feat(llm-gateway): 新增对外 LLM API 代理服务 — 完整的监管注入、内容审查和管理后台

## 新增微服务: llm-gateway (端口 3008) 对外提供与 Anthropic/OpenAI 完全兼容的 API 接口，中间拦截实现： - API Key 认证：由我们分配 Key 给外部用户，SHA-256 哈希存储 - System Prompt 注入：在请求转发前注入监管合规内容（支持 prepend/append） - 内容审查过滤：对用户消息进行关键词/正则匹配，支持 block/warn/log 三种动作 - 用量记录：异步批量写入，跟踪 token 消耗和费用估算 - 审计日志：记录每次请求的来源 IP、过滤状态、注入状态等 - 速率限制：基于内存滑动窗口的 RPM 限制 ### 技术选型 - Fastify (非 NestJS)：纯代理场景无需 DI 容器，路由开销 ~2ms - SSE 流式管道：零缓冲直通，支持 Anthropic streaming 和 OpenAI streaming - 规则缓存：30 秒 TTL，避免每次请求查库 ### API 端点 - POST /v1/messages — Anthropic Messages API 代理（流式+非流式） - POST /v1/embeddings — OpenAI Embeddings API 代理 - POST /v1/chat/completions — OpenAI Chat Completions API 代理 - GET /health — 健康检查 ## 数据库 (5 张新表) - gateway_api_keys: 外部用户 API Key（权限、限速、预算、过期时间） - gateway_injection_rules: 监管内容注入规则（位置、匹配模型、匹配 Key） - gateway_content_rules: 内容审查规则（关键词/正则、block/warn/log） - gateway_usage_logs: Token 用量记录（按 Key、模型、提供商统计） - gateway_audit_logs: 请求审计日志（IP、过滤状态、注入状态） ## Admin 后端 (conversation-service) 4 个 NestJS 控制器，挂载在 /conversations/admin/gateway/ 下： - AdminGatewayKeysController: Key 的 CRUD + toggle - AdminGatewayInjectionRulesController: 注入规则 CRUD + toggle - AdminGatewayContentRulesController: 内容审查规则 CRUD + toggle - AdminGatewayDashboardController: 仪表盘汇总、用量查询、审计日志查询 5 个 ORM 实体文件对应 5 张数据库表。 ## Admin 前端 (admin-client) 新增 features/llm-gateway 模块，Tabs 布局包含 5 个管理面板： - API Key Tab: 创建/删除/启停 Key，创建时一次性显示完整 Key - 注入规则 Tab: 配置监管内容（前置/追加到 system prompt） - 内容审查 Tab: 配置关键词/正则过滤规则 - 用量统计 Tab: 查看 token 消耗、费用、响应时间 - 审计日志 Tab: 查看请求记录、过滤命中、注入状态菜单项: GatewayOutlined + "LLM 网关"，位于"系统总监"和"数据分析"之间。 ## 基础设施 - docker-compose.yml: 新增 llm-gateway 服务定义 - kong.yml: 新增 /v1/messages、/v1/embeddings、/v1/chat/completions 路由 - 超时设置 300 秒（LLM 长响应） - CORS 新增 X-Api-Key、anthropic-version、anthropic-beta 头 - init-db.sql: 新增 5 张 gateway 表的建表语句 ## 架构说明内部服务（conversation-service、knowledge-service、evolution-service）继续直连 API， llm-gateway 仅服务外部用户。两者通过共享 PostgreSQL 数据库关联配置。 Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
2026-02-25 22:32:25 -08:00 · 2026-02-25 22:32:25 -08:00 · 6476bd868f
parent 021afd8677
commit 6476bd868f
40 changed files with 3724 additions and 2 deletions
--- a/docker-compose.yml
+++ b/docker-compose.yml
@ -426,6 +426,40 @@ services:
    networks:
      - iconsulting-network

+  #=============================================================================
+  # LLM Gateway - 对外 API 代理服务
+  #=============================================================================
+
+  llm-gateway:
+    build:
+      context: .
+      dockerfile: packages/services/llm-gateway/Dockerfile
+    container_name: iconsulting-llm-gateway
+    restart: unless-stopped
+    depends_on:
+      postgres:
+        condition: service_healthy
+    environment:
+      NODE_ENV: production
+      PORT: 3008
+      DATABASE_URL: postgresql://${POSTGRES_USER:-postgres}:${POSTGRES_PASSWORD:-postgres}@postgres:5432/${POSTGRES_DB:-iconsulting}
+      ANTHROPIC_API_KEY: ${ANTHROPIC_API_KEY}
+      ANTHROPIC_UPSTREAM_URL: ${ANTHROPIC_BASE_URL:-https://api.anthropic.com}
+      OPENAI_API_KEY: ${OPENAI_API_KEY}
+      OPENAI_UPSTREAM_URL: ${OPENAI_BASE_URL:-https://api.openai.com}
+      RULES_CACHE_TTL_MS: 30000
+      LOG_LEVEL: info
+    ports:
+      - "3008:3008"
+    healthcheck:
+      test: ["CMD", "wget", "--no-verbose", "--tries=1", "--spider", "http://127.0.0.1:3008/health"]
+      interval: 30s
+      timeout: 10s
+      retries: 3
+      start_period: 10s
+    networks:
+      - iconsulting-network
+
  #=============================================================================
  # 前端 Nginx
  #=============================================================================
--- a/kong/kong.yml
+++ b/kong/kong.yml
@ -169,6 +169,28 @@ services:
          - DELETE
          - OPTIONS

+  #-----------------------------------------------------------------------------
+  # LLM Gateway - 对外 API 代理服务
+  # 注意: 需要长超时以支持 LLM 流式响应
+  #-----------------------------------------------------------------------------
+  - name: llm-gateway
+    url: http://llm-gateway:3008
+    connect_timeout: 60000
+    write_timeout: 300000
+    read_timeout: 300000
+    retries: 2
+    routes:
+      - name: llm-gateway-routes
+        paths:
+          - /v1/messages
+          - /v1/embeddings
+          - /v1/chat/completions
+        strip_path: false
+        preserve_host: true
+        methods:
+          - POST
+          - OPTIONS
+
 #===============================================================================
 # 全局插件配置
 #===============================================================================
@ -198,6 +220,9 @@ plugins:
        - Authorization
        - X-User-Id
        - X-Request-Id
+        - X-Api-Key
+        - anthropic-version
+        - anthropic-beta
      exposed_headers:
        - X-Request-Id
      credentials: true
--- a/packages/admin-client/src/App.tsx
+++ b/packages/admin-client/src/App.tsx
@ -15,6 +15,7 @@ import { ObservabilityPage } from './features/observability';
 import { AssessmentConfigPage } from './features/assessment-config';
 import { CollectionConfigPage } from './features/collection-config';
 import { SupervisorPage } from './features/supervisor';
+import { LLMGatewayPage } from './features/llm-gateway';

 function App() {
  return (
@ -44,6 +45,7 @@ function App() {
        <Route path="tenants" element={<TenantsPage />} />
        <Route path="mcp" element={<McpPage />} />
        <Route path="supervisor" element={<SupervisorPage />} />
+        <Route path="llm-gateway" element={<LLMGatewayPage />} />
        <Route path="observability" element={<ObservabilityPage />} />
        <Route path="settings" element={<SettingsPage />} />
      </Route>
--- a/packages/admin-client/src/features/llm-gateway/application/useLLMGateway.ts
+++ b/packages/admin-client/src/features/llm-gateway/application/useLLMGateway.ts
@ -0,0 +1,209 @@
+import { useQuery, useMutation, useQueryClient } from '@tanstack/react-query';
+import { message } from 'antd';
+import { gatewayApi } from '../infrastructure/llm-gateway.api';
+
+const KEYS = {
+  dashboard: 'gateway-dashboard',
+  keys: 'gateway-keys',
+  injectionRules: 'gateway-injection-rules',
+  contentRules: 'gateway-content-rules',
+  usage: 'gateway-usage',
+  auditLogs: 'gateway-audit-logs',
+};
+
+// ─── Dashboard ───
+
+export function useDashboard() {
+  return useQuery({
+    queryKey: [KEYS.dashboard],
+    queryFn: () => gatewayApi.getDashboard(),
+    refetchInterval: 30_000,
+  });
+}
+
+// ─── API Keys ───
+
+export function useApiKeys() {
+  return useQuery({
+    queryKey: [KEYS.keys],
+    queryFn: () => gatewayApi.listKeys(),
+  });
+}
+
+export function useCreateKey() {
+  const qc = useQueryClient();
+  return useMutation({
+    mutationFn: gatewayApi.createKey,
+    onSuccess: () => {
+      qc.invalidateQueries({ queryKey: [KEYS.keys] });
+      qc.invalidateQueries({ queryKey: [KEYS.dashboard] });
+      message.success('API Key 创建成功');
+    },
+    onError: () => message.error('创建失败'),
+  });
+}
+
+export function useUpdateKey() {
+  const qc = useQueryClient();
+  return useMutation({
+    mutationFn: ({ id, dto }: { id: string; dto: any }) => gatewayApi.updateKey(id, dto),
+    onSuccess: () => {
+      qc.invalidateQueries({ queryKey: [KEYS.keys] });
+      message.success('更新成功');
+    },
+    onError: () => message.error('更新失败'),
+  });
+}
+
+export function useDeleteKey() {
+  const qc = useQueryClient();
+  return useMutation({
+    mutationFn: gatewayApi.deleteKey,
+    onSuccess: () => {
+      qc.invalidateQueries({ queryKey: [KEYS.keys] });
+      qc.invalidateQueries({ queryKey: [KEYS.dashboard] });
+      message.success('删除成功');
+    },
+    onError: () => message.error('删除失败'),
+  });
+}
+
+export function useToggleKey() {
+  const qc = useQueryClient();
+  return useMutation({
+    mutationFn: gatewayApi.toggleKey,
+    onSuccess: () => {
+      qc.invalidateQueries({ queryKey: [KEYS.keys] });
+      qc.invalidateQueries({ queryKey: [KEYS.dashboard] });
+    },
+    onError: () => message.error('操作失败'),
+  });
+}
+
+// ─── Injection Rules ───
+
+export function useInjectionRules() {
+  return useQuery({
+    queryKey: [KEYS.injectionRules],
+    queryFn: () => gatewayApi.listInjectionRules(),
+  });
+}
+
+export function useCreateInjectionRule() {
+  const qc = useQueryClient();
+  return useMutation({
+    mutationFn: gatewayApi.createInjectionRule,
+    onSuccess: () => {
+      qc.invalidateQueries({ queryKey: [KEYS.injectionRules] });
+      message.success('注入规则创建成功');
+    },
+    onError: () => message.error('创建失败'),
+  });
+}
+
+export function useUpdateInjectionRule() {
+  const qc = useQueryClient();
+  return useMutation({
+    mutationFn: ({ id, dto }: { id: string; dto: any }) => gatewayApi.updateInjectionRule(id, dto),
+    onSuccess: () => {
+      qc.invalidateQueries({ queryKey: [KEYS.injectionRules] });
+      message.success('更新成功');
+    },
+    onError: () => message.error('更新失败'),
+  });
+}
+
+export function useDeleteInjectionRule() {
+  const qc = useQueryClient();
+  return useMutation({
+    mutationFn: gatewayApi.deleteInjectionRule,
+    onSuccess: () => {
+      qc.invalidateQueries({ queryKey: [KEYS.injectionRules] });
+      message.success('删除成功');
+    },
+    onError: () => message.error('删除失败'),
+  });
+}
+
+export function useToggleInjectionRule() {
+  const qc = useQueryClient();
+  return useMutation({
+    mutationFn: gatewayApi.toggleInjectionRule,
+    onSuccess: () => {
+      qc.invalidateQueries({ queryKey: [KEYS.injectionRules] });
+    },
+    onError: () => message.error('操作失败'),
+  });
+}
+
+// ─── Content Rules ───
+
+export function useContentRules() {
+  return useQuery({
+    queryKey: [KEYS.contentRules],
+    queryFn: () => gatewayApi.listContentRules(),
+  });
+}
+
+export function useCreateContentRule() {
+  const qc = useQueryClient();
+  return useMutation({
+    mutationFn: gatewayApi.createContentRule,
+    onSuccess: () => {
+      qc.invalidateQueries({ queryKey: [KEYS.contentRules] });
+      message.success('审查规则创建成功');
+    },
+    onError: () => message.error('创建失败'),
+  });
+}
+
+export function useUpdateContentRule() {
+  const qc = useQueryClient();
+  return useMutation({
+    mutationFn: ({ id, dto }: { id: string; dto: any }) => gatewayApi.updateContentRule(id, dto),
+    onSuccess: () => {
+      qc.invalidateQueries({ queryKey: [KEYS.contentRules] });
+      message.success('更新成功');
+    },
+    onError: () => message.error('更新失败'),
+  });
+}
+
+export function useDeleteContentRule() {
+  const qc = useQueryClient();
+  return useMutation({
+    mutationFn: gatewayApi.deleteContentRule,
+    onSuccess: () => {
+      qc.invalidateQueries({ queryKey: [KEYS.contentRules] });
+      message.success('删除成功');
+    },
+    onError: () => message.error('删除失败'),
+  });
+}
+
+export function useToggleContentRule() {
+  const qc = useQueryClient();
+  return useMutation({
+    mutationFn: gatewayApi.toggleContentRule,
+    onSuccess: () => {
+      qc.invalidateQueries({ queryKey: [KEYS.contentRules] });
+    },
+    onError: () => message.error('操作失败'),
+  });
+}
+
+// ─── Usage & Audit ───
+
+export function useUsageLogs(params?: { keyId?: string; startDate?: string; endDate?: string }) {
+  return useQuery({
+    queryKey: [KEYS.usage, params],
+    queryFn: () => gatewayApi.getUsage(params),
+  });
+}
+
+export function useAuditLogs(params?: { keyId?: string; filtered?: boolean; startDate?: string; endDate?: string }) {
+  return useQuery({
+    queryKey: [KEYS.auditLogs, params],
+    queryFn: () => gatewayApi.getAuditLogs(params),
+  });
+}
--- a/packages/admin-client/src/features/llm-gateway/index.ts
+++ b/packages/admin-client/src/features/llm-gateway/index.ts
@ -0,0 +1 @@
+export { LLMGatewayPage } from './presentation/pages/LLMGatewayPage';
--- a/packages/admin-client/src/features/llm-gateway/infrastructure/llm-gateway.api.ts
+++ b/packages/admin-client/src/features/llm-gateway/infrastructure/llm-gateway.api.ts
@ -0,0 +1,176 @@
+import api from '../../../shared/utils/api';
+
+const BASE = '/conversations/admin/gateway';
+
+// ─── Types ───
+
+export interface GatewayApiKey {
+  id: string;
+  tenantId: string | null;
+  keyHash: string;
+  keyPrefix: string;
+  name: string;
+  owner: string;
+  permissions: {
+    allowedModels: string[];
+    allowStreaming: boolean;
+    allowTools: boolean;
+  };
+  rateLimitRpm: number;
+  rateLimitTpd: number;
+  monthlyBudget: number | null;
+  enabled: boolean;
+  expiresAt: string | null;
+  lastUsedAt: string | null;
+  createdBy: string | null;
+  createdAt: string;
+  rawKey?: string; // only on creation
+}
+
+export interface InjectionRule {
+  id: string;
+  name: string;
+  description: string | null;
+  position: 'prepend' | 'append';
+  content: string;
+  matchModels: string[];
+  matchKeyIds: string[] | null;
+  priority: number;
+  enabled: boolean;
+  createdAt: string;
+  updatedAt: string;
+}
+
+export interface ContentRule {
+  id: string;
+  name: string;
+  type: 'keyword' | 'regex';
+  pattern: string;
+  action: 'block' | 'warn' | 'log';
+  rejectMessage: string | null;
+  priority: number;
+  enabled: boolean;
+  createdAt: string;
+  updatedAt: string;
+}
+
+export interface UsageLog {
+  id: string;
+  apiKeyId: string;
+  model: string;
+  provider: string;
+  inputTokens: number;
+  outputTokens: number;
+  totalTokens: number;
+  costUsd: number | null;
+  durationMs: number;
+  statusCode: number;
+  createdAt: string;
+}
+
+export interface AuditLog {
+  id: string;
+  apiKeyId: string;
+  requestMethod: string;
+  requestPath: string;
+  requestModel: string | null;
+  requestIp: string;
+  contentFiltered: boolean;
+  filterRuleId: string | null;
+  injectionApplied: boolean;
+  responseStatus: number;
+  durationMs: number;
+  createdAt: string;
+}
+
+export interface DashboardData {
+  keys: { total: number; active: number };
+  today: { requests: number; tokens: number; cost: number; filtered: number };
+  month: { requests: number; tokens: number; cost: number };
+}
+
+// ─── API ───
+
+export const gatewayApi = {
+  // Dashboard
+  getDashboard: async (): Promise<DashboardData> => {
+    const res = await api.get(`${BASE}/dashboard`);
+    return res.data.data;
+  },
+
+  // Keys
+  listKeys: async () => {
+    const res = await api.get(`${BASE}/keys`);
+    return res.data.data;
+  },
+  createKey: async (dto: { name: string; owner?: string; rateLimitRpm?: number; rateLimitTpd?: number; monthlyBudget?: number; expiresAt?: string }) => {
+    const res = await api.post(`${BASE}/keys`, dto);
+    return res.data.data;
+  },
+  updateKey: async (id: string, dto: Partial<GatewayApiKey>) => {
+    const res = await api.put(`${BASE}/keys/${id}`, dto);
+    return res.data.data;
+  },
+  deleteKey: async (id: string) => {
+    const res = await api.delete(`${BASE}/keys/${id}`);
+    return res.data;
+  },
+  toggleKey: async (id: string) => {
+    const res = await api.post(`${BASE}/keys/${id}/toggle`);
+    return res.data.data;
+  },
+
+  // Injection Rules
+  listInjectionRules: async () => {
+    const res = await api.get(`${BASE}/injection-rules`);
+    return res.data.data;
+  },
+  createInjectionRule: async (dto: { name: string; position?: string; content: string; description?: string; matchModels?: string[]; priority?: number }) => {
+    const res = await api.post(`${BASE}/injection-rules`, dto);
+    return res.data.data;
+  },
+  updateInjectionRule: async (id: string, dto: Partial<InjectionRule>) => {
+    const res = await api.put(`${BASE}/injection-rules/${id}`, dto);
+    return res.data.data;
+  },
+  deleteInjectionRule: async (id: string) => {
+    const res = await api.delete(`${BASE}/injection-rules/${id}`);
+    return res.data;
+  },
+  toggleInjectionRule: async (id: string) => {
+    const res = await api.post(`${BASE}/injection-rules/${id}/toggle`);
+    return res.data.data;
+  },
+
+  // Content Rules
+  listContentRules: async () => {
+    const res = await api.get(`${BASE}/content-rules`);
+    return res.data.data;
+  },
+  createContentRule: async (dto: { name: string; type?: string; pattern: string; action?: string; rejectMessage?: string; priority?: number }) => {
+    const res = await api.post(`${BASE}/content-rules`, dto);
+    return res.data.data;
+  },
+  updateContentRule: async (id: string, dto: Partial<ContentRule>) => {
+    const res = await api.put(`${BASE}/content-rules/${id}`, dto);
+    return res.data.data;
+  },
+  deleteContentRule: async (id: string) => {
+    const res = await api.delete(`${BASE}/content-rules/${id}`);
+    return res.data;
+  },
+  toggleContentRule: async (id: string) => {
+    const res = await api.post(`${BASE}/content-rules/${id}/toggle`);
+    return res.data.data;
+  },
+
+  // Usage & Audit
+  getUsage: async (params?: { keyId?: string; startDate?: string; endDate?: string; limit?: number }) => {
+    const res = await api.get(`${BASE}/usage`, { params });
+    return res.data.data;
+  },
+  getAuditLogs: async (params?: { keyId?: string; filtered?: boolean; startDate?: string; endDate?: string; limit?: number }) => {
+    const res = await api.get(`${BASE}/audit-logs`, { params });
+    return res.data.data;
+  },
+};
--- a/packages/admin-client/src/features/llm-gateway/presentation/components/ApiKeysTab.tsx
+++ b/packages/admin-client/src/features/llm-gateway/presentation/components/ApiKeysTab.tsx
@ -0,0 +1,137 @@
+import { useState } from 'react';
+import { Table, Button, Modal, Form, Input, InputNumber, Switch, Space, Tag, Popconfirm, Typography, message as antMsg } from 'antd';
+import { PlusOutlined, CopyOutlined } from '@ant-design/icons';
+import { useApiKeys, useCreateKey, useDeleteKey, useToggleKey } from '../../application/useLLMGateway';
+import type { GatewayApiKey } from '../../infrastructure/llm-gateway.api';
+
+const { Text } = Typography;
+
+export function ApiKeysTab() {
+  const { data, isLoading } = useApiKeys();
+  const createMutation = useCreateKey();
+  const deleteMutation = useDeleteKey();
+  const toggleMutation = useToggleKey();
+  const [createOpen, setCreateOpen] = useState(false);
+  const [newKeyModal, setNewKeyModal] = useState<{ rawKey: string } | null>(null);
+  const [form] = Form.useForm();
+
+  const handleCreate = async () => {
+    const values = await form.validateFields();
+    const result = await createMutation.mutateAsync(values);
+    setCreateOpen(false);
+    form.resetFields();
+    if (result.rawKey) {
+      setNewKeyModal({ rawKey: result.rawKey });
+    }
+  };
+
+  const columns = [
+    {
+      title: '名称',
+      dataIndex: 'name',
+      key: 'name',
+    },
+    {
+      title: 'Key 前缀',
+      dataIndex: 'keyPrefix',
+      key: 'keyPrefix',
+      render: (v: string) => <Text code>{v}...</Text>,
+    },
+    {
+      title: '所属',
+      dataIndex: 'owner',
+      key: 'owner',
+    },
+    {
+      title: '限速 (RPM)',
+      dataIndex: 'rateLimitRpm',
+      key: 'rateLimitRpm',
+    },
+    {
+      title: '状态',
+      dataIndex: 'enabled',
+      key: 'enabled',
+      render: (v: boolean, record: GatewayApiKey) => (
+        <Switch
+          checked={v}
+          size="small"
+          onChange={() => toggleMutation.mutate(record.id)}
+        />
+      ),
+    },
+    {
+      title: '最后使用',
+      dataIndex: 'lastUsedAt',
+      key: 'lastUsedAt',
+      render: (v: string | null) => v ? new Date(v).toLocaleString() : '-',
+    },
+    {
+      title: '操作',
+      key: 'actions',
+      render: (_: any, record: GatewayApiKey) => (
+        <Popconfirm title="确认删除？" onConfirm={() => deleteMutation.mutate(record.id)}>
+          <Button type="link" danger size="small">删除</Button>
+        </Popconfirm>
+      ),
+    },
+  ];
+
+  return (
+    <>
+      <div style={{ display: 'flex', justifyContent: 'flex-end', marginBottom: 16 }}>
+        <Button type="primary" icon={<PlusOutlined />} onClick={() => setCreateOpen(true)}>
+          创建 API Key
+        </Button>
+      </div>
+
+      <Table
+        columns={columns}
+        dataSource={data?.items}
+        rowKey="id"
+        loading={isLoading}
+        size="small"
+        pagination={false}
+      />
+
+      {/* Create Modal */}
+      <Modal
+        title="创建 API Key"
+        open={createOpen}
+        onOk={handleCreate}
+        onCancel={() => setCreateOpen(false)}
+        confirmLoading={createMutation.isPending}
+      >
+        <Form form={form} layout="vertical">
+          <Form.Item name="name" label="名称" rules={[{ required: true, message: '请输入名称' }]}>
+            <Input placeholder="例如: 客户A测试Key" />
+          </Form.Item>
+          <Form.Item name="owner" label="所属用户/组织">
+            <Input placeholder="例如: Company A" />
+          </Form.Item>
+          <Form.Item name="rateLimitRpm" label="每分钟请求限制" initialValue={60}>
+            <InputNumber min={1} max={10000} style={{ width: '100%' }} />
+          </Form.Item>
+          <Form.Item name="rateLimitTpd" label="每日 Token 限制" initialValue={1000000}>
+            <InputNumber min={1000} max={100000000} style={{ width: '100%' }} />
+          </Form.Item>
+        </Form>
+      </Modal>
+
+      {/* New Key Display Modal */}
+      <Modal
+        title="API Key 已创建"
+        open={!!newKeyModal}
+        onOk={() => setNewKeyModal(null)}
+        onCancel={() => setNewKeyModal(null)}
+        cancelButtonProps={{ style: { display: 'none' } }}
+      >
+        <p>请立即复制保存此 Key，关闭后将无法再次查看：</p>
+        <div style={{ background: '#f5f5f5', padding: 12, borderRadius: 6, display: 'flex', alignItems: 'center', gap: 8 }}>
+          <Text code copyable style={{ flex: 1, wordBreak: 'break-all' }}>
+            {newKeyModal?.rawKey}
+          </Text>
+        </div>
+      </Modal>
+    </>
+  );
+}
--- a/packages/admin-client/src/features/llm-gateway/presentation/components/AuditLogsTab.tsx
+++ b/packages/admin-client/src/features/llm-gateway/presentation/components/AuditLogsTab.tsx
@ -0,0 +1,55 @@
+import { Table, Tag } from 'antd';
+import { useAuditLogs } from '../../application/useLLMGateway';
+
+export function AuditLogsTab() {
+  const { data, isLoading } = useAuditLogs();
+
+  const columns = [
+    {
+      title: '时间',
+      dataIndex: 'createdAt',
+      key: 'createdAt',
+      render: (v: string) => new Date(v).toLocaleString(),
+      width: 180,
+    },
+    { title: '方法', dataIndex: 'requestMethod', key: 'requestMethod', width: 70 },
+    { title: '路径', dataIndex: 'requestPath', key: 'requestPath', width: 160 },
+    { title: '模型', dataIndex: 'requestModel', key: 'requestModel', ellipsis: true },
+    { title: 'IP', dataIndex: 'requestIp', key: 'requestIp', width: 130 },
+    {
+      title: '内容过滤',
+      dataIndex: 'contentFiltered',
+      key: 'contentFiltered',
+      render: (v: boolean) => v ? <Tag color="red">已过滤</Tag> : <Tag color="green">通过</Tag>,
+    },
+    {
+      title: '注入',
+      dataIndex: 'injectionApplied',
+      key: 'injectionApplied',
+      render: (v: boolean) => v ? <Tag color="blue">已注入</Tag> : '-',
+    },
+    {
+      title: '状态码',
+      dataIndex: 'responseStatus',
+      key: 'responseStatus',
+      render: (v: number) => <Tag color={v < 400 ? 'green' : 'red'}>{v}</Tag>,
+    },
+    {
+      title: '耗时',
+      dataIndex: 'durationMs',
+      key: 'durationMs',
+      render: (v: number) => `${(v / 1000).toFixed(1)}s`,
+    },
+  ];
+
+  return (
+    <Table
+      columns={columns}
+      dataSource={data?.items}
+      rowKey="id"
+      loading={isLoading}
+      size="small"
+      pagination={{ pageSize: 20, showSizeChanger: false }}
+    />
+  );
+}
--- a/packages/admin-client/src/features/llm-gateway/presentation/components/ContentRulesTab.tsx
+++ b/packages/admin-client/src/features/llm-gateway/presentation/components/ContentRulesTab.tsx
@ -0,0 +1,109 @@
+import { useState } from 'react';
+import { Table, Button, Modal, Form, Input, Select, Switch, Popconfirm, Tag } from 'antd';
+import { PlusOutlined } from '@ant-design/icons';
+import { useContentRules, useCreateContentRule, useDeleteContentRule, useToggleContentRule } from '../../application/useLLMGateway';
+import type { ContentRule } from '../../infrastructure/llm-gateway.api';
+
+export function ContentRulesTab() {
+  const { data, isLoading } = useContentRules();
+  const createMutation = useCreateContentRule();
+  const deleteMutation = useDeleteContentRule();
+  const toggleMutation = useToggleContentRule();
+  const [createOpen, setCreateOpen] = useState(false);
+  const [form] = Form.useForm();
+
+  const handleCreate = async () => {
+    const values = await form.validateFields();
+    await createMutation.mutateAsync(values);
+    setCreateOpen(false);
+    form.resetFields();
+  };
+
+  const actionColors: Record<string, string> = { block: 'red', warn: 'orange', log: 'blue' };
+  const actionLabels: Record<string, string> = { block: '拦截', warn: '警告', log: '记录' };
+
+  const columns = [
+    { title: '名称', dataIndex: 'name', key: 'name' },
+    {
+      title: '类型',
+      dataIndex: 'type',
+      key: 'type',
+      render: (v: string) => <Tag>{v === 'keyword' ? '关键词' : '正则'}</Tag>,
+    },
+    {
+      title: '匹配模式',
+      dataIndex: 'pattern',
+      key: 'pattern',
+      ellipsis: true,
+      width: 200,
+    },
+    {
+      title: '动作',
+      dataIndex: 'action',
+      key: 'action',
+      render: (v: string) => <Tag color={actionColors[v]}>{actionLabels[v]}</Tag>,
+    },
+    {
+      title: '拒绝消息',
+      dataIndex: 'rejectMessage',
+      key: 'rejectMessage',
+      ellipsis: true,
+      width: 200,
+    },
+    {
+      title: '状态',
+      dataIndex: 'enabled',
+      key: 'enabled',
+      render: (v: boolean, record: ContentRule) => (
+        <Switch checked={v} size="small" onChange={() => toggleMutation.mutate(record.id)} />
+      ),
+    },
+    {
+      title: '操作',
+      key: 'actions',
+      render: (_: any, record: ContentRule) => (
+        <Popconfirm title="确认删除？" onConfirm={() => deleteMutation.mutate(record.id)}>
+          <Button type="link" danger size="small">删除</Button>
+        </Popconfirm>
+      ),
+    },
+  ];
+
+  return (
+    <>
+      <div style={{ display: 'flex', justifyContent: 'flex-end', marginBottom: 16 }}>
+        <Button type="primary" icon={<PlusOutlined />} onClick={() => setCreateOpen(true)}>
+          创建审查规则
+        </Button>
+      </div>
+
+      <Table columns={columns} dataSource={data?.items} rowKey="id" loading={isLoading} size="small" pagination={false} />
+
+      <Modal
+        title="创建内容审查规则"
+        open={createOpen}
+        onOk={handleCreate}
+        onCancel={() => setCreateOpen(false)}
+        confirmLoading={createMutation.isPending}
+      >
+        <Form form={form} layout="vertical">
+          <Form.Item name="name" label="规则名称" rules={[{ required: true }]}>
+            <Input placeholder="例如: 禁止政治敏感话题" />
+          </Form.Item>
+          <Form.Item name="type" label="匹配类型" initialValue="keyword">
+            <Select options={[{ label: '关键词匹配', value: 'keyword' }, { label: '正则表达式', value: 'regex' }]} />
+          </Form.Item>
+          <Form.Item name="pattern" label="匹配模式" rules={[{ required: true }]}>
+            <Input placeholder="关键词或正则表达式" />
+          </Form.Item>
+          <Form.Item name="action" label="触发动作" initialValue="block">
+            <Select options={[{ label: '拦截请求', value: 'block' }, { label: '警告并放行', value: 'warn' }, { label: '仅记录', value: 'log' }]} />
+          </Form.Item>
+          <Form.Item name="rejectMessage" label="拒绝消息（拦截时返回给用户）">
+            <Input placeholder="Content blocked by policy." />
+          </Form.Item>
+        </Form>
+      </Modal>
+    </>
+  );
+}
--- a/packages/admin-client/src/features/llm-gateway/presentation/components/InjectionRulesTab.tsx
+++ b/packages/admin-client/src/features/llm-gateway/presentation/components/InjectionRulesTab.tsx
@ -0,0 +1,99 @@
+import { useState } from 'react';
+import { Table, Button, Modal, Form, Input, Select, Switch, Space, Popconfirm, Tag } from 'antd';
+import { PlusOutlined } from '@ant-design/icons';
+import { useInjectionRules, useCreateInjectionRule, useDeleteInjectionRule, useToggleInjectionRule } from '../../application/useLLMGateway';
+import type { InjectionRule } from '../../infrastructure/llm-gateway.api';
+
+const { TextArea } = Input;
+
+export function InjectionRulesTab() {
+  const { data, isLoading } = useInjectionRules();
+  const createMutation = useCreateInjectionRule();
+  const deleteMutation = useDeleteInjectionRule();
+  const toggleMutation = useToggleInjectionRule();
+  const [createOpen, setCreateOpen] = useState(false);
+  const [form] = Form.useForm();
+
+  const handleCreate = async () => {
+    const values = await form.validateFields();
+    await createMutation.mutateAsync(values);
+    setCreateOpen(false);
+    form.resetFields();
+  };
+
+  const columns = [
+    { title: '名称', dataIndex: 'name', key: 'name' },
+    {
+      title: '位置',
+      dataIndex: 'position',
+      key: 'position',
+      render: (v: string) => <Tag color={v === 'prepend' ? 'blue' : 'green'}>{v === 'prepend' ? '前置' : '追加'}</Tag>,
+    },
+    {
+      title: '内容预览',
+      dataIndex: 'content',
+      key: 'content',
+      ellipsis: true,
+      width: 300,
+    },
+    {
+      title: '匹配模型',
+      dataIndex: 'matchModels',
+      key: 'matchModels',
+      render: (v: string[]) => v?.includes('*') ? <Tag>全部</Tag> : v?.map((m: string) => <Tag key={m}>{m}</Tag>),
+    },
+    {
+      title: '状态',
+      dataIndex: 'enabled',
+      key: 'enabled',
+      render: (v: boolean, record: InjectionRule) => (
+        <Switch checked={v} size="small" onChange={() => toggleMutation.mutate(record.id)} />
+      ),
+    },
+    {
+      title: '操作',
+      key: 'actions',
+      render: (_: any, record: InjectionRule) => (
+        <Popconfirm title="确认删除？" onConfirm={() => deleteMutation.mutate(record.id)}>
+          <Button type="link" danger size="small">删除</Button>
+        </Popconfirm>
+      ),
+    },
+  ];
+
+  return (
+    <>
+      <div style={{ display: 'flex', justifyContent: 'flex-end', marginBottom: 16 }}>
+        <Button type="primary" icon={<PlusOutlined />} onClick={() => setCreateOpen(true)}>
+          创建注入规则
+        </Button>
+      </div>
+
+      <Table columns={columns} dataSource={data?.items} rowKey="id" loading={isLoading} size="small" pagination={false} />
+
+      <Modal
+        title="创建注入规则"
+        open={createOpen}
+        onOk={handleCreate}
+        onCancel={() => setCreateOpen(false)}
+        confirmLoading={createMutation.isPending}
+        width={600}
+      >
+        <Form form={form} layout="vertical">
+          <Form.Item name="name" label="规则名称" rules={[{ required: true }]}>
+            <Input placeholder="例如: 合规声明" />
+          </Form.Item>
+          <Form.Item name="position" label="注入位置" initialValue="append">
+            <Select options={[{ label: '追加到 system prompt 末尾', value: 'append' }, { label: '前置到 system prompt 开头', value: 'prepend' }]} />
+          </Form.Item>
+          <Form.Item name="content" label="注入内容" rules={[{ required: true }]}>
+            <TextArea rows={6} placeholder="输入要注入到 system prompt 中的监管内容..." />
+          </Form.Item>
+          <Form.Item name="description" label="描述">
+            <Input placeholder="规则说明（可选）" />
+          </Form.Item>
+        </Form>
+      </Modal>
+    </>
+  );
+}
--- a/packages/admin-client/src/features/llm-gateway/presentation/components/UsageDashboardTab.tsx
+++ b/packages/admin-client/src/features/llm-gateway/presentation/components/UsageDashboardTab.tsx
@ -0,0 +1,55 @@
+import { Table, Tag } from 'antd';
+import { useUsageLogs } from '../../application/useLLMGateway';
+
+export function UsageDashboardTab() {
+  const { data, isLoading } = useUsageLogs();
+
+  const columns = [
+    {
+      title: '时间',
+      dataIndex: 'createdAt',
+      key: 'createdAt',
+      render: (v: string) => new Date(v).toLocaleString(),
+      width: 180,
+    },
+    {
+      title: '提供商',
+      dataIndex: 'provider',
+      key: 'provider',
+      render: (v: string) => <Tag color={v === 'anthropic' ? 'purple' : 'green'}>{v}</Tag>,
+    },
+    { title: '模型', dataIndex: 'model', key: 'model', ellipsis: true },
+    { title: '输入 Tokens', dataIndex: 'inputTokens', key: 'inputTokens' },
+    { title: '输出 Tokens', dataIndex: 'outputTokens', key: 'outputTokens' },
+    { title: '总 Tokens', dataIndex: 'totalTokens', key: 'totalTokens' },
+    {
+      title: '费用 (USD)',
+      dataIndex: 'costUsd',
+      key: 'costUsd',
+      render: (v: number | null) => v !== null ? `$${Number(v).toFixed(6)}` : '-',
+    },
+    {
+      title: '耗时',
+      dataIndex: 'durationMs',
+      key: 'durationMs',
+      render: (v: number) => `${(v / 1000).toFixed(1)}s`,
+    },
+    {
+      title: '状态码',
+      dataIndex: 'statusCode',
+      key: 'statusCode',
+      render: (v: number) => <Tag color={v === 200 ? 'green' : 'red'}>{v}</Tag>,
+    },
+  ];
+
+  return (
+    <Table
+      columns={columns}
+      dataSource={data?.items}
+      rowKey="id"
+      loading={isLoading}
+      size="small"
+      pagination={{ pageSize: 20, showSizeChanger: false }}
+    />
+  );
+}
--- a/packages/admin-client/src/features/llm-gateway/presentation/pages/LLMGatewayPage.tsx
+++ b/packages/admin-client/src/features/llm-gateway/presentation/pages/LLMGatewayPage.tsx
@ -0,0 +1,103 @@
+import { useState } from 'react';
+import { Typography, Tabs, Card, Row, Col, Statistic, Spin } from 'antd';
+import { KeyOutlined, EditOutlined, SafetyOutlined, BarChartOutlined, FileSearchOutlined } from '@ant-design/icons';
+import { useDashboard } from '../../application/useLLMGateway';
+import { ApiKeysTab } from '../components/ApiKeysTab';
+import { InjectionRulesTab } from '../components/InjectionRulesTab';
+import { ContentRulesTab } from '../components/ContentRulesTab';
+import { UsageDashboardTab } from '../components/UsageDashboardTab';
+import { AuditLogsTab } from '../components/AuditLogsTab';
+
+const { Title } = Typography;
+
+export function LLMGatewayPage() {
+  const [activeTab, setActiveTab] = useState('keys');
+  const { data: dashboard, isLoading: dashLoading } = useDashboard();
+
+  return (
+    <div style={{ padding: 24 }}>
+      <Title level={4} style={{ marginBottom: 16 }}>LLM 网关管理</Title>
+
+      {/* Dashboard Summary Cards */}
+      <Row gutter={[16, 16]} style={{ marginBottom: 16 }}>
+        <Col xs={12} sm={6}>
+          <Card size="small">
+            <Statistic
+              title="活跃 Key"
+              value={dashLoading ? '-' : `${dashboard?.keys.active || 0} / ${dashboard?.keys.total || 0}`}
+              prefix={<KeyOutlined />}
+            />
+          </Card>
+        </Col>
+        <Col xs={12} sm={6}>
+          <Card size="small">
+            <Statistic
+              title="今日请求"
+              value={dashboard?.today.requests || 0}
+              loading={dashLoading}
+            />
+          </Card>
+        </Col>
+        <Col xs={12} sm={6}>
+          <Card size="small">
+            <Statistic
+              title="今日 Tokens"
+              value={dashboard?.today.tokens || 0}
+              loading={dashLoading}
+              formatter={(val) => {
+                const n = Number(val);
+                return n >= 1000000 ? `${(n / 1000000).toFixed(1)}M` : n >= 1000 ? `${(n / 1000).toFixed(1)}K` : String(n);
+              }}
+            />
+          </Card>
+        </Col>
+        <Col xs={12} sm={6}>
+          <Card size="small">
+            <Statistic
+              title="今日费用"
+              value={dashboard?.today.cost || 0}
+              loading={dashLoading}
+              precision={4}
+              prefix="$"
+            />
+          </Card>
+        </Col>
+      </Row>
+
+      {/* Tabs */}
+      <Card>
+        <Tabs
+          activeKey={activeTab}
+          onChange={setActiveTab}
+          items={[
+            {
+              key: 'keys',
+              label: <span><KeyOutlined /> API Key</span>,
+              children: <ApiKeysTab />,
+            },
+            {
+              key: 'injection',
+              label: <span><EditOutlined /> 注入规则</span>,
+              children: <InjectionRulesTab />,
+            },
+            {
+              key: 'content',
+              label: <span><SafetyOutlined /> 内容审查</span>,
+              children: <ContentRulesTab />,
+            },
+            {
+              key: 'usage',
+              label: <span><BarChartOutlined /> 用量统计</span>,
+              children: <UsageDashboardTab />,
+            },
+            {
+              key: 'audit',
+              label: <span><FileSearchOutlined /> 审计日志</span>,
+              children: <AuditLogsTab />,
+            },
+          ]}
+        />
+      </Card>
+    </div>
+  );
+}
--- a/packages/admin-client/src/shared/components/MainLayout.tsx
+++ b/packages/admin-client/src/shared/components/MainLayout.tsx
@ -22,6 +22,7 @@ import {
  ExperimentOutlined,
  FormOutlined,
  EyeOutlined,
+  GatewayOutlined,
 } from '@ant-design/icons';
 import { useAuth } from '../hooks/useAuth';

@ -59,6 +60,11 @@ const menuItems: MenuProps['items'] = [
    icon: <EyeOutlined />,
    label: '系统总监',
  },
+  {
+    key: '/llm-gateway',
+    icon: <GatewayOutlined />,
+    label: 'LLM 网关',
+  },
  {
    key: 'analytics-group',
    icon: <BarChartOutlined />,
--- a/packages/services/conversation-service/src/adapters/inbound/admin-gateway.controller.ts
+++ b/packages/services/conversation-service/src/adapters/inbound/admin-gateway.controller.ts
@ -0,0 +1,605 @@
+/**
+ * Admin Gateway Controller
+ * LLM 网关管理 API — API Key / 注入规则 / 内容审查规则 / 用量 / 审计日志
+ *
+ * ⚠️ 路由顺序：静态路由（keys, injection-rules, content-rules, usage, audit-logs, dashboard）
+ *    必须在动态路由（:id）之前声明
+ */
+
+import {
+  Controller,
+  Get,
+  Post,
+  Put,
+  Delete,
+  Param,
+  Body,
+  Query,
+  Headers,
+  HttpCode,
+  HttpStatus,
+  UnauthorizedException,
+  BadRequestException,
+} from '@nestjs/common';
+import { InjectRepository } from '@nestjs/typeorm';
+import { Repository, IsNull, Between, MoreThanOrEqual } from 'typeorm';
+import * as jwt from 'jsonwebtoken';
+import { v4 as uuidv4 } from 'uuid';
+import { createHash, randomBytes } from 'crypto';
+import { GatewayApiKeyORM } from '../../infrastructure/database/postgres/entities/gateway-api-key.orm';
+import { GatewayInjectionRuleORM } from '../../infrastructure/database/postgres/entities/gateway-injection-rule.orm';
+import { GatewayContentRuleORM } from '../../infrastructure/database/postgres/entities/gateway-content-rule.orm';
+import { GatewayUsageLogORM } from '../../infrastructure/database/postgres/entities/gateway-usage-log.orm';
+import { GatewayAuditLogORM } from '../../infrastructure/database/postgres/entities/gateway-audit-log.orm';
+
+interface AdminPayload {
+  id: string;
+  username: string;
+  role: string;
+  tenantId?: string;
+}
+
+// ═══════════════════════════════════════════════════════════════
+// API Key 管理
+// ═══════════════════════════════════════════════════════════════
+
+@Controller('conversations/admin/gateway/keys')
+export class AdminGatewayKeysController {
+  constructor(
+    @InjectRepository(GatewayApiKeyORM)
+    private readonly repo: Repository<GatewayApiKeyORM>,
+  ) {}
+
+  private verifyAdmin(authorization: string): AdminPayload {
+    const token = authorization?.replace('Bearer ', '');
+    if (!token) throw new UnauthorizedException('Missing token');
+    try {
+      return jwt.verify(token, process.env.JWT_SECRET || 'your-jwt-secret-key') as AdminPayload;
+    } catch {
+      throw new UnauthorizedException('Invalid token');
+    }
+  }
+
+  @Get()
+  async listKeys(@Headers('authorization') auth: string) {
+    const admin = this.verifyAdmin(auth);
+    const tenantId = admin.tenantId || null;
+
+    const keys = await this.repo.find({
+      where: { tenantId: tenantId ?? IsNull() },
+      order: { createdAt: 'DESC' },
+    });
+
+    return {
+      success: true,
+      data: {
+        items: keys,
+        total: keys.length,
+        enabledCount: keys.filter((k) => k.enabled).length,
+      },
+    };
+  }
+
+  @Post()
+  @HttpCode(HttpStatus.CREATED)
+  async createKey(
+    @Headers('authorization') auth: string,
+    @Body() dto: {
+      name: string;
+      owner?: string;
+      permissions?: { allowedModels?: string[]; allowStreaming?: boolean; allowTools?: boolean };
+      rateLimitRpm?: number;
+      rateLimitTpd?: number;
+      monthlyBudget?: number;
+      expiresAt?: string;
+    },
+  ) {
+    const admin = this.verifyAdmin(auth);
+    const tenantId = admin.tenantId || null;
+
+    if (!dto.name?.trim()) {
+      throw new BadRequestException('name is required');
+    }
+
+    // Generate a unique API key: sk-gw-<32 random hex chars>
+    const rawKey = `sk-gw-${randomBytes(16).toString('hex')}`;
+    const keyHash = createHash('sha256').update(rawKey).digest('hex');
+    const keyPrefix = rawKey.slice(0, 12);
+
+    const entity = this.repo.create({
+      id: uuidv4(),
+      tenantId,
+      keyHash,
+      keyPrefix,
+      name: dto.name.trim(),
+      owner: dto.owner?.trim() || '',
+      permissions: {
+        allowedModels: dto.permissions?.allowedModels || ['*'],
+        allowStreaming: dto.permissions?.allowStreaming !== false,
+        allowTools: dto.permissions?.allowTools !== false,
+      },
+      rateLimitRpm: dto.rateLimitRpm ?? 60,
+      rateLimitTpd: dto.rateLimitTpd ?? 1000000,
+      monthlyBudget: dto.monthlyBudget ?? null,
+      enabled: true,
+      expiresAt: dto.expiresAt ? new Date(dto.expiresAt) : null,
+      createdBy: admin.id,
+    });
+
+    const saved = await this.repo.save(entity);
+
+    // Return the raw key only on creation (never stored, only the hash)
+    return {
+      success: true,
+      data: {
+        ...saved,
+        rawKey, // ⚠️ 仅此一次返回完整 Key
+      },
+    };
+  }
+
+  @Put(':id')
+  async updateKey(
+    @Headers('authorization') auth: string,
+    @Param('id') id: string,
+    @Body() dto: {
+      name?: string;
+      owner?: string;
+      permissions?: { allowedModels?: string[]; allowStreaming?: boolean; allowTools?: boolean };
+      rateLimitRpm?: number;
+      rateLimitTpd?: number;
+      monthlyBudget?: number | null;
+      expiresAt?: string | null;
+      enabled?: boolean;
+    },
+  ) {
+    this.verifyAdmin(auth);
+
+    const key = await this.repo.findOne({ where: { id } });
+    if (!key) return { success: false, error: 'Key not found' };
+
+    if (dto.name !== undefined) key.name = dto.name.trim();
+    if (dto.owner !== undefined) key.owner = dto.owner.trim();
+    if (dto.permissions) key.permissions = { ...key.permissions, ...dto.permissions };
+    if (dto.rateLimitRpm !== undefined) key.rateLimitRpm = dto.rateLimitRpm;
+    if (dto.rateLimitTpd !== undefined) key.rateLimitTpd = dto.rateLimitTpd;
+    if (dto.monthlyBudget !== undefined) key.monthlyBudget = dto.monthlyBudget;
+    if (dto.expiresAt !== undefined) key.expiresAt = dto.expiresAt ? new Date(dto.expiresAt) : null;
+    if (dto.enabled !== undefined) key.enabled = dto.enabled;
+
+    const updated = await this.repo.save(key);
+    return { success: true, data: updated };
+  }
+
+  @Delete(':id')
+  async deleteKey(
+    @Headers('authorization') auth: string,
+    @Param('id') id: string,
+  ) {
+    this.verifyAdmin(auth);
+    const key = await this.repo.findOne({ where: { id } });
+    if (!key) return { success: false, error: 'Key not found' };
+    await this.repo.delete(id);
+    return { success: true };
+  }
+
+  @Post(':id/toggle')
+  @HttpCode(HttpStatus.OK)
+  async toggleKey(
+    @Headers('authorization') auth: string,
+    @Param('id') id: string,
+  ) {
+    this.verifyAdmin(auth);
+    const key = await this.repo.findOne({ where: { id } });
+    if (!key) return { success: false, error: 'Key not found' };
+    key.enabled = !key.enabled;
+    const updated = await this.repo.save(key);
+    return { success: true, data: updated };
+  }
+}
+
+// ═══════════════════════════════════════════════════════════════
+// 注入规则管理
+// ═══════════════════════════════════════════════════════════════
+
+@Controller('conversations/admin/gateway/injection-rules')
+export class AdminGatewayInjectionRulesController {
+  constructor(
+    @InjectRepository(GatewayInjectionRuleORM)
+    private readonly repo: Repository<GatewayInjectionRuleORM>,
+  ) {}
+
+  private verifyAdmin(authorization: string): AdminPayload {
+    const token = authorization?.replace('Bearer ', '');
+    if (!token) throw new UnauthorizedException('Missing token');
+    try {
+      return jwt.verify(token, process.env.JWT_SECRET || 'your-jwt-secret-key') as AdminPayload;
+    } catch {
+      throw new UnauthorizedException('Invalid token');
+    }
+  }
+
+  @Get()
+  async listRules(@Headers('authorization') auth: string) {
+    const admin = this.verifyAdmin(auth);
+    const tenantId = admin.tenantId || null;
+
+    const rules = await this.repo.find({
+      where: { tenantId: tenantId ?? IsNull() },
+      order: { priority: 'ASC', createdAt: 'ASC' },
+    });
+
+    return {
+      success: true,
+      data: {
+        items: rules,
+        total: rules.length,
+        enabledCount: rules.filter((r) => r.enabled).length,
+      },
+    };
+  }
+
+  @Post()
+  @HttpCode(HttpStatus.CREATED)
+  async createRule(
+    @Headers('authorization') auth: string,
+    @Body() dto: {
+      name: string;
+      description?: string;
+      position?: 'prepend' | 'append';
+      content: string;
+      matchModels?: string[];
+      matchKeyIds?: string[];
+      priority?: number;
+    },
+  ) {
+    const admin = this.verifyAdmin(auth);
+    const tenantId = admin.tenantId || null;
+
+    if (!dto.name?.trim()) throw new BadRequestException('name is required');
+    if (!dto.content?.trim()) throw new BadRequestException('content is required');
+
+    const rule = this.repo.create({
+      id: uuidv4(),
+      tenantId,
+      name: dto.name.trim(),
+      description: dto.description || null,
+      position: dto.position || 'append',
+      content: dto.content.trim(),
+      matchModels: dto.matchModels || ['*'],
+      matchKeyIds: dto.matchKeyIds || null,
+      priority: dto.priority ?? 0,
+      enabled: true,
+      createdBy: admin.id,
+    });
+
+    const saved = await this.repo.save(rule);
+    return { success: true, data: saved };
+  }
+
+  @Put(':id')
+  async updateRule(
+    @Headers('authorization') auth: string,
+    @Param('id') id: string,
+    @Body() dto: {
+      name?: string;
+      description?: string | null;
+      position?: 'prepend' | 'append';
+      content?: string;
+      matchModels?: string[];
+      matchKeyIds?: string[] | null;
+      priority?: number;
+      enabled?: boolean;
+    },
+  ) {
+    this.verifyAdmin(auth);
+    const rule = await this.repo.findOne({ where: { id } });
+    if (!rule) return { success: false, error: 'Rule not found' };
+
+    if (dto.name !== undefined) rule.name = dto.name.trim();
+    if (dto.description !== undefined) rule.description = dto.description;
+    if (dto.position !== undefined) rule.position = dto.position;
+    if (dto.content !== undefined) rule.content = dto.content.trim();
+    if (dto.matchModels !== undefined) rule.matchModels = dto.matchModels;
+    if (dto.matchKeyIds !== undefined) rule.matchKeyIds = dto.matchKeyIds;
+    if (dto.priority !== undefined) rule.priority = dto.priority;
+    if (dto.enabled !== undefined) rule.enabled = dto.enabled;
+
+    const updated = await this.repo.save(rule);
+    return { success: true, data: updated };
+  }
+
+  @Delete(':id')
+  async deleteRule(@Headers('authorization') auth: string, @Param('id') id: string) {
+    this.verifyAdmin(auth);
+    const rule = await this.repo.findOne({ where: { id } });
+    if (!rule) return { success: false, error: 'Rule not found' };
+    await this.repo.delete(id);
+    return { success: true };
+  }
+
+  @Post(':id/toggle')
+  @HttpCode(HttpStatus.OK)
+  async toggleRule(@Headers('authorization') auth: string, @Param('id') id: string) {
+    this.verifyAdmin(auth);
+    const rule = await this.repo.findOne({ where: { id } });
+    if (!rule) return { success: false, error: 'Rule not found' };
+    rule.enabled = !rule.enabled;
+    const updated = await this.repo.save(rule);
+    return { success: true, data: updated };
+  }
+}
+
+// ═══════════════════════════════════════════════════════════════
+// 内容审查规则管理
+// ═══════════════════════════════════════════════════════════════
+
+@Controller('conversations/admin/gateway/content-rules')
+export class AdminGatewayContentRulesController {
+  constructor(
+    @InjectRepository(GatewayContentRuleORM)
+    private readonly repo: Repository<GatewayContentRuleORM>,
+  ) {}
+
+  private verifyAdmin(authorization: string): AdminPayload {
+    const token = authorization?.replace('Bearer ', '');
+    if (!token) throw new UnauthorizedException('Missing token');
+    try {
+      return jwt.verify(token, process.env.JWT_SECRET || 'your-jwt-secret-key') as AdminPayload;
+    } catch {
+      throw new UnauthorizedException('Invalid token');
+    }
+  }
+
+  @Get()
+  async listRules(@Headers('authorization') auth: string) {
+    const admin = this.verifyAdmin(auth);
+    const tenantId = admin.tenantId || null;
+
+    const rules = await this.repo.find({
+      where: { tenantId: tenantId ?? IsNull() },
+      order: { priority: 'ASC', createdAt: 'ASC' },
+    });
+
+    return {
+      success: true,
+      data: {
+        items: rules,
+        total: rules.length,
+        enabledCount: rules.filter((r) => r.enabled).length,
+      },
+    };
+  }
+
+  @Post()
+  @HttpCode(HttpStatus.CREATED)
+  async createRule(
+    @Headers('authorization') auth: string,
+    @Body() dto: {
+      name: string;
+      type?: 'keyword' | 'regex';
+      pattern: string;
+      action?: 'block' | 'warn' | 'log';
+      rejectMessage?: string;
+      priority?: number;
+    },
+  ) {
+    const admin = this.verifyAdmin(auth);
+    const tenantId = admin.tenantId || null;
+
+    if (!dto.name?.trim()) throw new BadRequestException('name is required');
+    if (!dto.pattern?.trim()) throw new BadRequestException('pattern is required');
+
+    // Validate regex if type is regex
+    if (dto.type === 'regex') {
+      try {
+        new RegExp(dto.pattern);
+      } catch {
+        throw new BadRequestException('Invalid regex pattern');
+      }
+    }
+
+    const rule = this.repo.create({
+      id: uuidv4(),
+      tenantId,
+      name: dto.name.trim(),
+      type: dto.type || 'keyword',
+      pattern: dto.pattern.trim(),
+      action: dto.action || 'block',
+      rejectMessage: dto.rejectMessage || null,
+      priority: dto.priority ?? 0,
+      enabled: true,
+      createdBy: admin.id,
+    });
+
+    const saved = await this.repo.save(rule);
+    return { success: true, data: saved };
+  }
+
+  @Put(':id')
+  async updateRule(
+    @Headers('authorization') auth: string,
+    @Param('id') id: string,
+    @Body() dto: {
+      name?: string;
+      type?: 'keyword' | 'regex';
+      pattern?: string;
+      action?: 'block' | 'warn' | 'log';
+      rejectMessage?: string | null;
+      priority?: number;
+      enabled?: boolean;
+    },
+  ) {
+    this.verifyAdmin(auth);
+    const rule = await this.repo.findOne({ where: { id } });
+    if (!rule) return { success: false, error: 'Rule not found' };
+
+    if (dto.name !== undefined) rule.name = dto.name.trim();
+    if (dto.type !== undefined) rule.type = dto.type;
+    if (dto.pattern !== undefined) {
+      if (dto.type === 'regex' || rule.type === 'regex') {
+        try { new RegExp(dto.pattern); } catch { throw new BadRequestException('Invalid regex pattern'); }
+      }
+      rule.pattern = dto.pattern.trim();
+    }
+    if (dto.action !== undefined) rule.action = dto.action;
+    if (dto.rejectMessage !== undefined) rule.rejectMessage = dto.rejectMessage;
+    if (dto.priority !== undefined) rule.priority = dto.priority;
+    if (dto.enabled !== undefined) rule.enabled = dto.enabled;
+
+    const updated = await this.repo.save(rule);
+    return { success: true, data: updated };
+  }
+
+  @Delete(':id')
+  async deleteRule(@Headers('authorization') auth: string, @Param('id') id: string) {
+    this.verifyAdmin(auth);
+    const rule = await this.repo.findOne({ where: { id } });
+    if (!rule) return { success: false, error: 'Rule not found' };
+    await this.repo.delete(id);
+    return { success: true };
+  }
+
+  @Post(':id/toggle')
+  @HttpCode(HttpStatus.OK)
+  async toggleRule(@Headers('authorization') auth: string, @Param('id') id: string) {
+    this.verifyAdmin(auth);
+    const rule = await this.repo.findOne({ where: { id } });
+    if (!rule) return { success: false, error: 'Rule not found' };
+    rule.enabled = !rule.enabled;
+    const updated = await this.repo.save(rule);
+    return { success: true, data: updated };
+  }
+}
+
+// ═══════════════════════════════════════════════════════════════
+// 用量统计 + 审计日志 + 仪表盘
+// ═══════════════════════════════════════════════════════════════
+
+@Controller('conversations/admin/gateway')
+export class AdminGatewayDashboardController {
+  constructor(
+    @InjectRepository(GatewayUsageLogORM)
+    private readonly usageRepo: Repository<GatewayUsageLogORM>,
+    @InjectRepository(GatewayAuditLogORM)
+    private readonly auditRepo: Repository<GatewayAuditLogORM>,
+    @InjectRepository(GatewayApiKeyORM)
+    private readonly keyRepo: Repository<GatewayApiKeyORM>,
+  ) {}
+
+  private verifyAdmin(authorization: string): AdminPayload {
+    const token = authorization?.replace('Bearer ', '');
+    if (!token) throw new UnauthorizedException('Missing token');
+    try {
+      return jwt.verify(token, process.env.JWT_SECRET || 'your-jwt-secret-key') as AdminPayload;
+    } catch {
+      throw new UnauthorizedException('Invalid token');
+    }
+  }
+
+  @Get('dashboard')
+  async getDashboard(@Headers('authorization') auth: string) {
+    this.verifyAdmin(auth);
+
+    const now = new Date();
+    const todayStart = new Date(now.getFullYear(), now.getMonth(), now.getDate());
+    const monthStart = new Date(now.getFullYear(), now.getMonth(), 1);
+
+    // Parallel queries
+    const [totalKeys, activeKeys, todayUsage, monthUsage, recentAudits] = await Promise.all([
+      this.keyRepo.count(),
+      this.keyRepo.count({ where: { enabled: true } }),
+      this.usageRepo
+        .createQueryBuilder('u')
+        .select('COUNT(*)', 'requestCount')
+        .addSelect('COALESCE(SUM(u.total_tokens), 0)', 'totalTokens')
+        .addSelect('COALESCE(SUM(u.cost_usd), 0)', 'totalCost')
+        .where('u.created_at >= :start', { start: todayStart })
+        .getRawOne(),
+      this.usageRepo
+        .createQueryBuilder('u')
+        .select('COUNT(*)', 'requestCount')
+        .addSelect('COALESCE(SUM(u.total_tokens), 0)', 'totalTokens')
+        .addSelect('COALESCE(SUM(u.cost_usd), 0)', 'totalCost')
+        .where('u.created_at >= :start', { start: monthStart })
+        .getRawOne(),
+      this.auditRepo.count({
+        where: { contentFiltered: true, createdAt: MoreThanOrEqual(todayStart) },
+      }),
+    ]);
+
+    return {
+      success: true,
+      data: {
+        keys: { total: totalKeys, active: activeKeys },
+        today: {
+          requests: parseInt(todayUsage?.requestCount || '0'),
+          tokens: parseInt(todayUsage?.totalTokens || '0'),
+          cost: parseFloat(todayUsage?.totalCost || '0'),
+          filtered: recentAudits,
+        },
+        month: {
+          requests: parseInt(monthUsage?.requestCount || '0'),
+          tokens: parseInt(monthUsage?.totalTokens || '0'),
+          cost: parseFloat(monthUsage?.totalCost || '0'),
+        },
+      },
+    };
+  }
+
+  @Get('usage')
+  async getUsage(
+    @Headers('authorization') auth: string,
+    @Query('keyId') keyId?: string,
+    @Query('startDate') startDate?: string,
+    @Query('endDate') endDate?: string,
+    @Query('limit') limit?: string,
+  ) {
+    this.verifyAdmin(auth);
+
+    const qb = this.usageRepo.createQueryBuilder('u');
+
+    if (keyId) qb.andWhere('u.api_key_id = :keyId', { keyId });
+    if (startDate) qb.andWhere('u.created_at >= :startDate', { startDate: new Date(startDate) });
+    if (endDate) qb.andWhere('u.created_at <= :endDate', { endDate: new Date(endDate) });
+
+    qb.orderBy('u.created_at', 'DESC');
+    qb.take(parseInt(limit || '100'));
+
+    const [items, total] = await qb.getManyAndCount();
+
+    return {
+      success: true,
+      data: { items, total },
+    };
+  }
+
+  @Get('audit-logs')
+  async getAuditLogs(
+    @Headers('authorization') auth: string,
+    @Query('keyId') keyId?: string,
+    @Query('filtered') filtered?: string,
+    @Query('startDate') startDate?: string,
+    @Query('endDate') endDate?: string,
+    @Query('limit') limit?: string,
+  ) {
+    this.verifyAdmin(auth);
+
+    const qb = this.auditRepo.createQueryBuilder('a');
+
+    if (keyId) qb.andWhere('a.api_key_id = :keyId', { keyId });
+    if (filtered === 'true') qb.andWhere('a.content_filtered = true');
+    if (startDate) qb.andWhere('a.created_at >= :startDate', { startDate: new Date(startDate) });
+    if (endDate) qb.andWhere('a.created_at <= :endDate', { endDate: new Date(endDate) });
+
+    qb.orderBy('a.created_at', 'DESC');
+    qb.take(parseInt(limit || '100'));
+
+    const [items, total] = await qb.getManyAndCount();
+
+    return {
+      success: true,
+      data: { items, total },
+    };
+  }
+}
--- a/packages/services/conversation-service/src/conversation/conversation.module.ts
+++ b/packages/services/conversation-service/src/conversation/conversation.module.ts
@ -6,6 +6,11 @@ import { TokenUsageORM } from '../infrastructure/database/postgres/entities/toke
 import { AgentExecutionORM } from '../infrastructure/database/postgres/entities/agent-execution.orm';
 import { AssessmentDirectiveORM } from '../infrastructure/database/postgres/entities/assessment-directive.orm';
 import { CollectionDirectiveORM } from '../infrastructure/database/postgres/entities/collection-directive.orm';
+import { GatewayApiKeyORM } from '../infrastructure/database/postgres/entities/gateway-api-key.orm';
+import { GatewayInjectionRuleORM } from '../infrastructure/database/postgres/entities/gateway-injection-rule.orm';
+import { GatewayContentRuleORM } from '../infrastructure/database/postgres/entities/gateway-content-rule.orm';
+import { GatewayUsageLogORM } from '../infrastructure/database/postgres/entities/gateway-usage-log.orm';
+import { GatewayAuditLogORM } from '../infrastructure/database/postgres/entities/gateway-audit-log.orm';
 import { ConversationPostgresRepository } from '../adapters/outbound/persistence/conversation-postgres.repository';
 import { MessagePostgresRepository } from '../adapters/outbound/persistence/message-postgres.repository';
 import { TokenUsagePostgresRepository } from '../adapters/outbound/persistence/token-usage-postgres.repository';
@ -22,12 +27,18 @@ import { AdminObservabilityController } from '../adapters/inbound/admin-observab
 import { AdminAssessmentDirectiveController } from '../adapters/inbound/admin-assessment-directive.controller';
 import { AdminCollectionDirectiveController } from '../adapters/inbound/admin-collection-directive.controller';
 import { AdminSupervisorController } from '../adapters/inbound/admin-supervisor.controller';
+import {
+  AdminGatewayKeysController,
+  AdminGatewayInjectionRulesController,
+  AdminGatewayContentRulesController,
+  AdminGatewayDashboardController,
+} from '../adapters/inbound/admin-gateway.controller';
 import { ConversationGateway } from '../adapters/inbound/conversation.gateway';
 import { PaymentModule } from '../infrastructure/payment/payment.module';

@Module({
-  imports: [TypeOrmModule.forFeature([ConversationORM, MessageORM, TokenUsageORM, AgentExecutionORM, AssessmentDirectiveORM, CollectionDirectiveORM]), PaymentModule],
-  controllers: [ConversationController, InternalConversationController, AdminMcpController, AdminEvaluationRuleController, AdminAssessmentDirectiveController, AdminCollectionDirectiveController, AdminSupervisorController, AdminConversationController, AdminObservabilityController],
+  imports: [TypeOrmModule.forFeature([ConversationORM, MessageORM, TokenUsageORM, AgentExecutionORM, AssessmentDirectiveORM, CollectionDirectiveORM, GatewayApiKeyORM, GatewayInjectionRuleORM, GatewayContentRuleORM, GatewayUsageLogORM, GatewayAuditLogORM]), PaymentModule],
+  controllers: [ConversationController, InternalConversationController, AdminMcpController, AdminEvaluationRuleController, AdminAssessmentDirectiveController, AdminCollectionDirectiveController, AdminSupervisorController, AdminConversationController, AdminObservabilityController, AdminGatewayKeysController, AdminGatewayInjectionRulesController, AdminGatewayContentRulesController, AdminGatewayDashboardController],
  providers: [
    ConversationService,
    ConversationGateway,
--- a/packages/services/conversation-service/src/infrastructure/database/postgres/entities/gateway-api-key.orm.ts
+++ b/packages/services/conversation-service/src/infrastructure/database/postgres/entities/gateway-api-key.orm.ts
@ -0,0 +1,60 @@
+import {
+  Entity,
+  PrimaryGeneratedColumn,
+  Column,
+  Index,
+} from 'typeorm';
+
+@Entity('gateway_api_keys')
+@Index('idx_gateway_api_keys_tenant', ['tenantId'])
+@Index('idx_gateway_api_keys_enabled', ['enabled'])
+export class GatewayApiKeyORM {
+  @PrimaryGeneratedColumn('uuid')
+  id: string;
+
+  @Column({ name: 'tenant_id', type: 'uuid', nullable: true })
+  tenantId: string | null;
+
+  @Column({ name: 'key_hash', type: 'varchar', length: 64 })
+  keyHash: string;
+
+  @Column({ name: 'key_prefix', type: 'varchar', length: 12 })
+  keyPrefix: string;
+
+  @Column({ type: 'varchar', length: 100 })
+  name: string;
+
+  @Column({ type: 'varchar', length: 200, default: '' })
+  owner: string;
+
+  @Column({ type: 'jsonb', default: () => `'{"allowedModels": ["*"], "allowStreaming": true, "allowTools": true}'` })
+  permissions: {
+    allowedModels: string[];
+    allowStreaming: boolean;
+    allowTools: boolean;
+  };
+
+  @Column({ name: 'rate_limit_rpm', type: 'int', default: 60 })
+  rateLimitRpm: number;
+
+  @Column({ name: 'rate_limit_tpd', type: 'int', default: 1000000 })
+  rateLimitTpd: number;
+
+  @Column({ name: 'monthly_budget', type: 'decimal', precision: 10, scale: 2, nullable: true })
+  monthlyBudget: number | null;
+
+  @Column({ type: 'boolean', default: true })
+  enabled: boolean;
+
+  @Column({ name: 'expires_at', type: 'timestamp', nullable: true })
+  expiresAt: Date | null;
+
+  @Column({ name: 'last_used_at', type: 'timestamp', nullable: true })
+  lastUsedAt: Date | null;
+
+  @Column({ name: 'created_by', type: 'uuid', nullable: true })
+  createdBy: string | null;
+
+  @Column({ name: 'created_at', type: 'timestamp', default: () => 'NOW()' })
+  createdAt: Date;
+}
--- a/packages/services/conversation-service/src/infrastructure/database/postgres/entities/gateway-audit-log.orm.ts
+++ b/packages/services/conversation-service/src/infrastructure/database/postgres/entities/gateway-audit-log.orm.ts
@ -0,0 +1,47 @@
+import {
+  Entity,
+  PrimaryGeneratedColumn,
+  Column,
+  Index,
+} from 'typeorm';
+
+@Entity('gateway_audit_logs')
+@Index('idx_gateway_audit_logs_key', ['apiKeyId'])
+@Index('idx_gateway_audit_logs_created', ['createdAt'])
+export class GatewayAuditLogORM {
+  @PrimaryGeneratedColumn('uuid')
+  id: string;
+
+  @Column({ name: 'api_key_id', type: 'uuid' })
+  apiKeyId: string;
+
+  @Column({ name: 'request_method', type: 'varchar', length: 10 })
+  requestMethod: string;
+
+  @Column({ name: 'request_path', type: 'varchar', length: 200 })
+  requestPath: string;
+
+  @Column({ name: 'request_model', type: 'varchar', length: 100, nullable: true })
+  requestModel: string | null;
+
+  @Column({ name: 'request_ip', type: 'varchar', length: 50, default: '' })
+  requestIp: string;
+
+  @Column({ name: 'content_filtered', type: 'boolean', default: false })
+  contentFiltered: boolean;
+
+  @Column({ name: 'filter_rule_id', type: 'uuid', nullable: true })
+  filterRuleId: string | null;
+
+  @Column({ name: 'injection_applied', type: 'boolean', default: false })
+  injectionApplied: boolean;
+
+  @Column({ name: 'response_status', type: 'int', default: 200 })
+  responseStatus: number;
+
+  @Column({ name: 'duration_ms', type: 'int', default: 0 })
+  durationMs: number;
+
+  @Column({ name: 'created_at', type: 'timestamp', default: () => 'NOW()' })
+  createdAt: Date;
+}
--- a/packages/services/conversation-service/src/infrastructure/database/postgres/entities/gateway-content-rule.orm.ts
+++ b/packages/services/conversation-service/src/infrastructure/database/postgres/entities/gateway-content-rule.orm.ts
@ -0,0 +1,49 @@
+import {
+  Entity,
+  PrimaryGeneratedColumn,
+  Column,
+  CreateDateColumn,
+  UpdateDateColumn,
+  Index,
+} from 'typeorm';
+
+@Entity('gateway_content_rules')
+@Index('idx_gateway_content_rules_tenant', ['tenantId'])
+@Index('idx_gateway_content_rules_enabled', ['tenantId', 'enabled'])
+export class GatewayContentRuleORM {
+  @PrimaryGeneratedColumn('uuid')
+  id: string;
+
+  @Column({ name: 'tenant_id', type: 'uuid', nullable: true })
+  tenantId: string | null;
+
+  @Column({ type: 'varchar', length: 100 })
+  name: string;
+
+  @Column({ type: 'varchar', length: 20, default: 'keyword' })
+  type: 'keyword' | 'regex';
+
+  @Column({ type: 'text' })
+  pattern: string;
+
+  @Column({ type: 'varchar', length: 20, default: 'block' })
+  action: 'block' | 'warn' | 'log';
+
+  @Column({ name: 'reject_message', type: 'text', nullable: true })
+  rejectMessage: string | null;
+
+  @Column({ type: 'int', default: 0 })
+  priority: number;
+
+  @Column({ type: 'boolean', default: true })
+  enabled: boolean;
+
+  @Column({ name: 'created_by', type: 'uuid', nullable: true })
+  createdBy: string | null;
+
+  @CreateDateColumn({ name: 'created_at' })
+  createdAt: Date;
+
+  @UpdateDateColumn({ name: 'updated_at' })
+  updatedAt: Date;
+}
--- a/packages/services/conversation-service/src/infrastructure/database/postgres/entities/gateway-injection-rule.orm.ts
+++ b/packages/services/conversation-service/src/infrastructure/database/postgres/entities/gateway-injection-rule.orm.ts
@ -0,0 +1,52 @@
+import {
+  Entity,
+  PrimaryGeneratedColumn,
+  Column,
+  CreateDateColumn,
+  UpdateDateColumn,
+  Index,
+} from 'typeorm';
+
+@Entity('gateway_injection_rules')
+@Index('idx_gateway_injection_rules_tenant', ['tenantId'])
+@Index('idx_gateway_injection_rules_enabled', ['tenantId', 'enabled'])
+export class GatewayInjectionRuleORM {
+  @PrimaryGeneratedColumn('uuid')
+  id: string;
+
+  @Column({ name: 'tenant_id', type: 'uuid', nullable: true })
+  tenantId: string | null;
+
+  @Column({ type: 'varchar', length: 100 })
+  name: string;
+
+  @Column({ type: 'text', nullable: true })
+  description: string | null;
+
+  @Column({ type: 'varchar', length: 10, default: 'append' })
+  position: 'prepend' | 'append';
+
+  @Column({ type: 'text' })
+  content: string;
+
+  @Column({ name: 'match_models', type: 'jsonb', default: () => `'["*"]'` })
+  matchModels: string[];
+
+  @Column({ name: 'match_key_ids', type: 'jsonb', nullable: true })
+  matchKeyIds: string[] | null;
+
+  @Column({ type: 'int', default: 0 })
+  priority: number;
+
+  @Column({ type: 'boolean', default: true })
+  enabled: boolean;
+
+  @Column({ name: 'created_by', type: 'uuid', nullable: true })
+  createdBy: string | null;
+
+  @CreateDateColumn({ name: 'created_at' })
+  createdAt: Date;
+
+  @UpdateDateColumn({ name: 'updated_at' })
+  updatedAt: Date;
+}
--- a/packages/services/conversation-service/src/infrastructure/database/postgres/entities/gateway-usage-log.orm.ts
+++ b/packages/services/conversation-service/src/infrastructure/database/postgres/entities/gateway-usage-log.orm.ts
@ -0,0 +1,44 @@
+import {
+  Entity,
+  PrimaryGeneratedColumn,
+  Column,
+  Index,
+} from 'typeorm';
+
+@Entity('gateway_usage_logs')
+@Index('idx_gateway_usage_logs_key', ['apiKeyId'])
+@Index('idx_gateway_usage_logs_created', ['createdAt'])
+export class GatewayUsageLogORM {
+  @PrimaryGeneratedColumn('uuid')
+  id: string;
+
+  @Column({ name: 'api_key_id', type: 'uuid' })
+  apiKeyId: string;
+
+  @Column({ type: 'varchar', length: 100 })
+  model: string;
+
+  @Column({ type: 'varchar', length: 20 })
+  provider: string;
+
+  @Column({ name: 'input_tokens', type: 'int', default: 0 })
+  inputTokens: number;
+
+  @Column({ name: 'output_tokens', type: 'int', default: 0 })
+  outputTokens: number;
+
+  @Column({ name: 'total_tokens', type: 'int', default: 0 })
+  totalTokens: number;
+
+  @Column({ name: 'cost_usd', type: 'decimal', precision: 10, scale: 6, nullable: true })
+  costUsd: number | null;
+
+  @Column({ name: 'duration_ms', type: 'int', default: 0 })
+  durationMs: number;
+
+  @Column({ name: 'status_code', type: 'int', default: 200 })
+  statusCode: number;
+
+  @Column({ name: 'created_at', type: 'timestamp', default: () => 'NOW()' })
+  createdAt: Date;
+}
--- a/packages/services/llm-gateway/Dockerfile
+++ b/packages/services/llm-gateway/Dockerfile
@ -0,0 +1,57 @@
+# ===========================================
+# iConsulting LLM Gateway Dockerfile
+# ===========================================
+
+# 构建阶段
+FROM node:20-alpine AS builder
+
+WORKDIR /app
+
+# 安装 pnpm
+RUN corepack enable && corepack prepare pnpm@latest --activate
+
+# 复制 workspace 配置
+COPY pnpm-workspace.yaml package.json pnpm-lock.yaml ./
+COPY packages/services/llm-gateway/package.json ./packages/services/llm-gateway/
+
+# 安装依赖
+RUN pnpm install --frozen-lockfile --filter @iconsulting/llm-gateway
+
+# 复制源代码
+COPY packages/services/llm-gateway ./packages/services/llm-gateway
+
+# 构建服务
+RUN cd packages/services/llm-gateway && npx tsc
+
+# 运行阶段
+FROM node:20-alpine AS runner
+
+WORKDIR /app
+
+# 创建非 root 用户
+RUN addgroup --system --gid 1001 nodejs && \
+    adduser --system --uid 1001 gateway
+
+# 复制构建产物和依赖配置
+COPY --from=builder /app/packages/services/llm-gateway/dist ./dist
+COPY --from=builder /app/packages/services/llm-gateway/package.json ./
+
+# 安装生产依赖
+RUN npm install --omit=dev --ignore-scripts
+
+# 设置环境变量
+ENV NODE_ENV=production
+ENV PORT=3008
+
+# 切换用户
+USER gateway
+
+# 暴露端口
+EXPOSE 3008
+
+# 健康检查
+HEALTHCHECK --interval=30s --timeout=10s --start-period=10s --retries=3 \
+    CMD wget --no-verbose --tries=1 --spider http://localhost:3008/health || exit 1
+
+# 启动服务
+CMD ["node", "dist/main.js"]
--- a/packages/services/llm-gateway/package.json
+++ b/packages/services/llm-gateway/package.json
@ -0,0 +1,26 @@
+{
+  "name": "@iconsulting/llm-gateway",
+  "version": "1.0.0",
+  "description": "LLM API Gateway - Anthropic/OpenAI compatible proxy with regulatory injection and content filtering",
+  "main": "dist/main.js",
+  "scripts": {
+    "build": "tsc",
+    "dev": "tsx watch src/main.ts",
+    "start": "node dist/main.js",
+    "lint": "eslint \"src/**/*.ts\"",
+    "test": "jest",
+    "clean": "rm -rf dist"
+  },
+  "dependencies": {
+    "fastify": "^5.2.0",
+    "pg": "^8.11.0",
+    "uuid": "^9.0.0"
+  },
+  "devDependencies": {
+    "@types/node": "^20.10.0",
+    "@types/pg": "^8.10.0",
+    "@types/uuid": "^9.0.0",
+    "tsx": "^4.7.0",
+    "typescript": "^5.3.0"
+  }
+}
--- a/packages/services/llm-gateway/sql/init-gateway-tables.sql
+++ b/packages/services/llm-gateway/sql/init-gateway-tables.sql
@ -0,0 +1,104 @@
+-- ===========================================
+-- LLM Gateway Database Tables
+-- ===========================================
+
+-- 1. API Keys for external users
+CREATE TABLE IF NOT EXISTS gateway_api_keys (
+    id UUID PRIMARY KEY DEFAULT gen_random_uuid(),
+    tenant_id UUID,
+    key_hash VARCHAR(64) NOT NULL UNIQUE,
+    key_prefix VARCHAR(12) NOT NULL,
+    name VARCHAR(100) NOT NULL,
+    owner VARCHAR(200) NOT NULL DEFAULT '',
+    permissions JSONB NOT NULL DEFAULT '{"allowedModels": ["*"], "allowStreaming": true, "allowTools": true}',
+    rate_limit_rpm INTEGER NOT NULL DEFAULT 60,
+    rate_limit_tpd INTEGER NOT NULL DEFAULT 1000000,
+    monthly_budget DECIMAL(10,2),
+    enabled BOOLEAN NOT NULL DEFAULT true,
+    expires_at TIMESTAMP,
+    last_used_at TIMESTAMP,
+    created_by UUID,
+    created_at TIMESTAMP NOT NULL DEFAULT NOW()
+);
+
+CREATE INDEX IF NOT EXISTS idx_gateway_api_keys_hash ON gateway_api_keys (key_hash);
+CREATE INDEX IF NOT EXISTS idx_gateway_api_keys_tenant ON gateway_api_keys (tenant_id);
+CREATE INDEX IF NOT EXISTS idx_gateway_api_keys_enabled ON gateway_api_keys (enabled);
+
+-- 2. Injection rules (regulatory content to inject into system prompts)
+CREATE TABLE IF NOT EXISTS gateway_injection_rules (
+    id UUID PRIMARY KEY DEFAULT gen_random_uuid(),
+    tenant_id UUID,
+    name VARCHAR(100) NOT NULL,
+    description TEXT,
+    position VARCHAR(10) NOT NULL DEFAULT 'append' CHECK (position IN ('prepend', 'append')),
+    content TEXT NOT NULL,
+    match_models JSONB NOT NULL DEFAULT '["*"]',
+    match_key_ids JSONB,
+    priority INTEGER NOT NULL DEFAULT 0,
+    enabled BOOLEAN NOT NULL DEFAULT true,
+    created_by UUID,
+    created_at TIMESTAMP NOT NULL DEFAULT NOW(),
+    updated_at TIMESTAMP NOT NULL DEFAULT NOW()
+);
+
+CREATE INDEX IF NOT EXISTS idx_gateway_injection_rules_tenant ON gateway_injection_rules (tenant_id);
+CREATE INDEX IF NOT EXISTS idx_gateway_injection_rules_enabled ON gateway_injection_rules (tenant_id, enabled);
+
+-- 3. Content filter rules (message auditing)
+CREATE TABLE IF NOT EXISTS gateway_content_rules (
+    id UUID PRIMARY KEY DEFAULT gen_random_uuid(),
+    tenant_id UUID,
+    name VARCHAR(100) NOT NULL,
+    type VARCHAR(20) NOT NULL DEFAULT 'keyword' CHECK (type IN ('keyword', 'regex')),
+    pattern TEXT NOT NULL,
+    action VARCHAR(20) NOT NULL DEFAULT 'block' CHECK (action IN ('block', 'warn', 'log')),
+    reject_message TEXT,
+    priority INTEGER NOT NULL DEFAULT 0,
+    enabled BOOLEAN NOT NULL DEFAULT true,
+    created_by UUID,
+    created_at TIMESTAMP NOT NULL DEFAULT NOW(),
+    updated_at TIMESTAMP NOT NULL DEFAULT NOW()
+);
+
+CREATE INDEX IF NOT EXISTS idx_gateway_content_rules_tenant ON gateway_content_rules (tenant_id);
+CREATE INDEX IF NOT EXISTS idx_gateway_content_rules_enabled ON gateway_content_rules (tenant_id, enabled);
+
+-- 4. Usage logs (token consumption tracking)
+CREATE TABLE IF NOT EXISTS gateway_usage_logs (
+    id UUID PRIMARY KEY DEFAULT gen_random_uuid(),
+    api_key_id UUID NOT NULL REFERENCES gateway_api_keys(id) ON DELETE CASCADE,
+    model VARCHAR(100) NOT NULL,
+    provider VARCHAR(20) NOT NULL,
+    input_tokens INTEGER NOT NULL DEFAULT 0,
+    output_tokens INTEGER NOT NULL DEFAULT 0,
+    total_tokens INTEGER NOT NULL DEFAULT 0,
+    cost_usd DECIMAL(10,6),
+    duration_ms INTEGER NOT NULL DEFAULT 0,
+    status_code INTEGER NOT NULL DEFAULT 200,
+    created_at TIMESTAMP NOT NULL DEFAULT NOW()
+);
+
+CREATE INDEX IF NOT EXISTS idx_gateway_usage_logs_key ON gateway_usage_logs (api_key_id);
+CREATE INDEX IF NOT EXISTS idx_gateway_usage_logs_created ON gateway_usage_logs (created_at);
+CREATE INDEX IF NOT EXISTS idx_gateway_usage_logs_key_created ON gateway_usage_logs (api_key_id, created_at);
+
+-- 5. Audit logs (request/response auditing)
+CREATE TABLE IF NOT EXISTS gateway_audit_logs (
+    id UUID PRIMARY KEY DEFAULT gen_random_uuid(),
+    api_key_id UUID NOT NULL REFERENCES gateway_api_keys(id) ON DELETE CASCADE,
+    request_method VARCHAR(10) NOT NULL,
+    request_path VARCHAR(200) NOT NULL,
+    request_model VARCHAR(100),
+    request_ip VARCHAR(50) NOT NULL DEFAULT '',
+    content_filtered BOOLEAN NOT NULL DEFAULT false,
+    filter_rule_id UUID,
+    injection_applied BOOLEAN NOT NULL DEFAULT false,
+    response_status INTEGER NOT NULL DEFAULT 200,
+    duration_ms INTEGER NOT NULL DEFAULT 0,
+    created_at TIMESTAMP NOT NULL DEFAULT NOW()
+);
+
+CREATE INDEX IF NOT EXISTS idx_gateway_audit_logs_key ON gateway_audit_logs (api_key_id);
+CREATE INDEX IF NOT EXISTS idx_gateway_audit_logs_created ON gateway_audit_logs (created_at);
+CREATE INDEX IF NOT EXISTS idx_gateway_audit_logs_filtered ON gateway_audit_logs (content_filtered) WHERE content_filtered = true;
--- a/packages/services/llm-gateway/src/config.ts
+++ b/packages/services/llm-gateway/src/config.ts
@ -0,0 +1,29 @@
+export interface GatewayConfig {
+  port: number;
+  databaseUrl: string;
+  anthropicApiKey: string;
+  anthropicUpstreamUrl: string;
+  openaiApiKey: string;
+  openaiUpstreamUrl: string;
+  rulesCacheTtlMs: number;
+  logLevel: string;
+}
+
+export function loadConfig(): GatewayConfig {
+  const required = (key: string): string => {
+    const val = process.env[key];
+    if (!val) throw new Error(`Missing required env var: ${key}`);
+    return val;
+  };
+
+  return {
+    port: parseInt(process.env.PORT || '3008', 10),
+    databaseUrl: required('DATABASE_URL'),
+    anthropicApiKey: required('ANTHROPIC_API_KEY'),
+    anthropicUpstreamUrl: (process.env.ANTHROPIC_UPSTREAM_URL || 'https://api.anthropic.com').replace(/\/$/, ''),
+    openaiApiKey: required('OPENAI_API_KEY'),
+    openaiUpstreamUrl: (process.env.OPENAI_UPSTREAM_URL || 'https://api.openai.com').replace(/\/$/, ''),
+    rulesCacheTtlMs: parseInt(process.env.RULES_CACHE_TTL_MS || '30000', 10),
+    logLevel: process.env.LOG_LEVEL || 'info',
+  };
+}
--- a/packages/services/llm-gateway/src/db.ts
+++ b/packages/services/llm-gateway/src/db.ts
@ -0,0 +1,43 @@
+import { Pool, PoolClient } from 'pg';
+import { GatewayConfig } from './config';
+
+let pool: Pool;
+
+export function initDb(config: GatewayConfig): Pool {
+  pool = new Pool({
+    connectionString: config.databaseUrl,
+    max: 20,
+    idleTimeoutMillis: 30000,
+    connectionTimeoutMillis: 5000,
+  });
+
+  pool.on('error', (err) => {
+    console.error('[DB] Unexpected error on idle client', err);
+  });
+
+  return pool;
+}
+
+export function getPool(): Pool {
+  if (!pool) throw new Error('Database not initialized');
+  return pool;
+}
+
+export async function query<T = any>(text: string, params?: any[]): Promise<T[]> {
+  const result = await getPool().query(text, params);
+  return result.rows as T[];
+}
+
+export async function queryOne<T = any>(text: string, params?: any[]): Promise<T | null> {
+  const rows = await query<T>(text, params);
+  return rows[0] || null;
+}
+
+export async function withClient<T>(fn: (client: PoolClient) => Promise<T>): Promise<T> {
+  const client = await getPool().connect();
+  try {
+    return await fn(client);
+  } finally {
+    client.release();
+  }
+}
--- a/packages/services/llm-gateway/src/injection/system-prompt-injector.ts
+++ b/packages/services/llm-gateway/src/injection/system-prompt-injector.ts
@ -0,0 +1,119 @@
+import { query } from '../db';
+import { InjectionRule, AnthropicSystem, AnthropicSystemBlock } from '../types';
+
+// ─── Cache ───
+
+let cachedRules: InjectionRule[] = [];
+let cacheTimestamp = 0;
+let cacheTtlMs = 30_000;
+
+export function setInjectionCacheTtl(ttlMs: number): void {
+  cacheTtlMs = ttlMs;
+}
+
+async function loadRules(): Promise<InjectionRule[]> {
+  if (Date.now() - cacheTimestamp < cacheTtlMs && cachedRules.length > 0) {
+    return cachedRules;
+  }
+
+  const rows = await query<any>(
+    `SELECT id, tenant_id, name, position, content, match_models, match_key_ids, priority, enabled
+     FROM gateway_injection_rules
+     WHERE enabled = true
+     ORDER BY priority ASC`,
+  );
+
+  cachedRules = rows.map((row) => ({
+    id: row.id,
+    tenantId: row.tenant_id,
+    name: row.name,
+    position: row.position,
+    content: row.content,
+    matchModels: row.match_models || ['*'],
+    matchKeyIds: row.match_key_ids,
+    priority: row.priority,
+    enabled: row.enabled,
+  }));
+
+  cacheTimestamp = Date.now();
+  return cachedRules;
+}
+
+// ─── Match logic ───
+
+function matchesModel(ruleModels: string[], requestModel: string): boolean {
+  if (ruleModels.includes('*')) return true;
+  return ruleModels.some((pattern) => {
+    if (pattern.endsWith('*')) {
+      return requestModel.startsWith(pattern.slice(0, -1));
+    }
+    return requestModel === pattern;
+  });
+}
+
+function matchesKey(ruleKeyIds: string[] | null, apiKeyId: string): boolean {
+  if (!ruleKeyIds) return true; // null = matches all keys
+  return ruleKeyIds.includes(apiKeyId);
+}
+
+// ─── Inject regulatory content into system prompt ───
+
+export async function injectSystemPrompt(
+  system: AnthropicSystem | undefined,
+  model: string,
+  apiKeyId: string,
+): Promise<{ system: AnthropicSystem | undefined; applied: boolean; ruleIds: string[] }> {
+  const allRules = await loadRules();
+
+  const applicableRules = allRules.filter(
+    (r) => matchesModel(r.matchModels, model) && matchesKey(r.matchKeyIds, apiKeyId),
+  );
+
+  if (applicableRules.length === 0) {
+    return { system, applied: false, ruleIds: [] };
+  }
+
+  const prependRules = applicableRules.filter((r) => r.position === 'prepend');
+  const appendRules = applicableRules.filter((r) => r.position === 'append');
+
+  const prependText = prependRules.map((r) => r.content).join('\n\n');
+  const appendText = appendRules.map((r) => r.content).join('\n\n');
+  const ruleIds = applicableRules.map((r) => r.id);
+
+  if (!prependText && !appendText) {
+    return { system, applied: false, ruleIds: [] };
+  }
+
+  // Handle the three possible system formats
+  if (system === undefined || system === null) {
+    // No existing system prompt — create one
+    const combined = [prependText, appendText].filter(Boolean).join('\n\n');
+    return { system: combined || undefined, applied: true, ruleIds };
+  }
+
+  if (typeof system === 'string') {
+    // String format
+    const parts = [prependText, system, appendText].filter(Boolean);
+    return { system: parts.join('\n\n'), applied: true, ruleIds };
+  }
+
+  if (Array.isArray(system)) {
+    // Array of content blocks format
+    const blocks: AnthropicSystemBlock[] = [...system];
+    if (prependText) {
+      blocks.unshift({ type: 'text', text: prependText });
+    }
+    if (appendText) {
+      blocks.push({ type: 'text', text: appendText });
+    }
+    return { system: blocks, applied: true, ruleIds };
+  }
+
+  // Fallback — unknown format, don't modify
+  return { system, applied: false, ruleIds: [] };
+}
+
+export function clearInjectionRulesCache(): void {
+  cachedRules = [];
+  cacheTimestamp = 0;
+}
--- a/packages/services/llm-gateway/src/logging/audit-logger.ts
+++ b/packages/services/llm-gateway/src/logging/audit-logger.ts
@ -0,0 +1,79 @@
+import { query } from '../db';
+import { v4 as uuidv4 } from 'uuid';
+import { AuditLogEntry } from '../types';
+
+// ─── Batch queue (async, non-blocking) ───
+
+const pendingAudits: AuditLogEntry[] = [];
+let flushTimer: ReturnType<typeof setTimeout> | null = null;
+const FLUSH_INTERVAL_MS = 5_000;
+const BATCH_SIZE = 50;
+
+async function flushAudits(): Promise<void> {
+  if (pendingAudits.length === 0) return;
+
+  const batch = pendingAudits.splice(0, BATCH_SIZE);
+
+  try {
+    const values: any[] = [];
+    const placeholders: string[] = [];
+
+    for (let i = 0; i < batch.length; i++) {
+      const entry = batch[i];
+      const offset = i * 11;
+      placeholders.push(
+        `($${offset + 1}, $${offset + 2}, $${offset + 3}, $${offset + 4}, $${offset + 5}, $${offset + 6}, $${offset + 7}, $${offset + 8}, $${offset + 9}, $${offset + 10}, $${offset + 11})`,
+      );
+      values.push(
+        uuidv4(),
+        entry.apiKeyId,
+        entry.requestMethod,
+        entry.requestPath,
+        entry.requestModel,
+        entry.requestIp,
+        entry.contentFiltered,
+        entry.filterRuleId,
+        entry.injectionApplied,
+        entry.responseStatus,
+        entry.durationMs,
+      );
+    }
+
+    await query(
+      `INSERT INTO gateway_audit_logs (id, api_key_id, request_method, request_path, request_model, request_ip, content_filtered, filter_rule_id, injection_applied, response_status, duration_ms)
+       VALUES ${placeholders.join(', ')}`,
+      values,
+    );
+  } catch (err: any) {
+    console.error('[AuditLogger] Failed to flush audits:', err.message);
+    if (pendingAudits.length < 500) {
+      pendingAudits.unshift(...batch);
+    }
+  }
+}
+
+function scheduleFlush(): void {
+  if (flushTimer) return;
+  flushTimer = setTimeout(async () => {
+    flushTimer = null;
+    await flushAudits();
+    if (pendingAudits.length > 0) scheduleFlush();
+  }, FLUSH_INTERVAL_MS);
+}
+
+// ─── Public API ───
+
+export function recordAudit(entry: AuditLogEntry): void {
+  pendingAudits.push(entry);
+
+  if (pendingAudits.length >= BATCH_SIZE) {
+    flushAudits();
+  } else {
+    scheduleFlush();
+  }
+}
+
+export async function shutdown(): Promise<void> {
+  if (flushTimer) clearTimeout(flushTimer);
+  await flushAudits();
+}
--- a/packages/services/llm-gateway/src/logging/usage-tracker.ts
+++ b/packages/services/llm-gateway/src/logging/usage-tracker.ts
@ -0,0 +1,142 @@
+import { query } from '../db';
+import { v4 as uuidv4 } from 'uuid';
+import { UsageLogEntry } from '../types';
+
+// ─── Cost estimation (USD per million tokens, approximate) ───
+
+const COST_TABLE: Record<string, { input: number; output: number }> = {
+  'claude-sonnet-4-5': { input: 3.0, output: 15.0 },
+  'claude-haiku-4-5': { input: 0.8, output: 4.0 },
+  'claude-opus-4': { input: 15.0, output: 75.0 },
+  'text-embedding-3-small': { input: 0.02, output: 0 },
+  'text-embedding-3-large': { input: 0.13, output: 0 },
+};
+
+function estimateCost(model: string, inputTokens: number, outputTokens: number): number | null {
+  // Find matching cost entry by prefix
+  for (const [prefix, costs] of Object.entries(COST_TABLE)) {
+    if (model.startsWith(prefix)) {
+      return (inputTokens * costs.input + outputTokens * costs.output) / 1_000_000;
+    }
+  }
+  return null;
+}
+
+// ─── Batch insert queue (async, non-blocking) ───
+
+const pendingLogs: UsageLogEntry[] = [];
+let flushTimer: ReturnType<typeof setTimeout> | null = null;
+const FLUSH_INTERVAL_MS = 5_000;
+const BATCH_SIZE = 50;
+
+async function flushLogs(): Promise<void> {
+  if (pendingLogs.length === 0) return;
+
+  const batch = pendingLogs.splice(0, BATCH_SIZE);
+
+  try {
+    const values: any[] = [];
+    const placeholders: string[] = [];
+
+    for (let i = 0; i < batch.length; i++) {
+      const entry = batch[i];
+      const offset = i * 10;
+      placeholders.push(
+        `($${offset + 1}, $${offset + 2}, $${offset + 3}, $${offset + 4}, $${offset + 5}, $${offset + 6}, $${offset + 7}, $${offset + 8}, $${offset + 9}, $${offset + 10})`,
+      );
+      values.push(
+        uuidv4(),
+        entry.apiKeyId,
+        entry.model,
+        entry.provider,
+        entry.inputTokens,
+        entry.outputTokens,
+        entry.totalTokens,
+        entry.costUsd,
+        entry.durationMs,
+        entry.statusCode,
+      );
+    }
+
+    await query(
+      `INSERT INTO gateway_usage_logs (id, api_key_id, model, provider, input_tokens, output_tokens, total_tokens, cost_usd, duration_ms, status_code)
+       VALUES ${placeholders.join(', ')}`,
+      values,
+    );
+  } catch (err: any) {
+    console.error('[UsageTracker] Failed to flush logs:', err.message);
+    // Re-queue failed entries (at the front, max 500 pending)
+    if (pendingLogs.length < 500) {
+      pendingLogs.unshift(...batch);
+    }
+  }
+}
+
+function scheduleFlush(): void {
+  if (flushTimer) return;
+  flushTimer = setTimeout(async () => {
+    flushTimer = null;
+    await flushLogs();
+    if (pendingLogs.length > 0) scheduleFlush();
+  }, FLUSH_INTERVAL_MS);
+}
+
+// ─── Public API ───
+
+export function recordUsage(entry: UsageLogEntry): void {
+  if (!entry.costUsd && entry.costUsd !== 0) {
+    entry.costUsd = estimateCost(entry.model, entry.inputTokens, entry.outputTokens);
+  }
+  pendingLogs.push(entry);
+
+  if (pendingLogs.length >= BATCH_SIZE) {
+    flushLogs();
+  } else {
+    scheduleFlush();
+  }
+}
+
+export function recordFromAnthropicResponse(
+  apiKeyId: string,
+  model: string,
+  usage: { input_tokens?: number; output_tokens?: number } | undefined,
+  statusCode: number,
+  durationMs: number,
+): void {
+  recordUsage({
+    apiKeyId,
+    model,
+    provider: 'anthropic',
+    inputTokens: usage?.input_tokens || 0,
+    outputTokens: usage?.output_tokens || 0,
+    totalTokens: (usage?.input_tokens || 0) + (usage?.output_tokens || 0),
+    costUsd: null,
+    durationMs,
+    statusCode,
+  });
+}
+
+export function recordFromOpenAIResponse(
+  apiKeyId: string,
+  model: string,
+  usage: { prompt_tokens?: number; total_tokens?: number } | undefined,
+  statusCode: number,
+  durationMs: number,
+): void {
+  recordUsage({
+    apiKeyId,
+    model,
+    provider: 'openai',
+    inputTokens: usage?.prompt_tokens || 0,
+    outputTokens: 0,
+    totalTokens: usage?.total_tokens || 0,
+    costUsd: null,
+    durationMs,
+    statusCode,
+  });
+}
+
+export async function shutdown(): Promise<void> {
+  if (flushTimer) clearTimeout(flushTimer);
+  await flushLogs();
+}
--- a/packages/services/llm-gateway/src/main.ts
+++ b/packages/services/llm-gateway/src/main.ts
@ -0,0 +1,73 @@
+import Fastify from 'fastify';
+import { loadConfig } from './config';
+import { initDb, getPool } from './db';
+import { registerHealthRoutes } from './routes/health';
+import { registerAnthropicRoutes } from './routes/anthropic';
+import { registerOpenAIRoutes } from './routes/openai';
+import { setContentFilterCacheTtl } from './middleware/content-filter';
+import { setInjectionCacheTtl } from './injection/system-prompt-injector';
+import { shutdown as shutdownUsageTracker } from './logging/usage-tracker';
+import { shutdown as shutdownAuditLogger } from './logging/audit-logger';
+
+async function main() {
+  // Load configuration
+  const config = loadConfig();
+
+  // Initialize database
+  initDb(config);
+
+  // Set cache TTLs
+  setContentFilterCacheTtl(config.rulesCacheTtlMs);
+  setInjectionCacheTtl(config.rulesCacheTtlMs);
+
+  // Create Fastify instance
+  const app = Fastify({
+    logger: {
+      level: config.logLevel,
+    },
+    bodyLimit: 50 * 1024 * 1024, // 50MB for base64 PDFs
+    trustProxy: true,
+  });
+
+  // Disable Fastify's default JSON body parsing — we parse manually
+  // to avoid double-parse overhead and to handle raw body forwarding
+  app.addContentTypeParser(
+    'application/json',
+    { parseAs: 'string' },
+    (_req, body, done) => {
+      done(null, body);
+    },
+  );
+
+  // Register routes
+  registerHealthRoutes(app);
+  registerAnthropicRoutes(app, config);
+  registerOpenAIRoutes(app, config);
+
+  // Graceful shutdown
+  const shutdown = async (signal: string) => {
+    console.log(`[LLM Gateway] Received ${signal}, shutting down...`);
+    await shutdownUsageTracker();
+    await shutdownAuditLogger();
+    await app.close();
+    const pool = getPool();
+    await pool.end();
+    process.exit(0);
+  };
+
+  process.on('SIGTERM', () => shutdown('SIGTERM'));
+  process.on('SIGINT', () => shutdown('SIGINT'));
+
+  // Start server
+  try {
+    await app.listen({ port: config.port, host: '0.0.0.0' });
+    console.log(`[LLM Gateway] Running on port ${config.port}`);
+    console.log(`[LLM Gateway] Anthropic upstream: ${config.anthropicUpstreamUrl}`);
+    console.log(`[LLM Gateway] OpenAI upstream: ${config.openaiUpstreamUrl}`);
+  } catch (err) {
+    console.error('[LLM Gateway] Failed to start:', err);
+    process.exit(1);
+  }
+}
+
+main();
--- a/packages/services/llm-gateway/src/middleware/auth.ts
+++ b/packages/services/llm-gateway/src/middleware/auth.ts
@ -0,0 +1,213 @@
+import { createHash } from 'crypto';
+import { FastifyRequest, FastifyReply } from 'fastify';
+import { query, queryOne } from '../db';
+import { ApiKeyRecord, ApiKeyPermissions } from '../types';
+
+// ─── Cache ───
+
+interface CachedKey {
+  record: ApiKeyRecord;
+  cachedAt: number;
+}
+
+const keyCache = new Map<string, CachedKey>();
+const KEY_CACHE_TTL_MS = 60_000; // 1 minute
+
+// ─── Hash ───
+
+function hashApiKey(key: string): string {
+  return createHash('sha256').update(key).digest('hex');
+}
+
+// ─── Extract API Key from request ───
+// Supports both Anthropic (X-Api-Key) and OpenAI (Authorization: Bearer) formats
+
+function extractApiKey(request: FastifyRequest): string | null {
+  // Anthropic style
+  const xApiKey = request.headers['x-api-key'];
+  if (typeof xApiKey === 'string' && xApiKey) return xApiKey;
+
+  // OpenAI style
+  const auth = request.headers['authorization'];
+  if (typeof auth === 'string' && auth.startsWith('Bearer ')) {
+    return auth.slice(7);
+  }
+
+  return null;
+}
+
+// ─── Lookup API Key record ───
+
+async function lookupApiKey(apiKey: string): Promise<ApiKeyRecord | null> {
+  const hash = hashApiKey(apiKey);
+
+  // Check cache
+  const cached = keyCache.get(hash);
+  if (cached && Date.now() - cached.cachedAt < KEY_CACHE_TTL_MS) {
+    return cached.record;
+  }
+
+  const row = await queryOne<any>(
+    `SELECT id, tenant_id, key_hash, key_prefix, name, owner,
+            permissions, rate_limit_rpm, rate_limit_tpd, monthly_budget,
+            enabled, expires_at, last_used_at
+     FROM gateway_api_keys
+     WHERE key_hash = $1`,
+    [hash],
+  );
+
+  if (!row) return null;
+
+  const record: ApiKeyRecord = {
+    id: row.id,
+    tenantId: row.tenant_id,
+    keyHash: row.key_hash,
+    keyPrefix: row.key_prefix,
+    name: row.name,
+    owner: row.owner,
+    permissions: row.permissions as ApiKeyPermissions,
+    rateLimitRpm: row.rate_limit_rpm,
+    rateLimitTpd: row.rate_limit_tpd,
+    monthlyBudget: row.monthly_budget ? parseFloat(row.monthly_budget) : null,
+    enabled: row.enabled,
+    expiresAt: row.expires_at ? new Date(row.expires_at) : null,
+    lastUsedAt: row.last_used_at ? new Date(row.last_used_at) : null,
+  };
+
+  keyCache.set(hash, { record, cachedAt: Date.now() });
+  return record;
+}
+
+// ─── Rate limit check (simple in-memory sliding window) ───
+
+interface RateWindow {
+  timestamps: number[];
+}
+
+const rateWindows = new Map<string, RateWindow>();
+
+function checkRateLimit(keyId: string, limitRpm: number): boolean {
+  if (limitRpm <= 0) return true; // no limit
+
+  const now = Date.now();
+  const windowStart = now - 60_000;
+
+  let window = rateWindows.get(keyId);
+  if (!window) {
+    window = { timestamps: [] };
+    rateWindows.set(keyId, window);
+  }
+
+  // Remove expired entries
+  window.timestamps = window.timestamps.filter((t) => t > windowStart);
+
+  if (window.timestamps.length >= limitRpm) {
+    return false; // rate limited
+  }
+
+  window.timestamps.push(now);
+  return true;
+}
+
+// Periodic cleanup of stale rate windows
+setInterval(() => {
+  const cutoff = Date.now() - 120_000;
+  for (const [key, window] of rateWindows.entries()) {
+    if (window.timestamps.length === 0 || window.timestamps[window.timestamps.length - 1] < cutoff) {
+      rateWindows.delete(key);
+    }
+  }
+}, 60_000);
+
+// ─── Update last_used_at (async, fire-and-forget) ───
+
+function touchLastUsed(keyId: string): void {
+  query('UPDATE gateway_api_keys SET last_used_at = NOW() WHERE id = $1', [keyId]).catch((err) => {
+    console.error('[Auth] Failed to update last_used_at:', err.message);
+  });
+}
+
+// ─── Auth middleware ───
+
+export async function authMiddleware(
+  request: FastifyRequest,
+  reply: FastifyReply,
+): Promise<void> {
+  const apiKey = extractApiKey(request);
+
+  if (!apiKey) {
+    reply.status(401).send({
+      error: {
+        type: 'authentication_error',
+        message: 'Missing API key. Provide via X-Api-Key header or Authorization: Bearer header.',
+      },
+    });
+    return;
+  }
+
+  const record = await lookupApiKey(apiKey);
+
+  if (!record) {
+    reply.status(401).send({
+      error: {
+        type: 'authentication_error',
+        message: 'Invalid API key.',
+      },
+    });
+    return;
+  }
+
+  if (!record.enabled) {
+    reply.status(403).send({
+      error: {
+        type: 'permission_error',
+        message: 'API key is disabled.',
+      },
+    });
+    return;
+  }
+
+  if (record.expiresAt && record.expiresAt < new Date()) {
+    reply.status(403).send({
+      error: {
+        type: 'permission_error',
+        message: 'API key has expired.',
+      },
+    });
+    return;
+  }
+
+  if (!checkRateLimit(record.id, record.rateLimitRpm)) {
+    reply.status(429).send({
+      error: {
+        type: 'rate_limit_error',
+        message: `Rate limit exceeded. Maximum ${record.rateLimitRpm} requests per minute.`,
+      },
+    });
+    return;
+  }
+
+  // Attach to request for downstream use
+  (request as any).apiKeyRecord = record;
+
+  // Fire-and-forget update
+  touchLastUsed(record.id);
+}
+
+// ─── Model permission check ───
+
+export function isModelAllowed(record: ApiKeyRecord, model: string): boolean {
+  const allowed = record.permissions?.allowedModels;
+  if (!allowed || allowed.length === 0 || allowed.includes('*')) return true;
+
+  return allowed.some((pattern) => {
+    if (pattern.endsWith('*')) {
+      return model.startsWith(pattern.slice(0, -1));
+    }
+    return model === pattern;
+  });
+}
+
+export function clearKeyCache(): void {
+  keyCache.clear();
+}
--- a/packages/services/llm-gateway/src/middleware/content-filter.ts
+++ b/packages/services/llm-gateway/src/middleware/content-filter.ts
@ -0,0 +1,116 @@
+import { query } from '../db';
+import { ContentRule, FilterResult, AnthropicMessage, AnthropicContentBlock } from '../types';
+
+// ─── Cache ───
+
+let cachedRules: ContentRule[] = [];
+let cacheTimestamp = 0;
+let cacheTtlMs = 30_000;
+
+export function setContentFilterCacheTtl(ttlMs: number): void {
+  cacheTtlMs = ttlMs;
+}
+
+async function loadRules(): Promise<ContentRule[]> {
+  if (Date.now() - cacheTimestamp < cacheTtlMs && cachedRules.length > 0) {
+    return cachedRules;
+  }
+
+  const rows = await query<any>(
+    `SELECT id, tenant_id, name, type, pattern, action, reject_message, priority, enabled
+     FROM gateway_content_rules
+     WHERE enabled = true
+     ORDER BY priority ASC`,
+  );
+
+  cachedRules = rows.map((row) => ({
+    id: row.id,
+    tenantId: row.tenant_id,
+    name: row.name,
+    type: row.type,
+    pattern: row.pattern,
+    action: row.action,
+    rejectMessage: row.reject_message,
+    priority: row.priority,
+    enabled: row.enabled,
+  }));
+
+  cacheTimestamp = Date.now();
+  return cachedRules;
+}
+
+// ─── Extract text from messages ───
+
+function extractTextFromContent(content: string | AnthropicContentBlock[]): string {
+  if (typeof content === 'string') return content;
+  if (!Array.isArray(content)) return '';
+
+  return content
+    .filter((block) => block.type === 'text' && block.text)
+    .map((block) => block.text!)
+    .join(' ');
+}
+
+// ─── Check messages against rules ───
+
+export async function checkContent(messages: AnthropicMessage[]): Promise<FilterResult> {
+  const rules = await loadRules();
+  if (rules.length === 0) return { blocked: false };
+
+  for (const msg of messages) {
+    if (msg.role !== 'user') continue;
+
+    const text = extractTextFromContent(msg.content);
+    if (!text) continue;
+
+    for (const rule of rules) {
+      let matched = false;
+
+      if (rule.type === 'keyword') {
+        matched = text.toLowerCase().includes(rule.pattern.toLowerCase());
+      } else if (rule.type === 'regex') {
+        try {
+          matched = new RegExp(rule.pattern, 'i').test(text);
+        } catch {
+          console.warn(`[ContentFilter] Invalid regex pattern in rule ${rule.id}: ${rule.pattern}`);
+          continue;
+        }
+      }
+
+      if (matched) {
+        if (rule.action === 'block') {
+          return {
+            blocked: true,
+            reason: rule.rejectMessage || 'Content blocked by policy.',
+            ruleId: rule.id,
+            action: 'block',
+          };
+        }
+        if (rule.action === 'warn') {
+          // Log but allow through
+          console.warn(`[ContentFilter] Warning triggered by rule "${rule.name}" (${rule.id})`);
+          return {
+            blocked: false,
+            ruleId: rule.id,
+            action: 'warn',
+          };
+        }
+        if (rule.action === 'log') {
+          console.info(`[ContentFilter] Log match by rule "${rule.name}" (${rule.id})`);
+          return {
+            blocked: false,
+            ruleId: rule.id,
+            action: 'log',
+          };
+        }
+      }
+    }
+  }
+
+  return { blocked: false };
+}
+
+export function clearContentRulesCache(): void {
+  cachedRules = [];
+  cacheTimestamp = 0;
+}
--- a/packages/services/llm-gateway/src/proxy/anthropic-proxy.ts
+++ b/packages/services/llm-gateway/src/proxy/anthropic-proxy.ts
@ -0,0 +1,205 @@
+import { FastifyRequest, FastifyReply } from 'fastify';
+import { GatewayConfig } from '../config';
+import { ApiKeyRecord, AnthropicRequestBody, FilterResult } from '../types';
+import { isModelAllowed } from '../middleware/auth';
+import { checkContent } from '../middleware/content-filter';
+import { injectSystemPrompt } from '../injection/system-prompt-injector';
+import { recordFromAnthropicResponse } from '../logging/usage-tracker';
+import { recordAudit } from '../logging/audit-logger';
+import { pipeSSEStream, createStreamUsageTracker } from './stream-pipe';
+
+export function createAnthropicProxy(config: GatewayConfig) {
+  return async function handleMessages(request: FastifyRequest, reply: FastifyReply): Promise<void> {
+    const startTime = Date.now();
+    const apiKeyRecord: ApiKeyRecord = (request as any).apiKeyRecord;
+    const clientIp = request.ip || request.headers['x-forwarded-for'] as string || 'unknown';
+
+    // 1. Parse request body
+    let body: AnthropicRequestBody;
+    try {
+      body = JSON.parse(request.body as string);
+    } catch {
+      reply.status(400).send({
+        error: { type: 'invalid_request_error', message: 'Invalid JSON body.' },
+      });
+      return;
+    }
+
+    const model = body.model || 'unknown';
+
+    // 2. Check model permission
+    if (!isModelAllowed(apiKeyRecord, model)) {
+      reply.status(403).send({
+        error: {
+          type: 'permission_error',
+          message: `Model "${model}" is not allowed for this API key.`,
+        },
+      });
+      return;
+    }
+
+    // 3. Check streaming permission
+    if (body.stream && apiKeyRecord.permissions?.allowStreaming === false) {
+      reply.status(403).send({
+        error: {
+          type: 'permission_error',
+          message: 'Streaming is not allowed for this API key.',
+        },
+      });
+      return;
+    }
+
+    // 4. Content filtering
+    let filterResult: FilterResult = { blocked: false };
+    if (body.messages) {
+      filterResult = await checkContent(body.messages);
+      if (filterResult.blocked) {
+        recordAudit({
+          apiKeyId: apiKeyRecord.id,
+          requestMethod: 'POST',
+          requestPath: '/v1/messages',
+          requestModel: model,
+          requestIp: clientIp,
+          contentFiltered: true,
+          filterRuleId: filterResult.ruleId || null,
+          injectionApplied: false,
+          responseStatus: 403,
+          durationMs: Date.now() - startTime,
+        });
+
+        reply.status(403).send({
+          error: {
+            type: 'content_policy_violation',
+            message: filterResult.reason || 'Content blocked by policy.',
+          },
+        });
+        return;
+      }
+    }
+
+    // 5. Inject regulatory content into system prompt
+    const injection = await injectSystemPrompt(body.system, model, apiKeyRecord.id);
+    body.system = injection.system;
+
+    // 6. Build upstream request headers
+    const upstreamHeaders: Record<string, string> = {
+      'Content-Type': 'application/json',
+      'anthropic-version': (request.headers['anthropic-version'] as string) || '2023-06-01',
+      'X-Api-Key': config.anthropicApiKey,
+    };
+
+    // Forward optional headers
+    const beta = request.headers['anthropic-beta'];
+    if (beta) upstreamHeaders['anthropic-beta'] = beta as string;
+
+    // 7. Forward to upstream
+    let upstreamResponse: Response;
+    try {
+      upstreamResponse = await fetch(`${config.anthropicUpstreamUrl}/v1/messages`, {
+        method: 'POST',
+        headers: upstreamHeaders,
+        body: JSON.stringify(body),
+      });
+    } catch (err: any) {
+      const durationMs = Date.now() - startTime;
+      recordAudit({
+        apiKeyId: apiKeyRecord.id,
+        requestMethod: 'POST',
+        requestPath: '/v1/messages',
+        requestModel: model,
+        requestIp: clientIp,
+        contentFiltered: filterResult.action === 'warn' || filterResult.action === 'log',
+        filterRuleId: filterResult.ruleId || null,
+        injectionApplied: injection.applied,
+        responseStatus: 502,
+        durationMs,
+      });
+
+      reply.status(502).send({
+        error: {
+          type: 'upstream_error',
+          message: `Failed to connect to upstream API: ${err.message}`,
+        },
+      });
+      return;
+    }
+
+    const durationMs = Date.now() - startTime;
+
+    // 8. Handle response
+    if (body.stream && upstreamResponse.ok && upstreamResponse.body) {
+      // Streaming response — pipe SSE directly
+      reply.raw.writeHead(upstreamResponse.status, {
+        'Content-Type': 'text/event-stream',
+        'Cache-Control': 'no-cache',
+        'Connection': 'keep-alive',
+        'X-Accel-Buffering': 'no',
+      });
+
+      const usageTracker = createStreamUsageTracker();
+
+      await pipeSSEStream(upstreamResponse.body, reply.raw, usageTracker.onDataLine);
+
+      // Record usage from stream (async)
+      const streamUsage = usageTracker.getUsage();
+      recordFromAnthropicResponse(
+        apiKeyRecord.id,
+        model,
+        { input_tokens: streamUsage.inputTokens, output_tokens: streamUsage.outputTokens },
+        upstreamResponse.status,
+        Date.now() - startTime,
+      );
+
+      recordAudit({
+        apiKeyId: apiKeyRecord.id,
+        requestMethod: 'POST',
+        requestPath: '/v1/messages',
+        requestModel: model,
+        requestIp: clientIp,
+        contentFiltered: filterResult.action === 'warn' || filterResult.action === 'log',
+        filterRuleId: filterResult.ruleId || null,
+        injectionApplied: injection.applied,
+        responseStatus: upstreamResponse.status,
+        durationMs: Date.now() - startTime,
+      });
+    } else {
+      // Non-streaming response — buffer and forward
+      const responseText = await upstreamResponse.text();
+
+      // Try to extract usage for logging
+      try {
+        const responseJson = JSON.parse(responseText);
+        if (responseJson.usage) {
+          recordFromAnthropicResponse(
+            apiKeyRecord.id,
+            model,
+            responseJson.usage,
+            upstreamResponse.status,
+            durationMs,
+          );
+        }
+      } catch {
+        // Not JSON — still forward
+      }
+
+      recordAudit({
+        apiKeyId: apiKeyRecord.id,
+        requestMethod: 'POST',
+        requestPath: '/v1/messages',
+        requestModel: model,
+        requestIp: clientIp,
+        contentFiltered: filterResult.action === 'warn' || filterResult.action === 'log',
+        filterRuleId: filterResult.ruleId || null,
+        injectionApplied: injection.applied,
+        responseStatus: upstreamResponse.status,
+        durationMs,
+      });
+
+      // Forward all response headers from upstream
+      reply.raw.writeHead(upstreamResponse.status, {
+        'Content-Type': upstreamResponse.headers.get('content-type') || 'application/json',
+      });
+      reply.raw.end(responseText);
+    }
+  };
+}
--- a/packages/services/llm-gateway/src/proxy/openai-proxy.ts
+++ b/packages/services/llm-gateway/src/proxy/openai-proxy.ts
@ -0,0 +1,216 @@
+import { FastifyRequest, FastifyReply } from 'fastify';
+import { GatewayConfig } from '../config';
+import { ApiKeyRecord } from '../types';
+import { isModelAllowed } from '../middleware/auth';
+import { recordFromOpenAIResponse } from '../logging/usage-tracker';
+import { recordAudit } from '../logging/audit-logger';
+import { pipeSSEStream, createStreamUsageTracker } from './stream-pipe';
+
+export function createOpenAIEmbeddingsProxy(config: GatewayConfig) {
+  return async function handleEmbeddings(request: FastifyRequest, reply: FastifyReply): Promise<void> {
+    const startTime = Date.now();
+    const apiKeyRecord: ApiKeyRecord = (request as any).apiKeyRecord;
+    const clientIp = request.ip || request.headers['x-forwarded-for'] as string || 'unknown';
+
+    let body: any;
+    try {
+      body = JSON.parse(request.body as string);
+    } catch {
+      reply.status(400).send({
+        error: { type: 'invalid_request_error', message: 'Invalid JSON body.' },
+      });
+      return;
+    }
+
+    const model = body.model || 'unknown';
+
+    if (!isModelAllowed(apiKeyRecord, model)) {
+      reply.status(403).send({
+        error: {
+          type: 'permission_error',
+          message: `Model "${model}" is not allowed for this API key.`,
+        },
+      });
+      return;
+    }
+
+    let upstreamResponse: Response;
+    try {
+      upstreamResponse = await fetch(`${config.openaiUpstreamUrl}/v1/embeddings`, {
+        method: 'POST',
+        headers: {
+          'Content-Type': 'application/json',
+          'Authorization': `Bearer ${config.openaiApiKey}`,
+        },
+        body: JSON.stringify(body),
+      });
+    } catch (err: any) {
+      recordAudit({
+        apiKeyId: apiKeyRecord.id,
+        requestMethod: 'POST',
+        requestPath: '/v1/embeddings',
+        requestModel: model,
+        requestIp: clientIp,
+        contentFiltered: false,
+        filterRuleId: null,
+        injectionApplied: false,
+        responseStatus: 502,
+        durationMs: Date.now() - startTime,
+      });
+
+      reply.status(502).send({
+        error: {
+          type: 'upstream_error',
+          message: `Failed to connect to upstream API: ${err.message}`,
+        },
+      });
+      return;
+    }
+
+    const durationMs = Date.now() - startTime;
+    const responseText = await upstreamResponse.text();
+
+    try {
+      const responseJson = JSON.parse(responseText);
+      if (responseJson.usage) {
+        recordFromOpenAIResponse(
+          apiKeyRecord.id,
+          model,
+          responseJson.usage,
+          upstreamResponse.status,
+          durationMs,
+        );
+      }
+    } catch {
+      // Not JSON
+    }
+
+    recordAudit({
+      apiKeyId: apiKeyRecord.id,
+      requestMethod: 'POST',
+      requestPath: '/v1/embeddings',
+      requestModel: model,
+      requestIp: clientIp,
+      contentFiltered: false,
+      filterRuleId: null,
+      injectionApplied: false,
+      responseStatus: upstreamResponse.status,
+      durationMs,
+    });
+
+    reply.raw.writeHead(upstreamResponse.status, {
+      'Content-Type': upstreamResponse.headers.get('content-type') || 'application/json',
+    });
+    reply.raw.end(responseText);
+  };
+}
+
+export function createOpenAIChatProxy(config: GatewayConfig) {
+  return async function handleChatCompletions(request: FastifyRequest, reply: FastifyReply): Promise<void> {
+    const startTime = Date.now();
+    const apiKeyRecord: ApiKeyRecord = (request as any).apiKeyRecord;
+    const clientIp = request.ip || request.headers['x-forwarded-for'] as string || 'unknown';
+
+    let body: any;
+    try {
+      body = JSON.parse(request.body as string);
+    } catch {
+      reply.status(400).send({
+        error: { type: 'invalid_request_error', message: 'Invalid JSON body.' },
+      });
+      return;
+    }
+
+    const model = body.model || 'unknown';
+
+    if (!isModelAllowed(apiKeyRecord, model)) {
+      reply.status(403).send({
+        error: {
+          type: 'permission_error',
+          message: `Model "${model}" is not allowed for this API key.`,
+        },
+      });
+      return;
+    }
+
+    let upstreamResponse: Response;
+    try {
+      upstreamResponse = await fetch(`${config.openaiUpstreamUrl}/v1/chat/completions`, {
+        method: 'POST',
+        headers: {
+          'Content-Type': 'application/json',
+          'Authorization': `Bearer ${config.openaiApiKey}`,
+        },
+        body: JSON.stringify(body),
+      });
+    } catch (err: any) {
+      recordAudit({
+        apiKeyId: apiKeyRecord.id,
+        requestMethod: 'POST',
+        requestPath: '/v1/chat/completions',
+        requestModel: model,
+        requestIp: clientIp,
+        contentFiltered: false,
+        filterRuleId: null,
+        injectionApplied: false,
+        responseStatus: 502,
+        durationMs: Date.now() - startTime,
+      });
+
+      reply.status(502).send({
+        error: {
+          type: 'upstream_error',
+          message: `Failed to connect to upstream API: ${err.message}`,
+        },
+      });
+      return;
+    }
+
+    const durationMs = Date.now() - startTime;
+
+    // Handle streaming chat completions
+    if (body.stream && upstreamResponse.ok && upstreamResponse.body) {
+      reply.raw.writeHead(upstreamResponse.status, {
+        'Content-Type': 'text/event-stream',
+        'Cache-Control': 'no-cache',
+        'Connection': 'keep-alive',
+        'X-Accel-Buffering': 'no',
+      });
+
+      await pipeSSEStream(upstreamResponse.body, reply.raw);
+
+      recordAudit({
+        apiKeyId: apiKeyRecord.id,
+        requestMethod: 'POST',
+        requestPath: '/v1/chat/completions',
+        requestModel: model,
+        requestIp: clientIp,
+        contentFiltered: false,
+        filterRuleId: null,
+        injectionApplied: false,
+        responseStatus: upstreamResponse.status,
+        durationMs: Date.now() - startTime,
+      });
+    } else {
+      const responseText = await upstreamResponse.text();
+
+      recordAudit({
+        apiKeyId: apiKeyRecord.id,
+        requestMethod: 'POST',
+        requestPath: '/v1/chat/completions',
+        requestModel: model,
+        requestIp: clientIp,
+        contentFiltered: false,
+        filterRuleId: null,
+        injectionApplied: false,
+        responseStatus: upstreamResponse.status,
+        durationMs,
+      });
+
+      reply.raw.writeHead(upstreamResponse.status, {
+        'Content-Type': upstreamResponse.headers.get('content-type') || 'application/json',
+      });
+      reply.raw.end(responseText);
+    }
+  };
+}
--- a/packages/services/llm-gateway/src/proxy/stream-pipe.ts
+++ b/packages/services/llm-gateway/src/proxy/stream-pipe.ts
@ -0,0 +1,94 @@
+import { Readable } from 'stream';
+import { ServerResponse } from 'http';
+
+/**
+ * Pipe an upstream SSE response body to the client response.
+ * Uses zero-buffering: chunks are forwarded as they arrive.
+ *
+ * Also provides a callback for each SSE data line, allowing
+ * the caller to extract usage info from the stream without buffering.
+ */
+export async function pipeSSEStream(
+  upstreamBody: ReadableStream<Uint8Array>,
+  clientResponse: ServerResponse,
+  onDataLine?: (line: string) => void,
+): Promise<void> {
+  const reader = upstreamBody.getReader();
+  const decoder = new TextDecoder();
+
+  try {
+    while (true) {
+      const { done, value } = await reader.read();
+      if (done) break;
+
+      const chunk = decoder.decode(value, { stream: true });
+
+      // Forward chunk immediately
+      const canContinue = clientResponse.write(chunk);
+
+      // Parse SSE data lines for usage extraction
+      if (onDataLine) {
+        const lines = chunk.split('\n');
+        for (const line of lines) {
+          if (line.startsWith('data: ')) {
+            onDataLine(line.slice(6).trim());
+          }
+        }
+      }
+
+      // Handle backpressure
+      if (!canContinue) {
+        await new Promise<void>((resolve) => clientResponse.once('drain', resolve));
+      }
+    }
+  } catch (err: any) {
+    // Connection closed by client — expected during cancellation
+    if (err.code === 'ERR_STREAM_PREMATURE_CLOSE' || clientResponse.destroyed) {
+      return;
+    }
+    throw err;
+  } finally {
+    reader.releaseLock();
+    clientResponse.end();
+  }
+}
+
+/**
+ * Extract usage from Anthropic's message_delta SSE event.
+ * The event format: {"type":"message_delta","delta":{...},"usage":{"output_tokens":N}}
+ * Final usage is in the message_start event's message.usage for input_tokens.
+ */
+export interface StreamUsage {
+  inputTokens: number;
+  outputTokens: number;
+}
+
+export function createStreamUsageTracker(): {
+  onDataLine: (line: string) => void;
+  getUsage: () => StreamUsage;
+} {
+  const usage: StreamUsage = { inputTokens: 0, outputTokens: 0 };
+
+  return {
+    onDataLine: (line: string) => {
+      if (line === '[DONE]') return;
+
+      try {
+        const event = JSON.parse(line);
+
+        // message_start contains input_tokens
+        if (event.type === 'message_start' && event.message?.usage) {
+          usage.inputTokens = event.message.usage.input_tokens || 0;
+        }
+
+        // message_delta contains output_tokens
+        if (event.type === 'message_delta' && event.usage) {
+          usage.outputTokens = event.usage.output_tokens || 0;
+        }
+      } catch {
+        // Not valid JSON — ignore
+      }
+    },
+    getUsage: () => usage,
+  };
+}
--- a/packages/services/llm-gateway/src/routes/anthropic.ts
+++ b/packages/services/llm-gateway/src/routes/anthropic.ts
@ -0,0 +1,13 @@
+import { FastifyInstance } from 'fastify';
+import { GatewayConfig } from '../config';
+import { authMiddleware } from '../middleware/auth';
+import { createAnthropicProxy } from '../proxy/anthropic-proxy';
+
+export function registerAnthropicRoutes(app: FastifyInstance, config: GatewayConfig): void {
+  const handleMessages = createAnthropicProxy(config);
+
+  app.post('/v1/messages', {
+    preHandler: authMiddleware,
+    handler: handleMessages,
+  });
+}
--- a/packages/services/llm-gateway/src/routes/health.ts
+++ b/packages/services/llm-gateway/src/routes/health.ts
@ -0,0 +1,15 @@
+import { FastifyInstance } from 'fastify';
+import { getPool } from '../db';
+
+export function registerHealthRoutes(app: FastifyInstance): void {
+  app.get('/health', async (_request, reply) => {
+    try {
+      const pool = getPool();
+      await pool.query('SELECT 1');
+      return { status: 'ok', service: 'llm-gateway', timestamp: new Date().toISOString() };
+    } catch (err: any) {
+      reply.status(503);
+      return { status: 'unhealthy', error: err.message };
+    }
+  });
+}
--- a/packages/services/llm-gateway/src/routes/openai.ts
+++ b/packages/services/llm-gateway/src/routes/openai.ts
@ -0,0 +1,19 @@
+import { FastifyInstance } from 'fastify';
+import { GatewayConfig } from '../config';
+import { authMiddleware } from '../middleware/auth';
+import { createOpenAIEmbeddingsProxy, createOpenAIChatProxy } from '../proxy/openai-proxy';
+
+export function registerOpenAIRoutes(app: FastifyInstance, config: GatewayConfig): void {
+  const handleEmbeddings = createOpenAIEmbeddingsProxy(config);
+  const handleChatCompletions = createOpenAIChatProxy(config);
+
+  app.post('/v1/embeddings', {
+    preHandler: authMiddleware,
+    handler: handleEmbeddings,
+  });
+
+  app.post('/v1/chat/completions', {
+    preHandler: authMiddleware,
+    handler: handleChatCompletions,
+  });
+}
--- a/packages/services/llm-gateway/src/types.ts
+++ b/packages/services/llm-gateway/src/types.ts
@ -0,0 +1,146 @@
+// ─── API Key Record (from gateway_api_keys table) ───
+
+export interface ApiKeyRecord {
+  id: string;
+  tenantId: string | null;
+  keyHash: string;
+  keyPrefix: string;
+  name: string;
+  owner: string;
+  permissions: ApiKeyPermissions;
+  rateLimitRpm: number;
+  rateLimitTpd: number;
+  monthlyBudget: number | null;
+  enabled: boolean;
+  expiresAt: Date | null;
+  lastUsedAt: Date | null;
+}
+
+export interface ApiKeyPermissions {
+  allowedModels: string[];   // ["*"] = all, or ["claude-haiku-4-5-*", "text-embedding-3-small"]
+  allowStreaming: boolean;
+  allowTools: boolean;
+}
+
+// ─── Injection Rule (from gateway_injection_rules table) ───
+
+export interface InjectionRule {
+  id: string;
+  tenantId: string | null;
+  name: string;
+  position: 'prepend' | 'append';
+  content: string;
+  matchModels: string[];     // ["*"] or specific model patterns
+  matchKeyIds: string[] | null;  // null = all keys
+  priority: number;
+  enabled: boolean;
+}
+
+// ─── Content Filter Rule (from gateway_content_rules table) ───
+
+export interface ContentRule {
+  id: string;
+  tenantId: string | null;
+  name: string;
+  type: 'keyword' | 'regex';
+  pattern: string;
+  action: 'block' | 'warn' | 'log';
+  rejectMessage: string | null;
+  priority: number;
+  enabled: boolean;
+}
+
+// ─── Anthropic Types ───
+
+export interface AnthropicSystemBlock {
+  type: 'text';
+  text: string;
+  cache_control?: { type: string };
+}
+
+export type AnthropicSystem = string | AnthropicSystemBlock[];
+
+export interface AnthropicRequestBody {
+  model: string;
+  system?: AnthropicSystem;
+  messages: AnthropicMessage[];
+  max_tokens: number;
+  stream?: boolean;
+  tools?: any[];
+  tool_choice?: any;
+  temperature?: number;
+  top_p?: number;
+  top_k?: number;
+  stop_sequences?: string[];
+  metadata?: any;
+  output_config?: any;
+  [key: string]: any;
+}
+
+export interface AnthropicMessage {
+  role: 'user' | 'assistant';
+  content: string | AnthropicContentBlock[];
+}
+
+export interface AnthropicContentBlock {
+  type: string;
+  text?: string;
+  source?: any;
+  [key: string]: any;
+}
+
+// ─── OpenAI Types ───
+
+export interface OpenAIEmbeddingRequest {
+  model: string;
+  input: string | string[];
+  encoding_format?: string;
+  dimensions?: number;
+}
+
+export interface OpenAIChatRequest {
+  model: string;
+  messages: Array<{ role: string; content: any }>;
+  stream?: boolean;
+  max_tokens?: number;
+  temperature?: number;
+  [key: string]: any;
+}
+
+// ─── Filter Result ───
+
+export interface FilterResult {
+  blocked: boolean;
+  reason?: string;
+  ruleId?: string;
+  action?: 'block' | 'warn' | 'log';
+}
+
+// ─── Usage Log ───
+
+export interface UsageLogEntry {
+  apiKeyId: string;
+  model: string;
+  provider: 'anthropic' | 'openai';
+  inputTokens: number;
+  outputTokens: number;
+  totalTokens: number;
+  costUsd: number | null;
+  durationMs: number;
+  statusCode: number;
+}
+
+// ─── Audit Log ───
+
+export interface AuditLogEntry {
+  apiKeyId: string;
+  requestMethod: string;
+  requestPath: string;
+  requestModel: string | null;
+  requestIp: string;
+  contentFiltered: boolean;
+  filterRuleId: string | null;
+  injectionApplied: boolean;
+  responseStatus: number;
+  durationMs: number;
+}
--- a/packages/services/llm-gateway/tsconfig.json
+++ b/packages/services/llm-gateway/tsconfig.json
@ -0,0 +1,23 @@
+{
+  "compilerOptions": {
+    "module": "commonjs",
+    "moduleResolution": "node",
+    "declaration": true,
+    "removeComments": true,
+    "allowSyntheticDefaultImports": true,
+    "esModuleInterop": true,
+    "target": "ES2021",
+    "sourceMap": true,
+    "outDir": "./dist",
+    "baseUrl": "./",
+    "incremental": true,
+    "skipLibCheck": true,
+    "strict": true,
+    "strictPropertyInitialization": false,
+    "noImplicitAny": true,
+    "forceConsistentCasingInFileNames": true,
+    "noFallthroughCasesInSwitch": true
+  },
+  "include": ["src/**/*"],
+  "exclude": ["node_modules", "dist"]
+}
--- a/scripts/init-db.sql
+++ b/scripts/init-db.sql
@ -1810,6 +1810,117 @@ COMMENT ON TABLE collection_directives IS '信息收集指令表 - 管理员配
 CREATE INDEX idx_collection_directives_tenant ON collection_directives(tenant_id);
 CREATE INDEX idx_collection_directives_tenant_enabled ON collection_directives(tenant_id, enabled);

+-- ===========================================
+-- LLM Gateway 相关表
+-- 对外 API 代理服务的配置、用量和审计
+-- ===========================================
+
+-- API Keys for external users
+CREATE TABLE IF NOT EXISTS gateway_api_keys (
+    id UUID PRIMARY KEY DEFAULT gen_random_uuid(),
+    tenant_id UUID,
+    key_hash VARCHAR(64) NOT NULL UNIQUE,
+    key_prefix VARCHAR(12) NOT NULL,
+    name VARCHAR(100) NOT NULL,
+    owner VARCHAR(200) NOT NULL DEFAULT '',
+    permissions JSONB NOT NULL DEFAULT '{"allowedModels": ["*"], "allowStreaming": true, "allowTools": true}',
+    rate_limit_rpm INTEGER NOT NULL DEFAULT 60,
+    rate_limit_tpd INTEGER NOT NULL DEFAULT 1000000,
+    monthly_budget DECIMAL(10,2),
+    enabled BOOLEAN NOT NULL DEFAULT true,
+    expires_at TIMESTAMP,
+    last_used_at TIMESTAMP,
+    created_by UUID,
+    created_at TIMESTAMP NOT NULL DEFAULT NOW()
+);
+
+COMMENT ON TABLE gateway_api_keys IS 'LLM Gateway - 外部用户 API Key 管理';
+CREATE INDEX IF NOT EXISTS idx_gateway_api_keys_hash ON gateway_api_keys (key_hash);
+CREATE INDEX IF NOT EXISTS idx_gateway_api_keys_tenant ON gateway_api_keys (tenant_id);
+CREATE INDEX IF NOT EXISTS idx_gateway_api_keys_enabled ON gateway_api_keys (enabled);
+
+-- Injection rules (regulatory content injected into system prompts)
+CREATE TABLE IF NOT EXISTS gateway_injection_rules (
+    id UUID PRIMARY KEY DEFAULT gen_random_uuid(),
+    tenant_id UUID,
+    name VARCHAR(100) NOT NULL,
+    description TEXT,
+    position VARCHAR(10) NOT NULL DEFAULT 'append' CHECK (position IN ('prepend', 'append')),
+    content TEXT NOT NULL,
+    match_models JSONB NOT NULL DEFAULT '["*"]',
+    match_key_ids JSONB,
+    priority INTEGER NOT NULL DEFAULT 0,
+    enabled BOOLEAN NOT NULL DEFAULT true,
+    created_by UUID,
+    created_at TIMESTAMP NOT NULL DEFAULT NOW(),
+    updated_at TIMESTAMP NOT NULL DEFAULT NOW()
+);
+
+COMMENT ON TABLE gateway_injection_rules IS 'LLM Gateway - 监管内容注入规则';
+CREATE INDEX IF NOT EXISTS idx_gateway_injection_rules_tenant ON gateway_injection_rules (tenant_id);
+CREATE INDEX IF NOT EXISTS idx_gateway_injection_rules_enabled ON gateway_injection_rules (tenant_id, enabled);
+
+-- Content filter rules (message auditing)
+CREATE TABLE IF NOT EXISTS gateway_content_rules (
+    id UUID PRIMARY KEY DEFAULT gen_random_uuid(),
+    tenant_id UUID,
+    name VARCHAR(100) NOT NULL,
+    type VARCHAR(20) NOT NULL DEFAULT 'keyword' CHECK (type IN ('keyword', 'regex')),
+    pattern TEXT NOT NULL,
+    action VARCHAR(20) NOT NULL DEFAULT 'block' CHECK (action IN ('block', 'warn', 'log')),
+    reject_message TEXT,
+    priority INTEGER NOT NULL DEFAULT 0,
+    enabled BOOLEAN NOT NULL DEFAULT true,
+    created_by UUID,
+    created_at TIMESTAMP NOT NULL DEFAULT NOW(),
+    updated_at TIMESTAMP NOT NULL DEFAULT NOW()
+);
+
+COMMENT ON TABLE gateway_content_rules IS 'LLM Gateway - 内容审查过滤规则';
+CREATE INDEX IF NOT EXISTS idx_gateway_content_rules_tenant ON gateway_content_rules (tenant_id);
+CREATE INDEX IF NOT EXISTS idx_gateway_content_rules_enabled ON gateway_content_rules (tenant_id, enabled);
+
+-- Usage logs (token consumption tracking)
+CREATE TABLE IF NOT EXISTS gateway_usage_logs (
+    id UUID PRIMARY KEY DEFAULT gen_random_uuid(),
+    api_key_id UUID NOT NULL REFERENCES gateway_api_keys(id) ON DELETE CASCADE,
+    model VARCHAR(100) NOT NULL,
+    provider VARCHAR(20) NOT NULL,
+    input_tokens INTEGER NOT NULL DEFAULT 0,
+    output_tokens INTEGER NOT NULL DEFAULT 0,
+    total_tokens INTEGER NOT NULL DEFAULT 0,
+    cost_usd DECIMAL(10,6),
+    duration_ms INTEGER NOT NULL DEFAULT 0,
+    status_code INTEGER NOT NULL DEFAULT 200,
+    created_at TIMESTAMP NOT NULL DEFAULT NOW()
+);
+
+COMMENT ON TABLE gateway_usage_logs IS 'LLM Gateway - API 用量记录';
+CREATE INDEX IF NOT EXISTS idx_gateway_usage_logs_key ON gateway_usage_logs (api_key_id);
+CREATE INDEX IF NOT EXISTS idx_gateway_usage_logs_created ON gateway_usage_logs (created_at);
+CREATE INDEX IF NOT EXISTS idx_gateway_usage_logs_key_created ON gateway_usage_logs (api_key_id, created_at);
+
+-- Audit logs (request/response auditing)
+CREATE TABLE IF NOT EXISTS gateway_audit_logs (
+    id UUID PRIMARY KEY DEFAULT gen_random_uuid(),
+    api_key_id UUID NOT NULL REFERENCES gateway_api_keys(id) ON DELETE CASCADE,
+    request_method VARCHAR(10) NOT NULL,
+    request_path VARCHAR(200) NOT NULL,
+    request_model VARCHAR(100),
+    request_ip VARCHAR(50) NOT NULL DEFAULT '',
+    content_filtered BOOLEAN NOT NULL DEFAULT false,
+    filter_rule_id UUID,
+    injection_applied BOOLEAN NOT NULL DEFAULT false,
+    response_status INTEGER NOT NULL DEFAULT 200,
+    duration_ms INTEGER NOT NULL DEFAULT 0,
+    created_at TIMESTAMP NOT NULL DEFAULT NOW()
+);
+
+COMMENT ON TABLE gateway_audit_logs IS 'LLM Gateway - 请求审计日志';
+CREATE INDEX IF NOT EXISTS idx_gateway_audit_logs_key ON gateway_audit_logs (api_key_id);
+CREATE INDEX IF NOT EXISTS idx_gateway_audit_logs_created ON gateway_audit_logs (created_at);
+CREATE INDEX IF NOT EXISTS idx_gateway_audit_logs_filtered ON gateway_audit_logs (content_filtered) WHERE content_filtered = true;
+
 -- ===========================================
 -- 结束
 -- ===========================================
				`@ -0,0 +1 @@`
				`export { LLMGatewayPage } from './presentation/pages/LLMGatewayPage';`