.items() if k not in stale_keys}
+            if cap_cleaned:
+                cleaned[model_key] = cap_cleaned
+        new_val = json.dumps(cleaned) if cleaned else None
+        conn.execute(
+            sa.text("UPDATE users SET model_capability_settings = :val WHERE id = :id"),
+            {"val": new_val, "id": row[0]},
+        )
+
+    # GlobalModel.supported_capabilities: JSON array, remove stale entries
+    rows = conn.execute(
+        sa.text(
+            "SELECT id, supported_capabilities FROM global_models"
+            " WHERE supported_capabilities IS NOT NULL"
+        )
+    ).fetchall()
+    for row in rows:
+        raw = row[1]
+        if raw is None:
+            continue
+        data = raw if isinstance(raw, list) else json.loads(raw)
+        cleaned = [c for c in data if c not in stale_keys]
+        new_val = json.dumps(cleaned) if cleaned else None
+        conn.execute(
+            sa.text("UPDATE global_models SET supported_capabilities = :val WHERE id = :id"),
+            {"val": new_val, "id": row[0]},
+        )
+
+    # --- 3. Create user_model_usage_counts table ---
+    if not table_exists("user_model_usage_counts"):
+        op.create_table(
+            "user_model_usage_counts",
+            sa.Column("id", sa.String(36), primary_key=True),
+            sa.Column(
+                "user_id",
+                sa.String(36),
+                sa.ForeignKey("users.id", ondelete="CASCADE"),
+                nullable=False,
+            ),
+            sa.Column("model", sa.String(100), nullable=False),
+            sa.Column("usage_count", sa.Integer, nullable=False, server_default="0"),
+            sa.Column(
+                "created_at",
+                sa.DateTime(timezone=True),
+                nullable=False,
+                server_default=sa.func.now(),
+            ),
+            sa.Column(
+                "updated_at",
+                sa.DateTime(timezone=True),
+                nullable=False,
+                server_default=sa.func.now(),
+            ),
+            sa.UniqueConstraint("user_id", "model", name="uq_user_model_usage_count"),
+        )
+    if not index_exists("user_model_usage_counts", "idx_user_model_usage_user"):
+        op.create_index("idx_user_model_usage_user", "user_model_usage_counts", ["user_id"])
+    if not index_exists("user_model_usage_counts", "idx_user_model_usage_model"):
+        op.create_index("idx_user_model_usage_model", "user_model_usage_counts", ["model"])
+
+    # Backfill from existing usage records (truncate first for idempotency)
+    conn.execute(sa.text("DELETE FROM user_model_usage_counts"))
+    rows = conn.execute(
+        sa.text(
+            "SELECT user_id, model, COUNT(*) AS cnt FROM usage"
+            " WHERE user_id IS NOT NULL GROUP BY user_id, model"
+        )
+    ).fetchall()
+    now = datetime.now(timezone.utc)
+    for row in rows:
+        conn.execute(
+            sa.text(
+                "INSERT INTO user_model_usage_counts"
+                " (id, user_id, model, usage_count, created_at, updated_at)"
+                " VALUES (:id, :user_id, :model, :cnt, :now, :now)"
+            ),
+            {
+                "id": str(uuid.uuid4()),
+                "user_id": row[0],
+                "model": row[1],
+                "cnt": row[2],
+                "now": now,
+            },
+        )
+
+    # --- 4. Enforce models.global_model_id NOT NULL ---
+    conn = op.get_bind()
+    insp = inspect(conn)
+    model_cols = {c["name"]: c for c in insp.get_columns("models")}
+    if model_cols.get("global_model_id", {}).get("nullable", True):
+        op.execute("DELETE FROM models WHERE global_model_id IS NULL")
+        op.alter_column("models", "global_model_id", existing_type=sa.String(36), nullable=False)
+
+
+def downgrade() -> None:
+    # Revert models.global_model_id to nullable
+    if column_exists("models", "global_model_id"):
+        op.alter_column("models", "global_model_id", existing_type=sa.String(36), nullable=True)
+
+    # Drop user_model_usage_counts
+    if table_exists("user_model_usage_counts"):
+        if index_exists("user_model_usage_counts", "idx_user_model_usage_model"):
+            op.drop_index("idx_user_model_usage_model", table_name="user_model_usage_counts")
+        if index_exists("user_model_usage_counts", "idx_user_model_usage_user"):
+            op.drop_index("idx_user_model_usage_user", table_name="user_model_usage_counts")
+        op.drop_table("user_model_usage_counts")
+
+    # Drop cache_creation columns
+    if column_exists("usage", "cache_creation_input_tokens_1h"):
+        op.drop_column("usage", "cache_creation_input_tokens_1h")
+    if column_exists("usage", "cache_creation_input_tokens_5m"):
+        op.drop_column("usage", "cache_creation_input_tokens_5m")
+    # capability settings cleanup is not reversible
@@ -0,0 +1,153 @@
+"""proxy_node_metrics_and_events
+
+Revision ID: 48afe197cc15
+Revises: b2c3d4e5f6a7
+Create Date: 2026-02-28 04:33:11.201185+00:00
+
+"""
+
+import sqlalchemy as sa
+from sqlalchemy import inspect
+
+from alembic import op
+
+# revision identifiers, used by Alembic.
+revision = "48afe197cc15"
+down_revision = "b2c3d4e5f6a7"
+branch_labels = None
+depends_on = None
+
+
+def _column_exists(table_name: str, column_name: str) -> bool:
+    bind = op.get_bind()
+    insp = inspect(bind)
+    columns = [c["name"] for c in insp.get_columns(table_name)]
+    return column_name in columns
+
+
+def _table_exists(table_name: str) -> bool:
+    bind = op.get_bind()
+    insp = inspect(bind)
+    return table_name in insp.get_table_names()
+
+
+def _enum_has_value(enum_name: str, value: str) -> bool:
+    """检查 PostgreSQL 枚举类型是否包含指定值"""
+    bind = op.get_bind()
+    result = bind.execute(
+        sa.text(
+            "SELECT 1 FROM pg_enum e JOIN pg_type t ON e.enumtypid = t.oid"
+            " WHERE t.typname = :enum_name AND e.enumlabel = :value"
+        ),
+        {"enum_name": enum_name, "value": value},
+    )
+    return result.fetchone() is not None
+
+
+def upgrade() -> None:
+    # proxy_nodes: 将已废弃的 unhealthy 状态迁移为 offline，然后从枚举中移除
+    if _enum_has_value("proxynodestatus", "unhealthy"):
+        op.execute("UPDATE proxy_nodes SET status = 'offline' WHERE status = 'unhealthy'")
+        op.execute("ALTER TYPE proxynodestatus RENAME TO proxynodestatus_old")
+        op.execute("CREATE TYPE proxynodestatus AS ENUM ('online', 'offline')")
+        # 必须先移除旧枚举类型的 DEFAULT，否则 ALTER TYPE 会因无法转换默认值而报错
+        op.execute("ALTER TABLE proxy_nodes ALTER COLUMN status DROP DEFAULT")
+        op.execute(
+            "ALTER TABLE proxy_nodes ALTER COLUMN status TYPE proxynodestatus"
+            " USING status::text::proxynodestatus"
+        )
+        op.execute("ALTER TABLE proxy_nodes ALTER COLUMN status SET DEFAULT 'online'::proxynodestatus")
+        op.execute("DROP TYPE proxynodestatus_old")
+
+    # proxy_nodes: 新增错误指标字段
+    if not _column_exists("proxy_nodes", "failed_requests"):
+        op.add_column(
+            "proxy_nodes",
+            sa.Column(
+                "failed_requests",
+                sa.BigInteger(),
+                nullable=False,
+                server_default="0",
+                comment="累计失败请求数",
+            ),
+        )
+    if not _column_exists("proxy_nodes", "dns_failures"):
+        op.add_column(
+            "proxy_nodes",
+            sa.Column(
+                "dns_failures",
+                sa.BigInteger(),
+                nullable=False,
+                server_default="0",
+                comment="累计 DNS 失败数",
+            ),
+        )
+    if not _column_exists("proxy_nodes", "stream_errors"):
+        op.add_column(
+            "proxy_nodes",
+            sa.Column(
+                "stream_errors",
+                sa.BigInteger(),
+                nullable=False,
+                server_default="0",
+                comment="累计流错误数",
+            ),
+        )
+
+    # proxy_node_events: 连接事件表
+    if not _table_exists("proxy_node_events"):
+        op.create_table(
+            "proxy_node_events",
+            sa.Column("id", sa.BigInteger(), autoincrement=True, nullable=False),
+            sa.Column("node_id", sa.String(length=36), nullable=False),
+            sa.Column(
+                "event_type",
+                sa.String(length=20),
+                nullable=False,
+                comment="事件类型: connected, disconnected, error",
+            ),
+            sa.Column(
+                "detail",
+                sa.String(length=500),
+                nullable=True,
+                comment="事件详情（如断开原因）",
+            ),
+            sa.Column("created_at", sa.DateTime(timezone=True), nullable=False),
+            sa.ForeignKeyConstraint(["node_id"], ["proxy_nodes.id"], ondelete="CASCADE"),
+            sa.PrimaryKeyConstraint("id"),
+        )
+        op.create_index(
+            "idx_proxy_node_events_node_created",
+            "proxy_node_events",
+            ["node_id", "created_at"],
+        )
+        op.create_index(
+            op.f("ix_proxy_node_events_node_id"),
+            "proxy_node_events",
+            ["node_id"],
+        )
+
+
+def downgrade() -> None:
+    # 恢复 proxynodestatus 枚举，加回 unhealthy
+    if not _enum_has_value("proxynodestatus", "unhealthy"):
+        op.execute("ALTER TYPE proxynodestatus RENAME TO proxynodestatus_old")
+        op.execute("CREATE TYPE proxynodestatus AS ENUM ('online', 'unhealthy', 'offline')")
+        op.execute("ALTER TABLE proxy_nodes ALTER COLUMN status DROP DEFAULT")
+        op.execute(
+            "ALTER TABLE proxy_nodes ALTER COLUMN status TYPE proxynodestatus"
+            " USING status::text::proxynodestatus"
+        )
+        op.execute("ALTER TABLE proxy_nodes ALTER COLUMN status SET DEFAULT 'online'::proxynodestatus")
+        op.execute("DROP TYPE proxynodestatus_old")
+
+    if _table_exists("proxy_node_events"):
+        op.drop_index(op.f("ix_proxy_node_events_node_id"), table_name="proxy_node_events")
+        op.drop_index("idx_proxy_node_events_node_created", table_name="proxy_node_events")
+        op.drop_table("proxy_node_events")
+    if _column_exists("proxy_nodes", "stream_errors"):
+        op.drop_column("proxy_nodes", "stream_errors")
+    if _column_exists("proxy_nodes", "dns_failures"):
+        op.drop_column("proxy_nodes", "dns_failures")
+    if _column_exists("proxy_nodes", "failed_requests"):
+        op.drop_column("proxy_nodes", "failed_requests")
@@ -0,0 +1,49 @@
+"""add_request_candidates_composite_indexes
+
+Revision ID: 00b9161b8729
+Revises: 48afe197cc15
+Create Date: 2026-02-28 14:48:00.000000+00:00
+
+"""
+
+from sqlalchemy import inspect
+
+from alembic import op
+
+# revision identifiers, used by Alembic.
+revision = "00b9161b8729"
+down_revision = "48afe197cc15"
+branch_labels = None
+depends_on = None
+
+
+def _index_exists(index_name: str) -> bool:
+    bind = op.get_bind()
+    insp = inspect(bind)
+    indexes = insp.get_indexes("request_candidates")
+    return any(idx["name"] == index_name for idx in indexes)
+
+
+def upgrade() -> None:
+    # (request_id, status) - fallback/retry 查询优化
+    if not _index_exists("idx_rc_request_id_status"):
+        op.create_index(
+            "idx_rc_request_id_status",
+            "request_candidates",
+            ["request_id", "status"],
+        )
+
+    # (provider_id, status, created_at) - provider 聚合统计优化
+    if not _index_exists("idx_rc_provider_status_created"):
+        op.create_index(
+            "idx_rc_provider_status_created",
+            "request_candidates",
+            ["provider_id", "status", "created_at"],
+        )
+
+
+def downgrade() -> None:
+    if _index_exists("idx_rc_provider_status_created"):
+        op.drop_index("idx_rc_provider_status_created", table_name="request_candidates")
+    if _index_exists("idx_rc_request_id_status"):
+        op.drop_index("idx_rc_request_id_status", table_name="request_candidates")
@@ -0,0 +1,263 @@
+"""vertex_ai_provider_type
+
+Migrate legacy Vertex auth_type/provider_type into the new model:
+- provider_type=vertex_ai
+- auth_type=service_account (legacy vertex_ai renamed)
+- fixed Vertex endpoints: gemini:chat + claude:chat
+
+Revision ID: 2a624af8dd3a
+Revises: 00b9161b8729
+Create Date: 2026-02-28 15:00:00.000000+00:00
+"""
+
+from __future__ import annotations
+
+import uuid
+
+import sqlalchemy as sa
+
+from alembic import op
+
+# revision identifiers, used by Alembic.
+revision = "2a624af8dd3a"
+down_revision = "00b9161b8729"
+branch_labels = None
+depends_on = None
+
+_VERTEX_BASE_URL = "https://aiplatform.googleapis.com"
+_VERTEX_ENDPOINTS: tuple[tuple[str, str, str], ...] = (
+    ("gemini:chat", "gemini", "chat"),
+    ("claude:chat", "claude", "chat"),
+)
+_VERTEX_KEY_FORMATS_SA = '["gemini:chat","claude:chat"]'
+_VERTEX_KEY_FORMATS_API_KEY = '["gemini:chat"]'
+
+
+def _select_vertex_provider_ids(conn: sa.Connection) -> list[str]:
+    """Collect providers that should be treated as Vertex after migration."""
+    rows = conn.execute(sa.text("""
+            SELECT DISTINCT p.id
+            FROM providers p
+            LEFT JOIN provider_api_keys pak ON pak.provider_id = p.id
+            WHERE lower(COALESCE(p.provider_type, '')) = 'vertex_ai'
+               OR pak.auth_type = 'vertex_ai'
+        """))
+    return [str(row[0]) for row in rows if row[0]]
+
+
+def _ensure_fixed_vertex_endpoints(conn: sa.Connection, provider_ids: list[str]) -> None:
+    """Ensure every Vertex provider has fixed gemini:chat + claude:chat endpoints."""
+    for provider_id in provider_ids:
+        provider_max_retries = (
+            conn.execute(
+                sa.text("""
+                    SELECT COALESCE(max_retries, 2)
+                    FROM providers
+                    WHERE id = :provider_id
+                """),
+                {"provider_id": provider_id},
+            ).scalar()
+            or 2
+        )
+
+        for api_format, api_family, endpoint_kind in _VERTEX_ENDPOINTS:
+            # Normalize existing fixed endpoint fields.
+            conn.execute(
+                sa.text("""
+                    UPDATE provider_endpoints
+                    SET
+                        api_family = :api_family,
+                        endpoint_kind = :endpoint_kind,
+                        base_url = :base_url,
+                        custom_path = NULL,
+                        is_active = TRUE,
+                        updated_at = CURRENT_TIMESTAMP
+                    WHERE provider_id = :provider_id
+                      AND api_format = :api_format
+                """),
+                {
+                    "provider_id": provider_id,
+                    "api_format": api_format,
+                    "api_family": api_family,
+                    "endpoint_kind": endpoint_kind,
+                    "base_url": _VERTEX_BASE_URL,
+                },
+            )
+
+            exists = conn.execute(
+                sa.text("""
+                    SELECT 1
+                    FROM provider_endpoints
+                    WHERE provider_id = :provider_id
+                      AND api_format = :api_format
+                    LIMIT 1
+                """),
+                {"provider_id": provider_id, "api_format": api_format},
+            ).first()
+
+            if not exists:
+                conn.execute(
+                    sa.text("""
+                        INSERT INTO provider_endpoints (
+                            id,
+                            provider_id,
+                            api_format,
+                            api_family,
+                            endpoint_kind,
+                            base_url,
+                            custom_path,
+                            header_rules,
+                            body_rules,
+                            max_retries,
+                            is_active,
+                            config,
+                            format_acceptance_config,
+                            proxy,
+                            created_at,
+                            updated_at
+                        )
+                        VALUES (
+                            :id,
+                            :provider_id,
+                            :api_format,
+                            :api_family,
+                            :endpoint_kind,
+                            :base_url,
+                            NULL,
+                            NULL,
+                            NULL,
+                            :max_retries,
+                            TRUE,
+                            NULL,
+                            NULL,
+                            NULL,
+                            CURRENT_TIMESTAMP,
+                            CURRENT_TIMESTAMP
+                        )
+                    """),
+                    {
+                        "id": str(uuid.uuid4()),
+                        "provider_id": provider_id,
+                        "api_format": api_format,
+                        "api_family": api_family,
+                        "endpoint_kind": endpoint_kind,
+                        "base_url": _VERTEX_BASE_URL,
+                        "max_retries": int(provider_max_retries),
+                    },
+                )
+
+        # Vertex fixed-provider model: disable non-fixed endpoints.
+        conn.execute(
+            sa.text("""
+                UPDATE provider_endpoints
+                SET
+                    is_active = FALSE,
+                    updated_at = CURRENT_TIMESTAMP
+                WHERE provider_id = :provider_id
+                  AND api_format NOT IN ('gemini:chat', 'claude:chat')
+            """),
+            {"provider_id": provider_id},
+        )
+
+
+def _normalize_vertex_key_formats(conn: sa.Connection, provider_ids: list[str]) -> None:
+    """Normalize key.api_formats for Vertex keys by auth type."""
+    for provider_id in provider_ids:
+        # Service Account (and legacy vertex_ai) keys: allow Gemini + Claude models.
+        conn.execute(
+            sa.text("""
+                UPDATE provider_api_keys
+                SET
+                    api_formats = CAST(:api_formats AS json),
+                    updated_at = CURRENT_TIMESTAMP
+                WHERE provider_id = :provider_id
+                  AND auth_type IN ('service_account', 'vertex_ai')
+            """),
+            {
+                "provider_id": provider_id,
+                "api_formats": _VERTEX_KEY_FORMATS_SA,
+            },
+        )
+
+        # API Key mode on Vertex 仅支持 Gemini（Google publisher）。
+        conn.execute(
+            sa.text("""
+                UPDATE provider_api_keys
+                SET
+                    api_formats = CAST(:api_formats AS json),
+                    updated_at = CURRENT_TIMESTAMP
+                WHERE provider_id = :provider_id
+                  AND auth_type = 'api_key'
+            """),
+            {
+                "provider_id": provider_id,
+                "api_formats": _VERTEX_KEY_FORMATS_API_KEY,
+            },
+        )
+
+
+def upgrade() -> None:
+    conn = op.get_bind()
+
+    # 1) 收集目标 Provider（兼容重复执行，先识别 legacy/new 两种来源）。
+    provider_ids = _select_vertex_provider_ids(conn)
+
+    # 2) 先重命名 auth_type（legacy vertex_ai -> service_account）。
+    conn.execute(sa.text("""
+            UPDATE provider_api_keys
+            SET auth_type = 'service_account'
+            WHERE auth_type = 'vertex_ai'
+        """))
+
+    if not provider_ids:
+        return
+
+    # 3) 归一 provider_type，并启用格式转换（Vertex 同时承载 Gemini/Claude）。
+    for provider_id in provider_ids:
+        conn.execute(
+            sa.text("""
+                UPDATE providers
+                SET
+                    provider_type = 'vertex_ai',
+                    enable_format_conversion = TRUE
+                WHERE id = :provider_id
+            """),
+            {"provider_id": provider_id},
+        )
+
+    # 4) 固定端点落地：gemini:chat + claude:chat。
+    _ensure_fixed_vertex_endpoints(conn, provider_ids)
+
+    # 5) 归一 key 的 api_formats，避免调度命中旧格式。
+    _normalize_vertex_key_formats(conn, provider_ids)
+
+
+def downgrade() -> None:
+    conn = op.get_bind()
+
+    provider_rows = conn.execute(sa.text("""
+            SELECT id
+            FROM providers
+            WHERE lower(COALESCE(provider_type, '')) = 'vertex_ai'
+        """))
+    provider_ids = [str(row[0]) for row in provider_rows if row[0]]
+
+    if provider_ids:
+        for provider_id in provider_ids:
+            conn.execute(
+                sa.text("""
+                    UPDATE provider_api_keys
+                    SET auth_type = 'vertex_ai'
+                    WHERE provider_id = :provider_id
+                      AND auth_type = 'service_account'
+                """),
+                {"provider_id": provider_id},
+            )
+            conn.execute(
+                sa.text("""
+                    UPDATE providers
+                    SET provider_type = 'custom'
+                    WHERE id = :provider_id
+                """),
+                {"provider_id": provider_id},
+            )
@@ -0,0 +1,199 @@
+"""backfill_codex_compact_endpoint
+
+Backfill Codex reverse-proxy endpoints:
+- ensure `openai:cli` endpoint is pinned to force_stream
+- ensure `openai:compact` endpoint exists
+
+Revision ID: f0c3a7b9d1e2
+Revises: 2a624af8dd3a
+Create Date: 2026-03-01 17:00:00.000000+00:00
+"""
+
+from __future__ import annotations
+
+import json
+import uuid
+from typing import Any
+
+import sqlalchemy as sa
+
+from alembic import op
+
+# revision identifiers, used by Alembic.
+revision = "f0c3a7b9d1e2"
+down_revision = "2a624af8dd3a"
+branch_labels = None
+depends_on = None
+
+_CODEX_BASE_URL = "https://chatgpt.com/backend-api/codex"
+_COMPACT_FORMAT = "openai:compact"
+_CLI_FORMAT = "openai:cli"
+_FORCE_STREAM = "force_stream"
+
+
+def _find_codex_provider_ids(conn: sa.Connection) -> list[str]:
+    """Find Codex providers (by provider_type or legacy base_url pattern)."""
+    rows = conn.execute(sa.text("""
+        SELECT DISTINCT p.id
+        FROM providers p
+        LEFT JOIN provider_endpoints pe ON pe.provider_id = p.id
+        WHERE lower(COALESCE(p.provider_type, '')) = 'codex'
+           OR (
+                lower(COALESCE(pe.api_format, '')) = 'openai:cli'
+            AND lower(COALESCE(pe.base_url, '')) LIKE '%/backend-api/codex%'
+           )
+    """))
+    return [str(r[0]) for r in rows if r[0]]
+
+
+def _get_cli_endpoint(conn: sa.Connection, provider_id: str) -> dict[str, Any] | None:
+    """Load existing openai:cli endpoint for the provider."""
+    row = (
+        conn.execute(
+            sa.text("""
+                SELECT base_url, header_rules, body_rules, max_retries, proxy, config
+                FROM provider_endpoints
+                WHERE provider_id = :pid AND api_format = :fmt
+                LIMIT 1
+            """),
+            {"pid": provider_id, "fmt": _CLI_FORMAT},
+        )
+        .mappings()
+        .first()
+    )
+    return dict(row) if row else None
+
+
+def _pin_cli_force_stream(conn: sa.Connection, provider_id: str, cli: dict[str, Any]) -> None:
+    """Set upstream_stream_policy=force_stream on existing cli endpoint."""
+    cfg = dict(cli.get("config") or {}) if isinstance(cli.get("config"), dict) else {}
+    cfg.pop("upstreamStreamPolicy", None)
+    cfg.pop("upstream_stream", None)
+    cfg["upstream_stream_policy"] = _FORCE_STREAM
+    conn.execute(
+        sa.text("""
+            UPDATE provider_endpoints
+            SET api_family = 'openai',
+                endpoint_kind = 'cli',
+                config = CAST(:config AS json),
+                updated_at = CURRENT_TIMESTAMP
+            WHERE provider_id = :pid AND api_format = :fmt
+        """),
+        {
+            "pid": provider_id,
+            "fmt": _CLI_FORMAT,
+            "config": json.dumps(cfg, ensure_ascii=False),
+        },
+    )
+
+
+def _ensure_compact_endpoint(conn: sa.Connection, provider_id: str, cli: dict[str, Any]) -> None:
+    """Create openai:compact endpoint if missing (clone from cli)."""
+    exists = conn.execute(
+        sa.text(
+            "SELECT 1 FROM provider_endpoints WHERE provider_id = :pid AND api_format = :fmt LIMIT 1"
+        ),
+        {"pid": provider_id, "fmt": _COMPACT_FORMAT},
+    ).first()
+    if exists:
+        # Already exists, just ensure api_family/endpoint_kind are set.
+        conn.execute(
+            sa.text("""
+                UPDATE provider_endpoints
+                SET api_family = 'openai', endpoint_kind = 'compact',
+                    updated_at = CURRENT_TIMESTAMP
+                WHERE provider_id = :pid AND api_format = :fmt
+            """),
+            {"pid": provider_id, "fmt": _COMPACT_FORMAT},
+        )
+        return
+
+    # Clone from cli endpoint, strip stream policy.
+    cfg = dict(cli.get("config") or {}) if isinstance(cli.get("config"), dict) else {}
+    for k in ("upstream_stream_policy", "upstreamStreamPolicy", "upstream_stream"):
+        cfg.pop(k, None)
+
+    def _json(val: Any) -> str | None:
+        return json.dumps(val, ensure_ascii=False) if val is not None else None
+
+    conn.execute(
+        sa.text("""
+            INSERT INTO provider_endpoints (
+                id, provider_id, api_format, api_family, endpoint_kind,
+                base_url, custom_path, header_rules, body_rules,
+                max_retries, is_active, config, format_acceptance_config,
+                proxy, created_at, updated_at
+            ) VALUES (
+                :id, :pid, :fmt, 'openai', 'compact',
+                :base_url, NULL, CAST(:header_rules AS json), CAST(:body_rules AS json),
+                :max_retries, TRUE, CAST(:config AS json), NULL,
+                CAST(:proxy AS jsonb), CURRENT_TIMESTAMP, CURRENT_TIMESTAMP
+            )
+        """),
+        {
+            "id": str(uuid.uuid4()),
+            "pid": provider_id,
+            "fmt": _COMPACT_FORMAT,
+            "base_url": cli.get("base_url") or _CODEX_BASE_URL,
+            "header_rules": _json(cli.get("header_rules")),
+            "body_rules": _json(cli.get("body_rules")),
+            "max_retries": cli.get("max_retries") or 2,
+            "config": _json(cfg or None),
+            "proxy": _json(cli.get("proxy")),
+        },
+    )
+
+
+def _add_compact_to_key_formats(conn: sa.Connection, provider_id: str) -> None:
+    """Ensure provider keys include openai:compact in api_formats."""
+    rows = (
+        conn.execute(
+            sa.text("SELECT id, api_formats FROM provider_api_keys WHERE provider_id = :pid"),
+            {"pid": provider_id},
+        )
+        .mappings()
+        .all()
+    )
+    for row in rows:
+        raw = row["api_formats"]
+        formats: list[str] = []
+        if isinstance(raw, list):
+            for item in raw:
+                v = str(item or "").strip().lower()
+                if v and v not in formats:
+                    formats.append(v)
+
+        if _COMPACT_FORMAT in formats:
+            continue
+
+        # Insert compact right after cli, or at end.
+        if _CLI_FORMAT in formats:
+            idx = formats.index(_CLI_FORMAT) + 1
+            formats.insert(idx, _COMPACT_FORMAT)
+        else:
+            formats.append(_COMPACT_FORMAT)
+
+        conn.execute(
+            sa.text("""
+                UPDATE provider_api_keys
+                SET api_formats = CAST(:fmts AS json), updated_at = CURRENT_TIMESTAMP
+                WHERE id = :id
+            """),
+            {"id": row["id"], "fmts": json.dumps(formats, ensure_ascii=False)},
+        )
+
+
+def upgrade() -> None:
+    conn = op.get_bind()
+    for provider_id in _find_codex_provider_ids(conn):
+        cli = _get_cli_endpoint(conn, provider_id)
+        if not cli:
+            continue  # No cli endpoint to clone from; skip.
+        _pin_cli_force_stream(conn, provider_id, cli)
+        _ensure_compact_endpoint(conn, provider_id, cli)
+        _add_compact_to_key_formats(conn, provider_id)
+
+
+def downgrade() -> None:
+    # Data backfill: no-op to avoid deleting user-managed data.
+    return
@@ -0,0 +1,29 @@
+"""add_proxy_metadata_to_proxy_nodes
+
+Revision ID: 1d2e3f4a5b6c
+Revises: f0c3a7b9d1e2
+Create Date: 2026-03-02 13:00:00.000000+00:00
+"""
+
+from __future__ import annotations
+
+from collections.abc import Sequence
+
+from alembic import op
+
+# revision identifiers, used by Alembic.
+revision: str = "1d2e3f4a5b6c"
+down_revision: str | None = "f0c3a7b9d1e2"
+branch_labels: str | Sequence[str] | None = None
+depends_on: str | Sequence[str] | None = None
+
+
+def upgrade() -> None:
+    op.execute("ALTER TABLE public.proxy_nodes ADD COLUMN IF NOT EXISTS proxy_metadata json")
+    op.execute(
+        "COMMENT ON COLUMN public.proxy_nodes.proxy_metadata IS 'aether-proxy 上报元数据（版本等）'"
+    )
+
+
+def downgrade() -> None:
+    op.execute("ALTER TABLE public.proxy_nodes DROP COLUMN IF EXISTS proxy_metadata")
@@ -0,0 +1,72 @@
+"""backfill_codex_default_body_rules
+
+Backfill default body_rules for codex providers with openai:cli endpoints
+that currently have body_rules IS NULL.
+
+Rules:
+- drop max_output_tokens
+- drop temperature
+- drop top_p
+- set store = false
+- set instructions = "You are GPT-5." (when instructions not exists)
+
+Revision ID: dd0278c0a28c
+Revises: 1d2e3f4a5b6c
+Create Date: 2026-03-02 15:00:00.000000+00:00
+"""
+
+from __future__ import annotations
+
+import json
+
+import sqlalchemy as sa
+
+from alembic import op
+
+# revision identifiers, used by Alembic.
+revision = "dd0278c0a28c"
+down_revision = "1d2e3f4a5b6c"
+branch_labels = None
+depends_on = None
+
+_TARGET_FORMATS = ("openai:cli",)
+
+_DEFAULT_BODY_RULES = [
+    {"action": "drop", "path": "max_output_tokens"},
+    {"action": "drop", "path": "temperature"},
+    {"action": "drop", "path": "top_p"},
+    {"action": "set", "path": "store", "value": False},
+    {
+        "action": "set",
+        "path": "instructions",
+        "value": "You are GPT-5.",
+        "condition": {"path": "instructions", "op": "not_exists"},
+    },
+]
+
+
+def upgrade() -> None:
+    conn = op.get_bind()
+
+    # 幂等性: 仅回填 codex 提供商中 body_rules 为空（SQL NULL 或 JSON null）的记录
+    rules_json = json.dumps(_DEFAULT_BODY_RULES, ensure_ascii=False)
+    result = conn.execute(
+        sa.text("""
+            UPDATE provider_endpoints pe
+            SET body_rules = CAST(:rules AS json),
+                updated_at = CURRENT_TIMESTAMP
+            FROM providers p
+            WHERE pe.provider_id = p.id
+              AND p.provider_type = :ptype
+              AND pe.api_format = :fmt
+              AND (pe.body_rules IS NULL OR pe.body_rules::text = 'null')
+        """),
+        {"rules": rules_json, "ptype": "codex", "fmt": _TARGET_FORMATS[0]},
+    )
+    if result.rowcount:
+        print(f"  backfilled body_rules for {result.rowcount} endpoint(s)")
+
+
+def downgrade() -> None:
+    # Data backfill: no-op to avoid removing user-customized rules.
+    return
@@ -0,0 +1,45 @@
+"""add_idx_usage_provider_key
+
+Add composite index on usage(provider_id, provider_api_key_id) to support
+the pool management page's per-key usage stats aggregation query.
+
+Revision ID: 0ba031f328de
+Revises: dd0278c0a28c
+Create Date: 2026-03-03 10:00:00.000000+00:00
+"""
+
+from __future__ import annotations
+
+import sqlalchemy as sa
+from alembic import op
+
+revision = "0ba031f328de"
+down_revision = "dd0278c0a28c"
+branch_labels = None
+depends_on = None
+
+INDEX_NAME = "idx_usage_provider_key"
+TABLE = "usage"
+COLUMNS = ["provider_id", "provider_api_key_id"]
+
+
+def upgrade() -> None:
+    bind = op.get_bind()
+    result = bind.execute(
+        sa.text("SELECT 1 FROM pg_indexes WHERE indexname = :name"),
+        {"name": INDEX_NAME},
+    ).fetchone()
+    if result:
+        return
+    op.create_index(INDEX_NAME, TABLE, COLUMNS)
+
+
+def downgrade() -> None:
+    bind = op.get_bind()
+    result = bind.execute(
+        sa.text("SELECT 1 FROM pg_indexes WHERE indexname = :name"),
+        {"name": INDEX_NAME},
+    ).fetchone()
+    if not result:
+        return
+    op.drop_index(INDEX_NAME, table_name=TABLE)
@@ -0,0 +1,45 @@
+"""add_idx_usage_status_user_created
+
+Add composite index on usage(status, user_id, created_at) to speed up
+interval timeline and active usage analytics queries.
+
+Revision ID: 5f1d2e3c4b5a
+Revises: 0ba031f328de
+Create Date: 2026-03-03 17:30:00.000000+00:00
+"""
+
+from __future__ import annotations
+
+import sqlalchemy as sa
+from alembic import op
+
+revision = "5f1d2e3c4b5a"
+down_revision = "0ba031f328de"
+branch_labels = None
+depends_on = None
+
+INDEX_NAME = "idx_usage_status_user_created"
+TABLE = "usage"
+COLUMNS = ["status", "user_id", "created_at"]
+
+
+def upgrade() -> None:
+    bind = op.get_bind()
+    result = bind.execute(
+        sa.text("SELECT 1 FROM pg_indexes WHERE indexname = :name"),
+        {"name": INDEX_NAME},
+    ).fetchone()
+    if result:
+        return
+    op.create_index(INDEX_NAME, TABLE, COLUMNS)
+
+
+def downgrade() -> None:
+    bind = op.get_bind()
+    result = bind.execute(
+        sa.text("SELECT 1 FROM pg_indexes WHERE indexname = :name"),
+        {"name": INDEX_NAME},
+    ).fetchone()
+    if not result:
+        return
+    op.drop_index(INDEX_NAME, table_name=TABLE)
@@ -0,0 +1,41 @@
+"""add fingerprint column to provider_api_keys
+
+Revision ID: 6a9b8c7d5e4f
+Revises: 5f1d2e3c4b5a
+Create Date: 2026-03-04 23:50:00.000000
+"""
+
+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 = "6a9b8c7d5e4f"
+down_revision: str | None = "5f1d2e3c4b5a"
+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", "fingerprint"):
+        op.add_column(
+            "provider_api_keys",
+            sa.Column("fingerprint", sa.JSON(), nullable=True),
+        )
+
+
+def downgrade() -> None:
+    if column_exists("provider_api_keys", "fingerprint"):
+        op.drop_column("provider_api_keys", "fingerprint")
@@ -0,0 +1,65 @@
+"""remove standalone api key locking
+
+Revision ID: 7c91d2e4f8a1
+Revises: 6f7a8b9c0d1e
+Create Date: 2026-03-05 17:00:00.000000+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 = "7c91d2e4f8a1"
+down_revision: str | None = "6f7a8b9c0d1e"
+branch_labels: str | Sequence[str] | None = None
+depends_on: str | Sequence[str] | None = None
+
+_CONSTRAINT_NAME = "ck_api_keys_standalone_not_locked"
+
+
+def _column_exists(table_name: str, column_name: str) -> bool:
+    bind = op.get_bind()
+    insp = inspect(bind)
+    insp.clear_cache()
+    return column_name in [c["name"] for c in insp.get_columns(table_name)]
+
+
+def _check_constraint_exists(table_name: str, constraint_name: str) -> bool:
+    bind = op.get_bind()
+    insp = inspect(bind)
+    insp.clear_cache()
+    return any(c.get("name") == constraint_name for c in insp.get_check_constraints(table_name))
+
+
+def upgrade() -> None:
+    if not (
+        _column_exists("api_keys", "is_standalone")
+        and _column_exists("api_keys", "is_locked")
+        and _column_exists("api_keys", "is_active")
+    ):
+        return
+
+    op.execute(sa.text("""
+            UPDATE api_keys
+            SET is_active = FALSE,
+                is_locked = FALSE
+            WHERE is_standalone IS TRUE AND is_locked IS TRUE
+            """))
+
+    if not _check_constraint_exists("api_keys", _CONSTRAINT_NAME):
+        op.create_check_constraint(
+            _CONSTRAINT_NAME,
+            "api_keys",
+            "(NOT is_standalone) OR (NOT is_locked)",
+        )
+
+
+def downgrade() -> None:
+    if _check_constraint_exists("api_keys", _CONSTRAINT_NAME):
+        op.drop_constraint(_CONSTRAINT_NAME, "api_keys", type_="check")
@@ -0,0 +1,220 @@
+"""tighten wallet transaction snapshots and remove wallet version
+
+Revision ID: 8e71f2a4c9b0
+Revises: 7c91d2e4f8a1
+Create Date: 2026-03-07 13:00:00.000000+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 = "8e71f2a4c9b0"
+down_revision: str | None = "7c91d2e4f8a1"
+branch_labels: str | Sequence[str] | None = None
+depends_on: str | Sequence[str] | None = None
+
+_WALLET_TX_BEFORE_CHECK = "ck_wallet_tx_balance_before_consistent"
+_WALLET_TX_AFTER_CHECK = "ck_wallet_tx_balance_after_consistent"
+_WALLET_LIMIT_MODE_INDEX = "idx_wallets_limit_mode"
+
+
+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 [c["name"] for c 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(index.get("name") == index_name for index in insp.get_indexes(table_name))
+
+
+def _check_constraint_exists(table_name: str, constraint_name: str) -> bool:
+    bind = op.get_bind()
+    insp = inspect(bind)
+    insp.clear_cache()
+    return any(c.get("name") == constraint_name for c in insp.get_check_constraints(table_name))
+
+
+def _tighten_wallet_transaction_snapshots() -> None:
+    if not _table_exists("wallet_transactions"):
+        return
+
+    required_columns = {
+        "balance_before",
+        "balance_after",
+        "recharge_balance_before",
+        "recharge_balance_after",
+        "gift_balance_before",
+        "gift_balance_after",
+    }
+    existing_columns = {
+        column["name"] for column in inspect(op.get_bind()).get_columns("wallet_transactions")
+    }
+    if not required_columns.issubset(existing_columns):
+        return
+
+    op.execute(
+        sa.text(
+            """
+            UPDATE wallet_transactions
+            SET recharge_balance_before = balance_before
+            WHERE recharge_balance_before IS NULL
+            """
+        )
+    )
+    op.execute(
+        sa.text(
+            """
+            UPDATE wallet_transactions
+            SET recharge_balance_after = balance_after
+            WHERE recharge_balance_after IS NULL
+            """
+        )
+    )
+    op.execute(
+        sa.text(
+            """
+            UPDATE wallet_transactions
+            SET gift_balance_before = 0
+            WHERE gift_balance_before IS NULL
+            """
+        )
+    )
+    op.execute(
+        sa.text(
+            """
+            UPDATE wallet_transactions
+            SET gift_balance_after = 0
+            WHERE gift_balance_after IS NULL
+            """
+        )
+    )
+    op.execute(
+        sa.text(
+            """
+            UPDATE wallet_transactions
+            SET balance_before = recharge_balance_before + gift_balance_before,
+                balance_after = recharge_balance_after + gift_balance_after
+            """
+        )
+    )
+
+    if not _check_constraint_exists("wallet_transactions", _WALLET_TX_BEFORE_CHECK):
+        op.create_check_constraint(
+            _WALLET_TX_BEFORE_CHECK,
+            "wallet_transactions",
+            "balance_before = recharge_balance_before + gift_balance_before",
+        )
+    if not _check_constraint_exists("wallet_transactions", _WALLET_TX_AFTER_CHECK):
+        op.create_check_constraint(
+            _WALLET_TX_AFTER_CHECK,
+            "wallet_transactions",
+            "balance_after = recharge_balance_after + gift_balance_after",
+        )
+
+    op.alter_column(
+        "wallet_transactions",
+        "recharge_balance_before",
+        existing_type=sa.Numeric(20, 8),
+        nullable=False,
+    )
+    op.alter_column(
+        "wallet_transactions",
+        "recharge_balance_after",
+        existing_type=sa.Numeric(20, 8),
+        nullable=False,
+    )
+    op.alter_column(
+        "wallet_transactions",
+        "gift_balance_before",
+        existing_type=sa.Numeric(20, 8),
+        nullable=False,
+    )
+    op.alter_column(
+        "wallet_transactions",
+        "gift_balance_after",
+        existing_type=sa.Numeric(20, 8),
+        nullable=False,
+    )
+
+
+def _drop_wallet_cleanup_artifacts() -> None:
+    if not _table_exists("wallets"):
+        return
+
+    if _index_exists("wallets", _WALLET_LIMIT_MODE_INDEX):
+        op.drop_index(_WALLET_LIMIT_MODE_INDEX, table_name="wallets")
+
+    if _column_exists("wallets", "version"):
+        op.drop_column("wallets", "version")
+
+
+def upgrade() -> None:
+    _tighten_wallet_transaction_snapshots()
+    _drop_wallet_cleanup_artifacts()
+
+
+def downgrade() -> None:
+    if _table_exists("wallets"):
+        if not _column_exists("wallets", "version"):
+            op.add_column(
+                "wallets",
+                sa.Column("version", sa.Integer(), nullable=False, server_default="0"),
+            )
+        if not _index_exists("wallets", _WALLET_LIMIT_MODE_INDEX):
+            op.create_index(_WALLET_LIMIT_MODE_INDEX, "wallets", ["limit_mode"])
+
+    if not _table_exists("wallet_transactions"):
+        return
+
+    if _column_exists("wallet_transactions", "recharge_balance_before"):
+        op.alter_column(
+            "wallet_transactions",
+            "recharge_balance_before",
+            existing_type=sa.Numeric(20, 8),
+            nullable=True,
+        )
+    if _column_exists("wallet_transactions", "recharge_balance_after"):
+        op.alter_column(
+            "wallet_transactions",
+            "recharge_balance_after",
+            existing_type=sa.Numeric(20, 8),
+            nullable=True,
+        )
+    if _column_exists("wallet_transactions", "gift_balance_before"):
+        op.alter_column(
+            "wallet_transactions",
+            "gift_balance_before",
+            existing_type=sa.Numeric(20, 8),
+            nullable=True,
+        )
+    if _column_exists("wallet_transactions", "gift_balance_after"):
+        op.alter_column(
+            "wallet_transactions",
+            "gift_balance_after",
+            existing_type=sa.Numeric(20, 8),
+            nullable=True,
+        )
+
+    if _check_constraint_exists("wallet_transactions", _WALLET_TX_AFTER_CHECK):
+        op.drop_constraint(_WALLET_TX_AFTER_CHECK, "wallet_transactions", type_="check")
+    if _check_constraint_exists("wallet_transactions", _WALLET_TX_BEFORE_CHECK):
+        op.drop_constraint(_WALLET_TX_BEFORE_CHECK, "wallet_transactions", type_="check")
@@ -0,0 +1,80 @@
+"""add missing foreign key indexes for cascade delete performance
+
+Revision ID: 2d932114930d
+Revises: 8e71f2a4c9b0
+Create Date: 2026-03-07 16:28:48.633531+00:00
+
+"""
+from alembic import op
+from sqlalchemy import inspect
+
+# revision identifiers, used by Alembic.
+revision = '2d932114930d'
+down_revision = '8e71f2a4c9b0'
+branch_labels = None
+depends_on = None
+
+
+def _index_exists(table_name: str, index_name: str) -> bool:
+    bind = op.get_bind()
+    insp = inspect(bind)
+    return any(idx["name"] == index_name for idx in insp.get_indexes(table_name))
+
+
+def _create_index_if_not_exists(index_name: str, table_name: str, columns: list[str]) -> None:
+    if not _index_exists(table_name, index_name):
+        op.create_index(op.f(index_name), table_name, columns, unique=False)
+
+
+def _drop_index_if_exists(index_name: str, table_name: str) -> None:
+    if _index_exists(table_name, index_name):
+        op.drop_index(op.f(index_name), table_name=table_name)
+
+
+# (index_name, table_name, columns)
+_INDEXES = [
+    # api_keys.user_id (CASCADE -> users.id)
+    ('ix_api_keys_user_id', 'api_keys', ['user_id']),
+    # usage: wallet_id, provider_endpoint_id, provider_api_key_id (SET NULL)
+    ('ix_usage_wallet_id', 'usage', ['wallet_id']),
+    ('ix_usage_provider_endpoint_id', 'usage', ['provider_endpoint_id']),
+    ('ix_usage_provider_api_key_id', 'usage', ['provider_api_key_id']),
+    # wallet_transactions.operator_id (SET NULL -> users.id)
+    ('ix_wallet_transactions_operator_id', 'wallet_transactions', ['operator_id']),
+    # payment_callbacks.payment_order_id (SET NULL -> payment_orders.id)
+    ('ix_payment_callbacks_payment_order_id', 'payment_callbacks', ['payment_order_id']),
+    # refund_requests: payment_order_id, requested_by, approved_by, processed_by (SET NULL)
+    ('ix_refund_requests_payment_order_id', 'refund_requests', ['payment_order_id']),
+    ('ix_refund_requests_requested_by', 'refund_requests', ['requested_by']),
+    ('ix_refund_requests_approved_by', 'refund_requests', ['approved_by']),
+    ('ix_refund_requests_processed_by', 'refund_requests', ['processed_by']),
+    # proxy_nodes.registered_by (SET NULL -> users.id)
+    ('ix_proxy_nodes_registered_by', 'proxy_nodes', ['registered_by']),
+    # video_tasks: api_key_id, provider_id, endpoint_id, key_id, remixed_from_task_id
+    ('ix_video_tasks_api_key_id', 'video_tasks', ['api_key_id']),
+    ('ix_video_tasks_provider_id', 'video_tasks', ['provider_id']),
+    ('ix_video_tasks_endpoint_id', 'video_tasks', ['endpoint_id']),
+    ('ix_video_tasks_key_id', 'video_tasks', ['key_id']),
+    ('ix_video_tasks_remixed_from_task_id', 'video_tasks', ['remixed_from_task_id']),
+    # user_preferences.default_provider_id (-> providers.id)
+    ('ix_user_preferences_default_provider_id', 'user_preferences', ['default_provider_id']),
+    # announcements.author_id (SET NULL -> users.id)
+    ('ix_announcements_author_id', 'announcements', ['author_id']),
+    # announcement_reads.announcement_id (-> announcements.id)
+    ('ix_announcement_reads_announcement_id', 'announcement_reads', ['announcement_id']),
+    # request_candidates: user_id, api_key_id, endpoint_id, key_id (CASCADE)
+    ('ix_request_candidates_user_id', 'request_candidates', ['user_id']),
+    ('ix_request_candidates_api_key_id', 'request_candidates', ['api_key_id']),
+    ('ix_request_candidates_endpoint_id', 'request_candidates', ['endpoint_id']),
+    ('ix_request_candidates_key_id', 'request_candidates', ['key_id']),
+]
+
+
+def upgrade() -> None:
+    for index_name, table_name, columns in _INDEXES:
+        _create_index_if_not_exists(index_name, table_name, columns)
+
+
+def downgrade() -> None:
+    for index_name, table_name, _columns in reversed(_INDEXES):
+        _drop_index_if_exists(index_name, table_name)
@@ -0,0 +1,224 @@
+"""usage stats retention: SET NULL on delete and add name snapshots
+
+Revision ID: 45b118150a78
+Revises: 2d932114930d
+Create Date: 2026-03-08 03:48:49.622091+00:00
+
+"""
+
+from __future__ import annotations
+
+import sqlalchemy as sa
+
+from alembic import op
+
+# revision identifiers, used by Alembic.
+revision = "45b118150a78"
+down_revision = "2d932114930d"
+branch_labels = None
+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 = _SchemaCache()
+    c.load_columns(_TABLES)
+    c.load_fk_rules(["stats_user_daily", "stats_daily_api_key"])
+
+    # --- Usage: add name snapshot columns ---
+    if not c.column_exists("usage", "username"):
+        op.add_column(
+            "usage", sa.Column("username", sa.String(100), nullable=True, comment="用户名快照")
+        )
+    if not c.column_exists("usage", "api_key_name"):
+        op.add_column(
+            "usage",
+            sa.Column("api_key_name", sa.String(200), nullable=True, comment="API Key 名称快照"),
+        )
+
+    # --- StatsUserDaily: CASCADE -> SET NULL, add username snapshot ---
+    _replace_fk_if_needed(
+        c,
+        "stats_user_daily_user_id_fkey",
+        "stats_user_daily",
+        "users",
+        ["user_id"],
+        ["id"],
+        "SET NULL",
+    )
+    op.alter_column("stats_user_daily", "user_id", existing_type=sa.String(36), nullable=True)
+    if not c.column_exists("stats_user_daily", "username"):
+        op.add_column(
+            "stats_user_daily",
+            sa.Column(
+                "username",
+                sa.String(100),
+                nullable=True,
+                comment="用户名快照（删除用户后仍可追溯）",
+            ),
+        )
+
+    # --- StatsDailyApiKey: CASCADE -> SET NULL, add api_key_name snapshot ---
+    _replace_fk_if_needed(
+        c,
+        "stats_daily_api_key_api_key_id_fkey",
+        "stats_daily_api_key",
+        "api_keys",
+        ["api_key_id"],
+        ["id"],
+        "SET NULL",
+    )
+    op.alter_column("stats_daily_api_key", "api_key_id", existing_type=sa.String(36), nullable=True)
+    if not c.column_exists("stats_daily_api_key", "api_key_name"):
+        op.add_column(
+            "stats_daily_api_key",
+            sa.Column(
+                "api_key_name",
+                sa.String(200),
+                nullable=True,
+                comment="API Key 名称快照（删除 Key 后仍可追溯）",
+            ),
+        )
+
+
+def downgrade() -> None:
+    c = _SchemaCache()
+    c.load_columns(["stats_daily_api_key", "stats_user_daily", "usage"])
+    c.load_fk_rules(["stats_daily_api_key", "stats_user_daily"])
+
+    # --- Remove snapshot columns ---
+    if c.column_exists("stats_daily_api_key", "api_key_name"):
+        op.drop_column("stats_daily_api_key", "api_key_name")
+    if c.column_exists("stats_user_daily", "username"):
+        op.drop_column("stats_user_daily", "username")
+    if c.column_exists("usage", "api_key_name"):
+        op.drop_column("usage", "api_key_name")
+    if c.column_exists("usage", "username"):
+        op.drop_column("usage", "username")
+
+    # --- StatsDailyApiKey: SET NULL -> CASCADE ---
+    _replace_fk_if_needed(
+        c,
+        "stats_daily_api_key_api_key_id_fkey",
+        "stats_daily_api_key",
+        "api_keys",
+        ["api_key_id"],
+        ["id"],
+        "CASCADE",
+    )
+    op.alter_column(
+        "stats_daily_api_key", "api_key_id", existing_type=sa.String(36), nullable=False
+    )
+
+    # --- StatsUserDaily: SET NULL -> CASCADE ---
+    _replace_fk_if_needed(
+        c,
+        "stats_user_daily_user_id_fkey",
+        "stats_user_daily",
+        "users",
+        ["user_id"],
+        ["id"],
+        "CASCADE",
+    )
+    op.alter_column("stats_user_daily", "user_id", existing_type=sa.String(36), nullable=False)
@@ -0,0 +1,257 @@
+"""request_candidates/video_tasks retention: SET NULL and add snapshots
+
+Revision ID: 13a4c8f6d9e0
+Revises: 45b118150a78
+Create Date: 2026-03-08 12:15:00.000000+00:00
+
+"""
+
+from __future__ import annotations
+
+import sqlalchemy as sa
+
+from alembic import op
+
+# revision identifiers, used by Alembic.
+revision = "13a4c8f6d9e0"
+down_revision = "45b118150a78"
+branch_labels = None
+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 = _SchemaCache()
+    c.load_columns(_TABLES)
+    c.load_fk_rules(_TABLES)
+
+    # --- request_candidates: add snapshot columns ---
+    if not c.column_exists("request_candidates", "username"):
+        op.add_column(
+            "request_candidates",
+            sa.Column("username", sa.String(length=100), nullable=True, comment="用户名快照"),
+        )
+    if not c.column_exists("request_candidates", "api_key_name"):
+        op.add_column(
+            "request_candidates",
+            sa.Column(
+                "api_key_name",
+                sa.String(length=200),
+                nullable=True,
+                comment="API Key 名称快照",
+            ),
+        )
+
+    # --- request_candidates: CASCADE -> SET NULL ---
+    _replace_fk_if_needed(
+        c,
+        "request_candidates_user_id_fkey",
+        "request_candidates",
+        "users",
+        ["user_id"],
+        ["id"],
+        "SET NULL",
+    )
+    _replace_fk_if_needed(
+        c,
+        "request_candidates_api_key_id_fkey",
+        "request_candidates",
+        "api_keys",
+        ["api_key_id"],
+        ["id"],
+        "SET NULL",
+    )
+
+    # --- video_tasks: add snapshot columns ---
+    if not c.column_exists("video_tasks", "username"):
+        op.add_column(
+            "video_tasks",
+            sa.Column("username", sa.String(length=100), nullable=True, comment="用户名快照"),
+        )
+    if not c.column_exists("video_tasks", "api_key_name"):
+        op.add_column(
+            "video_tasks",
+            sa.Column(
+                "api_key_name",
+                sa.String(length=200),
+                nullable=True,
+                comment="API Key 名称快照",
+            ),
+        )
+
+    # --- 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(
+        c,
+        "video_tasks_user_id_fkey",
+        "video_tasks",
+        "users",
+        ["user_id"],
+        ["id"],
+        "SET NULL",
+    )
+    _replace_fk_if_needed(
+        c,
+        "video_tasks_api_key_id_fkey",
+        "video_tasks",
+        "api_keys",
+        ["api_key_id"],
+        ["id"],
+        "SET NULL",
+    )
+
+
+def downgrade() -> None:
+    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(
+        c,
+        "video_tasks_api_key_id_fkey",
+        "video_tasks",
+        "api_keys",
+        ["api_key_id"],
+        ["id"],
+        "NO ACTION",
+    )
+    _replace_fk_if_needed(
+        c,
+        "video_tasks_user_id_fkey",
+        "video_tasks",
+        "users",
+        ["user_id"],
+        ["id"],
+        "NO ACTION",
+    )
+    op.alter_column("video_tasks", "user_id", existing_type=sa.String(length=36), nullable=False)
+    if c.column_exists("video_tasks", "api_key_name"):
+        op.drop_column("video_tasks", "api_key_name")
+    if c.column_exists("video_tasks", "username"):
+        op.drop_column("video_tasks", "username")
+
+    # --- request_candidates: SET NULL -> CASCADE ---
+    _replace_fk_if_needed(
+        c,
+        "request_candidates_api_key_id_fkey",
+        "request_candidates",
+        "api_keys",
+        ["api_key_id"],
+        ["id"],
+        "CASCADE",
+    )
+    _replace_fk_if_needed(
+        c,
+        "request_candidates_user_id_fkey",
+        "request_candidates",
+        "users",
+        ["user_id"],
+        ["id"],
+        "CASCADE",
+    )
+    if c.column_exists("request_candidates", "api_key_name"):
+        op.drop_column("request_candidates", "api_key_name")
+    if c.column_exists("request_candidates", "username"):
+        op.drop_column("request_candidates", "username")
@@ -0,0 +1,230 @@
+"""cost fields: Float -> Numeric(20,8) + provider_api_keys composite index
+
+Revision ID: 2053ab8ed764
+Revises: 13a4c8f6d9e0
+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
+
+# revision identifiers, used by Alembic.
+revision = "2053ab8ed764"
+down_revision = "13a4c8f6d9e0"
+branch_labels = None
+depends_on = None
+
+
+# (table_name, column_name, nullable, server_default)
+_COST_COLUMNS: list[tuple[str, str, bool, str | None]] = [
+    # api_keys
+    ("api_keys", "total_cost_usd", True, "0.0"),
+    # usage
+    ("usage", "input_cost_usd", True, "0.0"),
+    ("usage", "output_cost_usd", True, "0.0"),
+    ("usage", "cache_cost_usd", True, "0.0"),
+    ("usage", "cache_creation_cost_usd", True, "0.0"),
+    ("usage", "cache_read_cost_usd", True, "0.0"),
+    ("usage", "request_cost_usd", True, "0.0"),
+    ("usage", "total_cost_usd", True, "0.0"),
+    ("usage", "actual_input_cost_usd", True, "0.0"),
+    ("usage", "actual_output_cost_usd", True, "0.0"),
+    ("usage", "actual_cache_creation_cost_usd", True, "0.0"),
+    ("usage", "actual_cache_read_cost_usd", True, "0.0"),
+    ("usage", "actual_request_cost_usd", True, "0.0"),
+    ("usage", "actual_total_cost_usd", True, "0.0"),
+    ("usage", "rate_multiplier", True, "1.0"),
+    ("usage", "input_price_per_1m", True, None),
+    ("usage", "output_price_per_1m", True, None),
+    ("usage", "cache_creation_price_per_1m", True, None),
+    ("usage", "cache_read_price_per_1m", True, None),
+    ("usage", "price_per_request", True, None),
+    # providers
+    ("providers", "monthly_quota_usd", True, None),
+    ("providers", "monthly_used_usd", True, "0.0"),
+    # global_models
+    ("global_models", "default_price_per_request", True, None),
+    # models
+    ("models", "price_per_request", True, None),
+    # stats_hourly
+    ("stats_hourly", "total_cost", False, "0.0"),
+    ("stats_hourly", "actual_total_cost", False, "0.0"),
+    # stats_hourly_user
+    ("stats_hourly_user", "total_cost", False, "0.0"),
+    # stats_hourly_model
+    ("stats_hourly_model", "total_cost", False, "0.0"),
+    # stats_hourly_provider
+    ("stats_hourly_provider", "total_cost", False, "0.0"),
+    # stats_daily
+    ("stats_daily", "total_cost", False, "0.0"),
+    ("stats_daily", "actual_total_cost", False, "0.0"),
+    ("stats_daily", "input_cost", False, "0.0"),
+    ("stats_daily", "output_cost", False, "0.0"),
+    ("stats_daily", "cache_creation_cost", False, "0.0"),
+    ("stats_daily", "cache_read_cost", False, "0.0"),
+    # stats_daily_model
+    ("stats_daily_model", "total_cost", False, "0.0"),
+    # stats_daily_provider
+    ("stats_daily_provider", "total_cost", False, "0.0"),
+    # stats_daily_api_key
+    ("stats_daily_api_key", "total_cost", False, "0.0"),
+    # stats_summary
+    ("stats_summary", "all_time_cost", False, "0.0"),
+    ("stats_summary", "all_time_actual_cost", False, "0.0"),
+    # stats_user_daily
+    ("stats_user_daily", "total_cost", False, "0.0"),
+]
+
+_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 = _SchemaCache()
+    c.load_columns(_ALL_TABLES)
+
+    # -- 1. cost fields: Float -> Numeric  (batched per table)
+    cols_to_convert = [
+        (t, col, n, d)
+        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)
+
+    # -- 2. provider_api_keys composite index
+    if not _index_exists("idx_provider_api_keys_provider_active"):
+        op.create_index(
+            "idx_provider_api_keys_provider_active",
+            "provider_api_keys",
+            ["provider_id", "is_active"],
+        )
+
+
+def downgrade() -> None:
+    # -- 2. drop composite index
+    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 = _SchemaCache()
+    c.load_columns(_ALL_TABLES)
+
+    cols_to_revert = [
+        (t, col, n, d)
+        for t, col, n, d in _COST_COLUMNS
+        if c.column_exists(t, col) and c.is_numeric(t, col)
+    ]
+    _batch_alter_type(
+        c,
+        cols_to_revert,
+        cast_suffix="double precision",
+        type_fn=lambda _col: "DOUBLE PRECISION",
+    )
@@ -0,0 +1,127 @@
+"""video_tasks.key_id: add ondelete SET NULL
+
+Revision ID: d7649c1f8e21
+Revises: 2053ab8ed764
+Create Date: 2026-03-09 01:00:00.000000+00:00
+
+"""
+
+from __future__ import annotations
+
+import sqlalchemy as sa
+
+from alembic import op
+
+# revision identifiers, used by Alembic.
+revision = "d7649c1f8e21"
+down_revision = "2053ab8ed764"
+branch_labels = None
+depends_on = None
+
+_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 = _SchemaCache()
+    c.load_fk_rules([_TABLE])
+    _replace_fk_if_needed(
+        c,
+        _FK_NAME,
+        _TABLE,
+        "provider_api_keys",
+        ["key_id"],
+        ["id"],
+        "SET NULL",
+    )
+
+
+def downgrade() -> None:
+    c = _SchemaCache()
+    c.load_fk_rules([_TABLE])
+    _replace_fk_if_needed(
+        c,
+        _FK_NAME,
+        _TABLE,
+        "provider_api_keys",
+        ["key_id"],
+        ["id"],
+        "NO ACTION",
+    )
@@ -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,
+            )
@@ -15,7 +15,7 @@
 ### 用户系统
 - **users**: 用户账户管理
 - **api_keys**: API 密钥管理
- **user_quotas**: 用户配额管理
+- **wallets**: 统一钱包账户（充值余额/赠款余额/无限制模式）
 - **user_preferences**: 用户偏好设置

 ### Provider 三层架构
@@ -2,46 +2,132 @@
 # 智能部署脚本 - 自动检测依赖/代码/迁移变化
 #
 # 用法:
-#   部署/更新:    ./deploy.sh  (自动检测所有变化)
-#   强制重建:     ./deploy.sh --rebuild-base
-#   强制全部重建: ./deploy.sh --force
+#   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

-set -e
+set -euo pipefail
 cd "$(dirname "$0")"

 # 兼容 docker-compose 和 docker compose
 if command -v docker-compose &> /dev/null; then
    DC="docker-compose -f docker-compose.build.yml"
+    USE_LEGACY_COMPOSE=true
 else
    DC="docker compose -f docker-compose.build.yml"
+    USE_LEGACY_COMPOSE=false
 fi

+compose_up() {
+    if [ "$USE_LEGACY_COMPOSE" = true ]; then
+        $DC up -d --no-build "$@"
+    else
+        $DC up -d --no-build --pull never "$@"
+    fi
+}
+
 # 缓存文件
 HASH_FILE=".deps-hash"
 CODE_HASH_FILE=".code-hash"
 MIGRATION_HASH_FILE=".migration-hash"

-# 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。
-pyproject_deps_fingerprint() {
-    python3 - <<'PY'
-import json, pathlib, tomllib
+# Hub release 配置
+GITHUB_REPO="fawney19/Aether"
+HUB_TAG_STATE_FILE=".hub-tag"

-data = tomllib.loads(pathlib.Path("pyproject.toml").read_text("utf-8"))
-project = data.get("project") or {}
-build = data.get("build-system") or {}
+usage() {
+    cat <<'EOF'
+Usage: ./deploy.sh [options]

-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 []),
+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
+EOF
 }

-print(json.dumps(fingerprint, sort_keys=True, separators=(",", ":")))
-PY
+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
+    case "$1" in
+        --hub-tag)
+            if [ $# -lt 2 ]; then
+                echo "❌ --hub-tag 需要一个值，例如 hub-v0.1.0"
+                exit 1
+            fi
+            HUB_TAG="$2"
+            shift 2
+            ;;
+        --update-hub)
+            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
+            ;;
+        --force|-f)
+            FORCE_REBUILD_ALL=true
+            shift
+            ;;
+        -h|--help)
+            usage
+            exit 0
+            ;;
+        *)
+            echo "❌ 未知参数: $1"
+            usage
+            exit 1
+            ;;
+    esac
+done
+
+if [ -n "$HUB_TAG" ]; then
+    case "$HUB_TAG" in
+        hub-v*) ;;
+        *) echo "❌ --hub-tag 格式应为 hub-vX.Y.Z，例如 hub-v0.1.0"; exit 1 ;;
+    esac
+fi
+
+# 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() {
+    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）
@@ -63,6 +149,66 @@ calc_code_hash() {
    } | md5sum | cut -d' ' -f1
 }

+# 获取最新 hub release tag
+# 支持 GITHUB_TOKEN 环境变量以避免未认证 API 限流（60 次/小时 -> 5000 次/小时）
+get_latest_hub_tag() {
+    local auth_args=()
+    if [ -n "${GITHUB_TOKEN:-}" ]; then
+        auth_args=(-H "Authorization: token ${GITHUB_TOKEN}")
+    fi
+    curl -sL "${auth_args[@]}" "https://api.github.com/repos/$GITHUB_REPO/releases" | \
+        python3 -c "
+import json, sys
+releases = json.load(sys.stdin)
+for r in releases:
+    tag = r.get('tag_name', '')
+    if tag.startswith('hub-v') and not r.get('draft') and not r.get('prerelease'):
+        print(tag)
+        break
+" 2>/dev/null
+}
+
+# 解析当前应使用的 Hub release tag（优先使用指定值，否则拉取最新）
+resolve_hub_tag() {
+    local requested_tag="${1:-}"
+    local latest_tag
+
+    if [ -n "$requested_tag" ]; then
+        echo "$requested_tag"
+        return 0
+    fi
+
+    latest_tag="$(get_latest_hub_tag || true)"
+    if [ -n "$latest_tag" ]; then
+        echo "$latest_tag"
+        return 0
+    fi
+
+    if [ -f "$HUB_TAG_STATE_FILE" ]; then
+        echo "⚠️ 无法查询最新 Hub 版本，回退使用本地记录: $(cat "$HUB_TAG_STATE_FILE")" >&2
+        cat "$HUB_TAG_STATE_FILE"
+        return 0
+    fi
+
+    echo "❌ 无法获取 Hub Release tag，请检查网络或手动指定 --hub-tag" >&2
+    exit 1
+}
+
+# 确保本次构建的 Hub tag 已解析（默认追踪最新 release，也可通过 --hub-tag 固定版本）
+ensure_hub_tag() {
+    local requested_tag="${1:-}"
+    RESOLVED_HUB_TAG="$(resolve_hub_tag "$requested_tag")"
+
+    if [ -f "$HUB_TAG_STATE_FILE" ] && [ "$(cat "$HUB_TAG_STATE_FILE")" = "$RESOLVED_HUB_TAG" ]; then
+        echo ">>> Hub 版本未变化: $RESOLVED_HUB_TAG"
+        return 1
+    fi
+
+    echo "$RESOLVED_HUB_TAG" > "$HUB_TAG_STATE_FILE"
+    echo ">>> 使用 Hub 版本: $RESOLVED_HUB_TAG"
+    return 0
+}
+
 # 计算迁移文件的哈希值
 calc_migration_hash() {
    find alembic/versions -name "*.py" -type f 2>/dev/null | sort | xargs cat 2>/dev/null | md5sum | cut -d' ' -f1
@@ -92,6 +238,8 @@ check_code_changed() {
    return 0
 }

+
+
 # 检查迁移是否变化
 check_migration_changed() {
    local current_hash=$(calc_migration_hash)
@@ -112,10 +260,11 @@ save_migration_hash() { calc_migration_hash > "$MIGRATION_HASH_FILE"; }
 # 构建基础镜像
 build_base() {
    echo ">>> Building base image (dependencies)..."
-    docker build -f Dockerfile.base.local -t aether-base:latest .
+    docker build --pull=false -f Dockerfile.base.local -t aether-base:latest .
    save_deps_hash
 }

+
 # 生成版本文件
 generate_version_file() {
    # 从 git 获取版本号
@@ -137,8 +286,27 @@ EOF
 # 构建应用镜像
 build_app() {
    echo ">>> Building app image (code only)..."
+    if [ -z "${RESOLVED_HUB_TAG:-}" ]; then
+        echo ">>> RESOLVED_HUB_TAG 为空，无法构建 app 镜像"
+        exit 1
+    fi
+    echo ">>> Build args: HUB_TAG=$RESOLVED_HUB_TAG"
    generate_version_file
-    docker build -f Dockerfile.app.local -t aether-app:latest .
+    local token_args=()
+    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
 }

@@ -186,11 +354,15 @@ print('Old version cleared')
 }

 # 强制全部重建
-if [ "$1" = "--force" ] || [ "$1" = "-f" ]; then
+if [ "$FORCE_REBUILD_ALL" = true ]; then
    echo ">>> Force rebuilding everything..."
+    if [ "$FORCE_UPDATE_HUB" = true ]; then
+        rm -f "$HUB_TAG_STATE_FILE"
+    fi
+    ensure_hub_tag "$HUB_TAG" || true
    build_base
    build_app
-    $DC up -d --force-recreate
+    compose_up --force-recreate
    sleep 3
    run_migration
    docker image prune -f
@@ -200,19 +372,25 @@ if [ "$1" = "--force" ] || [ "$1" = "-f" ]; then
 fi

 # 强制重建基础镜像
-if [ "$1" = "--rebuild-base" ] || [ "$1" = "-r" ]; then
+if [ "$REBUILD_BASE_ONLY" = true ]; then
    build_base
    echo ">>> Base image rebuilt. Run ./deploy.sh to deploy."
    exit 0
 fi

-# 拉取最新代码
-echo ">>> Pulling latest code..."
-git pull
+# 更新 Hub 版本标记
+if [ "$FORCE_UPDATE_HUB" = true ]; then
+    rm -f "$HUB_TAG_STATE_FILE"
+    ensure_hub_tag "$HUB_TAG" || true
+    echo ">>> Hub tag updated: $RESOLVED_HUB_TAG"
+    echo ">>> Run ./deploy.sh to build app image with the new Hub release."
+    exit 0
+fi

 # 标记是否需要重启
 NEED_RESTART=false
 BASE_REBUILT=false
+HUB_UPDATED=false

 # 检查基础镜像是否存在，或依赖是否变化
 if ! docker image inspect aether-base:latest >/dev/null 2>&1; then
@@ -229,6 +407,14 @@ else
    echo ">>> Dependencies unchanged."
 fi

+# 解析/检查 Hub 版本（构建时由 Dockerfile 从 GitHub Release 下载）
+if ensure_hub_tag "$HUB_TAG"; then
+    HUB_UPDATED=true
+    NEED_RESTART=true
+else
+    echo ">>> Hub version unchanged."
+fi
+
 # 检查代码或迁移是否变化，或者 base 重建了（app 依赖 base）
 # 注意：迁移文件打包在镜像中，所以迁移变化也需要重建 app 镜像
 MIGRATION_CHANGED=false
@@ -244,6 +430,10 @@ elif [ "$BASE_REBUILT" = true ]; then
    echo ">>> Base image rebuilt, rebuilding app image..."
    build_app
    NEED_RESTART=true
+elif [ "$HUB_UPDATED" = true ]; then
+    echo ">>> Hub version updated, rebuilding app image..."
+    build_app
+    NEED_RESTART=true
 elif check_code_changed; then
    echo ">>> Code changed, rebuilding app image..."
    build_app
@@ -265,10 +455,10 @@ fi
 # 有变化时重启，或容器未运行时启动
 if [ "$NEED_RESTART" = true ]; then
    echo ">>> Restarting services..."
-    $DC up -d
+    compose_up
 elif [ "$CONTAINERS_RUNNING" = false ]; then
    echo ">>> Containers not running, starting services..."
-    $DC up -d
+    compose_up
 else
    echo ">>> No changes detected, skipping restart."
 fi
@@ -11,10 +11,16 @@ set +a
 export DATABASE_URL="postgresql://${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
@@ -1,7 +1,10 @@
 # Aether 部署配置 - 本地构建
 # 使用方法:
 #   首次构建 base: docker build -f Dockerfile.base -t aether-base:latest .
+#   Hub 二进制:   cd aether-hub && ./build.sh（默认 binary 模式）供发布；deploy.sh 构建 app 时会自动下载
+#   Hub 镜像发布: cd aether-hub && ./build.sh --image --tag <tag> --push（可选）
 #   启动服务: docker compose -f docker-compose.build.yml up -d --build
+#   或使用:    ./deploy.sh

 services:
  postgres:
@@ -26,9 +29,7 @@ services:
  redis:
    image: redis:7-alpine
    container_name: aether-redis
-    command: redis-server --appendonly yes --requirepass ${REDIS_PASSWORD}
-    volumes:
-      - redis_data:/data
+    command: redis-server --appendonly no --save "" --requirepass ${REDIS_PASSWORD}
    ports:
      - "${REDIS_PORT:-6379}:6379"
    healthcheck:
@@ -42,6 +43,8 @@ services:
    build:
      context: .
      dockerfile: Dockerfile.app.local
+      args:
+        GITHUB_TOKEN: ${GITHUB_TOKEN:-}
    image: aether-app:latest
    container_name: aether-app
    env_file:
@@ -51,7 +54,7 @@ services:
      DATABASE_URL: postgresql://postgres:${DB_PASSWORD}@postgres:5432/aether
      REDIS_URL: redis://:${REDIS_PASSWORD}@redis:6379/0
      # Supervisor 需要的变量
-      GUNICORN_WORKERS: ${GUNICORN_WORKERS:-4}
+      GUNICORN_WORKERS: ${GUNICORN_WORKERS:-2}
      # 容器级别设置
      TZ: Asia/Shanghai
    depends_on:
@@ -67,4 +70,3 @@ services:

 volumes:
  postgres_data:
-  redis_data:
@@ -22,9 +22,7 @@ services:
  redis:
    image: redis:7-alpine
    container_name: aether-redis
-    command: redis-server --appendonly yes --requirepass ${REDIS_PASSWORD}
-    volumes:
-      - redis_data:/data
+    command: redis-server --appendonly no --save "" --requirepass ${REDIS_PASSWORD}
    healthcheck:
      test: [ "CMD", "redis-cli", "--raw", "incr", "ping" ]
      interval: 5s
@@ -33,7 +31,7 @@ services:
    restart: unless-stopped

  app:
-    image: ghcr.io/fawney19/aether:latest
+    image: ${APP_IMAGE:-ghcr.io/fawney19/aether:latest}
    container_name: aether-app
    env_file:
      - .env
@@ -42,7 +40,7 @@ services:
      DATABASE_URL: postgresql://postgres:${DB_PASSWORD}@postgres:5432/aether
      REDIS_URL: redis://:${REDIS_PASSWORD}@redis:6379/0
      # Supervisor 需要的变量
-      GUNICORN_WORKERS: ${GUNICORN_WORKERS:-4}
+      GUNICORN_WORKERS: ${GUNICORN_WORKERS:-2}
      # 容器级别设置
      TZ: Asia/Shanghai
    depends_on:
@@ -58,4 +56,5 @@ services:

 volumes:
  postgres_data:
-  redis_data:
+
+
@@ -0,0 +1,117 @@
+<svg xmlns="http://www.w3.org/2000/svg" viewBox="0 0 760 420" width="760" height="420" font-family="'Inter', system-ui, -apple-system, sans-serif">
+  <defs>
+    <filter id="shadow-sm" x="-10%" y="-10%" width="120%" height="130%">
+      <feDropShadow dx="0" dy="2" stdDeviation="2" flood-opacity="0.5"/>
+    </filter>
+    <filter id="shadow-lg" x="-10%" y="-10%" width="120%" height="130%">
+      <feDropShadow dx="0" dy="8" stdDeviation="12" flood-opacity="0.7"/>
+    </filter>
+    <clipPath id="gateway-clip">
+      <path d="M 0 0 h 760 v 420 H 0 Z M 74 169 v 54 h 12 v -54 Z" clip-rule="evenodd"/>
+    </clipPath>
+  </defs>
+
+  <!-- Background -->
+  <rect width="760" height="420" rx="24" fill="#121212"/>
+
+  <!-- === Labels placeholder, drawn at top layer === -->
+
+  <!-- Source pills -->
+  <rect x="138" y="24" width="100" height="36" rx="18" fill="#1a1a1c" stroke="#2a2a2c" filter="url(#shadow-sm)"/>
+  <circle cx="163" cy="42" r="3" fill="#cc7154"/>
+  <text x="195" y="46" font-size="11" font-weight="700" fill="#f3f4f6" text-anchor="middle">Claude</text>
+  <rect x="330" y="24" width="100" height="36" rx="18" fill="#1a1a1c" stroke="#2a2a2c" filter="url(#shadow-sm)"/>
+  <circle cx="355" cy="42" r="3" fill="#cc7154"/>
+  <text x="387" y="46" font-size="11" font-weight="700" fill="#f3f4f6" text-anchor="middle">OpenAI</text>
+  <rect x="522" y="24" width="100" height="36" rx="18" fill="#1a1a1c" stroke="#2a2a2c" filter="url(#shadow-sm)"/>
+  <circle cx="547" cy="42" r="3" fill="#cc7154"/>
+  <text x="579" y="46" font-size="11" font-weight="700" fill="#f3f4f6" text-anchor="middle">Gemini</text>
+
+  <!-- Output pills -->
+  <rect x="128" y="360" width="120" height="36" rx="18" fill="#1a1a1c" stroke="#2a2a2c" filter="url(#shadow-sm)"/>
+  <circle cx="153" cy="378" r="3" fill="#cc7154"/>
+  <text x="195" y="382" font-size="10" font-weight="700" fill="#f3f4f6" text-anchor="middle">Claude 响应</text>
+  <rect x="320" y="360" width="120" height="36" rx="18" fill="#1a1a1c" stroke="#2a2a2c" filter="url(#shadow-sm)"/>
+  <circle cx="345" cy="378" r="3" fill="#cc7154"/>
+  <text x="387" y="382" font-size="10" font-weight="700" fill="#f3f4f6" text-anchor="middle">OpenAI 响应</text>
+  <rect x="512" y="360" width="120" height="36" rx="18" fill="#1a1a1c" stroke="#2a2a2c" filter="url(#shadow-sm)"/>
+  <circle cx="537" cy="378" r="3" fill="#cc7154"/>
+  <text x="579" y="382" font-size="10" font-weight="700" fill="#f3f4f6" text-anchor="middle">Gemini 响应</text>
+
+  <!-- === Layer 2: AETHER GATEWAY box === -->
+  <rect x="80" y="78" width="600" height="240" rx="24" fill="#1a1a1c" filter="url(#shadow-lg)"/>
+  <rect x="80" y="78" width="600" height="240" rx="24" fill="none" stroke="rgba(204,113,84,0.4)" clip-path="url(#gateway-clip)"/>
+  <g transform="translate(80, 176)" font-size="7" font-weight="900" fill="#cc7154" text-anchor="middle">
+    <text x="0" y="0">A</text><text x="0" y="9">E</text><text x="0" y="18">T</text>
+    <text x="0" y="27">H</text><text x="0" y="36">E</text><text x="0" y="45">R</text>
+  </g>
+
+  <!-- Dashed box -->
+  <rect x="180" y="98" width="400" height="112" rx="16" fill="none" stroke="rgba(204,113,84,0.4)" stroke-width="1.5" stroke-dasharray="6 4"/>
+  <rect x="200" y="108" width="360" height="26" rx="12" fill="none" stroke="#2a2a2c"/>
+  <text x="380" y="125" font-size="10" font-weight="600" fill="#f3f4f6" text-anchor="middle">统一模型规范 / 协议聚合</text>
+  <rect x="200" y="140" width="85" height="26" rx="13" fill="rgba(204,113,84,0.12)"/>
+  <text x="242" y="157" font-size="10" font-weight="600" fill="#f3f4f6" text-anchor="middle">多端鉴权</text>
+  <rect x="290" y="140" width="85" height="26" rx="13" fill="rgba(204,113,84,0.12)"/>
+  <text x="332" y="157" font-size="10" font-weight="600" fill="#f3f4f6" text-anchor="middle">配额管控</text>
+  <rect x="380" y="140" width="85" height="26" rx="13" fill="rgba(204,113,84,0.12)"/>
+  <text x="422" y="157" font-size="10" font-weight="600" fill="#f3f4f6" text-anchor="middle">全局并发</text>
+  <rect x="470" y="140" width="85" height="26" rx="13" fill="rgba(204,113,84,0.12)"/>
+  <text x="512" y="157" font-size="10" font-weight="600" fill="#f3f4f6" text-anchor="middle">缓存亲和</text>
+  <rect x="200" y="172" width="360" height="28" rx="14" fill="none" stroke="#2a2a2c"/>
+  <text x="380" y="190" font-size="10" font-weight="600" fill="#f3f4f6" text-anchor="middle">智能调度 / 故障转移</text>
+
+  <!-- Engine cards -->
+  <rect x="103" y="250" width="170" height="28" rx="14" fill="rgba(204,113,84,0.12)" stroke="#2a2a2c"/>
+  <circle cx="163" cy="264" r="3" fill="#cc7154"/>
+  <text x="192" y="268" font-size="10" font-weight="700" fill="#f3f4f6" text-anchor="middle">格式转换</text>
+  <rect x="295" y="250" width="170" height="28" rx="14" fill="rgba(204,113,84,0.12)" stroke="#2a2a2c"/>
+  <circle cx="355" cy="264" r="3" fill="#cc7154"/>
+  <text x="384" y="268" font-size="10" font-weight="700" fill="#f3f4f6" text-anchor="middle">反向代理</text>
+  <rect x="487" y="250" width="170" height="28" rx="14" fill="rgba(204,113,84,0.12)" stroke="#2a2a2c"/>
+  <circle cx="547" cy="264" r="3" fill="#cc7154"/>
+  <text x="576" y="268" font-size="10" font-weight="700" fill="#f3f4f6" text-anchor="middle">原生直通</text>
+
+  <!-- === Layer 3: Track lines === -->
+  <g stroke="rgba(255,255,255,0.08)" stroke-width="1.5" fill="none" stroke-linecap="round">
+    <path id="f-in-1" d="M 188 60 L 188 66 C 188 72, 380 72, 380 78 L 380 98"/>
+    <path id="f-in-2" d="M 380 60 L 380 98"/>
+    <path id="f-in-3" d="M 572 60 L 572 66 C 572 72, 380 72, 380 78 L 380 98"/>
+    <path id="f-sL" d="M 380 210 C 380 230, 188 230, 188 250"/>
+    <path id="f-sM" d="M 380 210 L 380 250"/>
+    <path id="f-sR" d="M 380 210 C 380 230, 572 230, 572 250"/>
+    <path id="f-oL" d="M 188 278 L 188 284 C 188 302, 380 302, 380 318"/>
+    <path id="f-oM" d="M 380 278 L 380 318"/>
+    <path id="f-oR" d="M 572 278 L 572 284 C 572 302, 380 302, 380 318"/>
+    <path id="f-oL2" d="M 380 318 C 380 338, 188 338, 188 360"/>
+    <path id="f-oM2" d="M 380 318 L 380 360"/>
+    <path id="f-oR2" d="M 380 318 C 380 338, 572 338, 572 360"/>
+  </g>
+
+  <!-- Animated dots -->
+  <g>
+    <circle r="2.5" fill="#cc7154"><animateMotion dur="2.2s" repeatCount="indefinite"><mpath href="#f-in-1"/></animateMotion></circle>
+    <circle r="2.5" fill="#cc7154"><animateMotion dur="2.2s" repeatCount="indefinite"><mpath href="#f-in-3"/></animateMotion></circle>
+    <circle r="2.5" fill="#cc7154"><animateMotion dur="1.4s" repeatCount="indefinite"><mpath href="#f-sL"/></animateMotion></circle>
+    <circle r="2.5" fill="#cc7154"><animateMotion dur="1.4s" repeatCount="indefinite"><mpath href="#f-sR"/></animateMotion></circle>
+    <circle r="2.5" fill="#cc7154"><animateMotion dur="1.4s" repeatCount="indefinite"><mpath href="#f-oL"/></animateMotion></circle>
+    <circle r="2.5" fill="#cc7154"><animateMotion dur="1.4s" repeatCount="indefinite"><mpath href="#f-oR"/></animateMotion></circle>
+    <circle r="2.5" fill="#cc7154"><animateMotion dur="1.2s" repeatCount="indefinite"><mpath href="#f-oL2"/></animateMotion></circle>
+    <circle r="2.5" fill="#cc7154"><animateMotion dur="1.2s" repeatCount="indefinite"><mpath href="#f-oR2"/></animateMotion></circle>
+  </g>
+
+  <!-- === Layer 4: Junction dots === -->
+  <circle cx="380" cy="78" r="4" fill="#cc7154" stroke="#1e1e1e"/>
+  <circle cx="380" cy="210" r="4" fill="#cc7154" stroke="#1e1e1e"/>
+  <circle cx="380" cy="318" r="4" fill="#cc7154" stroke="#1e1e1e"/>
+
+  <!-- === Layer 5: SOURCES / OUTPUTS labels === -->
+  <g>
+    <circle cx="355" cy="12" r="2" fill="#6b7280"/>
+    <text x="365" y="15" font-size="9" font-weight="800" letter-spacing="0.2em" fill="#6b7280">SOURCES</text>
+  </g>
+  <g>
+    <circle cx="353" cy="346" r="2" fill="#6b7280"/>
+    <text x="363" y="349" font-size="9" font-weight="800" letter-spacing="0.2em" fill="#6b7280">OUTPUTS</text>
+  </g>
+</svg>
@@ -0,0 +1,118 @@
+<svg xmlns="http://www.w3.org/2000/svg" viewBox="0 0 760 420" width="760" height="420" font-family="'Inter', system-ui, -apple-system, sans-serif">
+  <defs>
+    <filter id="shadow-sm" x="-10%" y="-10%" width="120%" height="130%">
+      <feDropShadow dx="0" dy="2" stdDeviation="3" flood-opacity="0.04"/>
+    </filter>
+    <filter id="shadow-lg" x="-10%" y="-10%" width="120%" height="130%">
+      <feDropShadow dx="0" dy="6" stdDeviation="10" flood-opacity="0.06"/>
+    </filter>
+    <!-- Clip path: full area minus AETHER label zone -->
+    <clipPath id="gateway-clip">
+      <path d="M 0 0 h 760 v 420 H 0 Z M 74 169 v 54 h 12 v -54 Z" clip-rule="evenodd"/>
+    </clipPath>
+  </defs>
+
+  <!-- Background -->
+  <rect width="760" height="420" rx="24" fill="#fcfcfc"/>
+
+  <!-- === Labels placeholder, drawn at top layer === -->
+
+  <!-- Source pills -->
+  <rect x="138" y="24" width="100" height="36" rx="18" fill="#ffffff" stroke="#e8e8ec" filter="url(#shadow-sm)"/>
+  <circle cx="163" cy="42" r="3" fill="#cc7154"/>
+  <text x="195" y="46" font-size="11" font-weight="700" fill="#1f2937" text-anchor="middle">Claude</text>
+  <rect x="330" y="24" width="100" height="36" rx="18" fill="#ffffff" stroke="#e8e8ec" filter="url(#shadow-sm)"/>
+  <circle cx="355" cy="42" r="3" fill="#cc7154"/>
+  <text x="387" y="46" font-size="11" font-weight="700" fill="#1f2937" text-anchor="middle">OpenAI</text>
+  <rect x="522" y="24" width="100" height="36" rx="18" fill="#ffffff" stroke="#e8e8ec" filter="url(#shadow-sm)"/>
+  <circle cx="547" cy="42" r="3" fill="#cc7154"/>
+  <text x="579" y="46" font-size="11" font-weight="700" fill="#1f2937" text-anchor="middle">Gemini</text>
+
+  <!-- Output pills -->
+  <rect x="128" y="360" width="120" height="36" rx="18" fill="#ffffff" stroke="#e8e8ec" filter="url(#shadow-sm)"/>
+  <circle cx="153" cy="378" r="3" fill="#cc7154"/>
+  <text x="195" y="382" font-size="10" font-weight="700" fill="#1f2937" text-anchor="middle">Claude 响应</text>
+  <rect x="320" y="360" width="120" height="36" rx="18" fill="#ffffff" stroke="#e8e8ec" filter="url(#shadow-sm)"/>
+  <circle cx="345" cy="378" r="3" fill="#cc7154"/>
+  <text x="387" y="382" font-size="10" font-weight="700" fill="#1f2937" text-anchor="middle">OpenAI 响应</text>
+  <rect x="512" y="360" width="120" height="36" rx="18" fill="#ffffff" stroke="#e8e8ec" filter="url(#shadow-sm)"/>
+  <circle cx="537" cy="378" r="3" fill="#cc7154"/>
+  <text x="579" y="382" font-size="10" font-weight="700" fill="#1f2937" text-anchor="middle">Gemini 响应</text>
+
+  <!-- === Layer 2: AETHER GATEWAY box (z-20) === -->
+  <rect x="80" y="78" width="600" height="240" rx="24" fill="#ffffff" filter="url(#shadow-lg)"/>
+  <rect x="80" y="78" width="600" height="240" rx="24" fill="none" stroke="rgba(204,113,84,0.25)" clip-path="url(#gateway-clip)"/>
+  <g transform="translate(80, 176)" font-size="7" font-weight="900" fill="#cc7154" text-anchor="middle">
+    <text x="0" y="0">A</text><text x="0" y="9">E</text><text x="0" y="18">T</text>
+    <text x="0" y="27">H</text><text x="0" y="36">E</text><text x="0" y="45">R</text>
+  </g>
+
+  <!-- Dashed box -->
+  <rect x="180" y="98" width="400" height="112" rx="16" fill="none" stroke="rgba(204,113,84,0.25)" stroke-width="1.5" stroke-dasharray="6 4"/>
+  <rect x="200" y="108" width="360" height="26" rx="12" fill="none" stroke="#e8e8ec"/>
+  <text x="380" y="125" font-size="10" font-weight="600" fill="#1f2937" text-anchor="middle">统一模型规范 / 协议聚合</text>
+  <rect x="200" y="140" width="85" height="26" rx="13" fill="rgba(204,113,84,0.06)"/>
+  <text x="242" y="157" font-size="10" font-weight="600" fill="#1f2937" text-anchor="middle">多端鉴权</text>
+  <rect x="290" y="140" width="85" height="26" rx="13" fill="rgba(204,113,84,0.06)"/>
+  <text x="332" y="157" font-size="10" font-weight="600" fill="#1f2937" text-anchor="middle">配额管控</text>
+  <rect x="380" y="140" width="85" height="26" rx="13" fill="rgba(204,113,84,0.06)"/>
+  <text x="422" y="157" font-size="10" font-weight="600" fill="#1f2937" text-anchor="middle">全局并发</text>
+  <rect x="470" y="140" width="85" height="26" rx="13" fill="rgba(204,113,84,0.06)"/>
+  <text x="512" y="157" font-size="10" font-weight="600" fill="#1f2937" text-anchor="middle">缓存亲和</text>
+  <rect x="200" y="172" width="360" height="28" rx="14" fill="none" stroke="#e8e8ec"/>
+  <text x="380" y="190" font-size="10" font-weight="600" fill="#1f2937" text-anchor="middle">智能调度 / 故障转移</text>
+
+  <!-- Engine cards -->
+  <rect x="103" y="250" width="170" height="28" rx="14" fill="rgba(204,113,84,0.06)" stroke="#e8e8ec"/>
+  <circle cx="163" cy="264" r="3" fill="#cc7154"/>
+  <text x="192" y="268" font-size="10" font-weight="700" fill="#1f2937" text-anchor="middle">格式转换</text>
+  <rect x="295" y="250" width="170" height="28" rx="14" fill="rgba(204,113,84,0.06)" stroke="#e8e8ec"/>
+  <circle cx="355" cy="264" r="3" fill="#cc7154"/>
+  <text x="384" y="268" font-size="10" font-weight="700" fill="#1f2937" text-anchor="middle">反向代理</text>
+  <rect x="487" y="250" width="170" height="28" rx="14" fill="rgba(204,113,84,0.06)" stroke="#e8e8ec"/>
+  <circle cx="547" cy="264" r="3" fill="#cc7154"/>
+  <text x="576" y="268" font-size="10" font-weight="700" fill="#1f2937" text-anchor="middle">原生直通</text>
+
+  <!-- === Layer 3: Track lines (z-30, above Gateway fill) === -->
+  <g stroke="#e5e7eb" stroke-width="1.5" fill="none" stroke-linecap="round">
+    <path id="f-in-1" d="M 188 60 L 188 66 C 188 72, 380 72, 380 78 L 380 98"/>
+    <path id="f-in-2" d="M 380 60 L 380 98"/>
+    <path id="f-in-3" d="M 572 60 L 572 66 C 572 72, 380 72, 380 78 L 380 98"/>
+    <path id="f-sL" d="M 380 210 C 380 230, 188 230, 188 250"/>
+    <path id="f-sM" d="M 380 210 L 380 250"/>
+    <path id="f-sR" d="M 380 210 C 380 230, 572 230, 572 250"/>
+    <path id="f-oL" d="M 188 278 L 188 284 C 188 302, 380 302, 380 318"/>
+    <path id="f-oM" d="M 380 278 L 380 318"/>
+    <path id="f-oR" d="M 572 278 L 572 284 C 572 302, 380 302, 380 318"/>
+    <path id="f-oL2" d="M 380 318 C 380 338, 188 338, 188 360"/>
+    <path id="f-oM2" d="M 380 318 L 380 360"/>
+    <path id="f-oR2" d="M 380 318 C 380 338, 572 338, 572 360"/>
+  </g>
+
+  <!-- Animated dots -->
+  <g>
+    <circle r="2.5" fill="#cc7154"><animateMotion dur="2.2s" repeatCount="indefinite"><mpath href="#f-in-1"/></animateMotion></circle>
+    <circle r="2.5" fill="#cc7154"><animateMotion dur="2.2s" repeatCount="indefinite"><mpath href="#f-in-3"/></animateMotion></circle>
+    <circle r="2.5" fill="#cc7154"><animateMotion dur="1.4s" repeatCount="indefinite"><mpath href="#f-sL"/></animateMotion></circle>
+    <circle r="2.5" fill="#cc7154"><animateMotion dur="1.4s" repeatCount="indefinite"><mpath href="#f-sR"/></animateMotion></circle>
+    <circle r="2.5" fill="#cc7154"><animateMotion dur="1.4s" repeatCount="indefinite"><mpath href="#f-oL"/></animateMotion></circle>
+    <circle r="2.5" fill="#cc7154"><animateMotion dur="1.4s" repeatCount="indefinite"><mpath href="#f-oR"/></animateMotion></circle>
+    <circle r="2.5" fill="#cc7154"><animateMotion dur="1.2s" repeatCount="indefinite"><mpath href="#f-oL2"/></animateMotion></circle>
+    <circle r="2.5" fill="#cc7154"><animateMotion dur="1.2s" repeatCount="indefinite"><mpath href="#f-oR2"/></animateMotion></circle>
+  </g>
+
+  <!-- === Layer 4: Junction dots (z-40, topmost) === -->
+  <circle cx="380" cy="78" r="4" fill="#cc7154" stroke="#ffffff"/>
+  <circle cx="380" cy="210" r="4" fill="#cc7154" stroke="#ffffff"/>
+  <circle cx="380" cy="318" r="4" fill="#cc7154" stroke="#ffffff"/>
+
+  <!-- === Layer 5: SOURCES / OUTPUTS labels (topmost) === -->
+  <g>
+    <circle cx="355" cy="12" r="2" fill="#8c94a1"/>
+    <text x="365" y="15" font-size="9" font-weight="800" letter-spacing="0.2em" fill="#8c94a1">SOURCES</text>
+  </g>
+  <g>
+    <circle cx="353" cy="346" r="2" fill="#8c94a1"/>
+    <text x="363" y="349" font-size="9" font-weight="800" letter-spacing="0.2em" fill="#8c94a1">OUTPUTS</text>
+  </g>
+</svg>
--- a/Show More
+++ b/Show More
@@ -1 +1 @@
 .14
 .13