fix(ai): V3 长期问题修复+规则合规+竞品对标

## P1 安全加固 - 原子化每日限额（tryConsumeDailyQuota）解决 TOCTOU 竞态 - 流式端点补齐 Zod 校验 + rate limit + 服务端强制 systemPrompt - 配额回退机制（refundDailyQuota）：过滤/失败不扣配额 - PII 最小化：移除 AI prompt 中的学生姓名 ## P1 数据一致性 - 修复 capability 埋点缺失 child_summary/study_path 类型 - 创建 data-access.ts：真实统计聚合替代硬编码零 - 修复 generateChildSummary/recommendStudyPath 的 capability 标记 ## P2 可靠性 - AI 调用重试机制（withRetry 指数退避，429/5xx，2 次重试） - 30s 超时配置 - 流式 controller 安全 enqueue（防已关闭抛错） - localStorage 防抖持久化（500ms，流式过程中跳过） ## P2 TypeScript/规则合规 - 移除 as 断言（VariantType 类型守卫、Permission 类型、StreamErrorKey） - 补齐返回类型标注（POST/getStatusFromError/DashboardLayout） - 拆分 use-ai-chat-stream hook（190→107 行，函数体≤80 行） - 抽取 stream-utils.ts（SSE 解析/错误映射/消息工具） - Tailwind 任意值添加注释说明（max-w-[80%] 聊天气泡） ## P3 竞品对标 - 苏格拉底式辅导强化（对标 Khanmigo）： - SOCRATIC_TUTOR_SYSTEM_PROMPT 3 级提示升级 - 强化 STUDENT_BLOCKED_PATTERNS 正则（中英文答案拦截） - validateSocraticOutput 服务端校验（问号结尾+连续陈述句限制） - socratic_warning SSE 事件类型 - 知识图谱集成（对标 Squirrel AI）： - StudyPathInput 新增 knowledgeGraph/textbookId 字段 - recommendStudyPathAction 自动从 textbooks 模块获取图谱+掌握度 - STUDY_PATH_SYSTEM_PROMPT 增加前置依赖链规则 - WEAKNESS_ANALYSIS_SYSTEM_PROMPT 增加 rootCause 字段 ## 架构文档同步 - 004 更新 AI 模块章节（V3 标记/新导出/依赖关系/安全机制/文件清单） - 005 更新 modules.ai 节点（dependsOn/exports/dataAccess/streamUtils/dependencyMatrix）
2026-06-23 09:39:18 +08:00
parent 036a2f2839
commit 696346dc08
22 changed files with 847 additions and 238 deletions
--- a/docs/architecture/004_architecture_impact_map.md
+++ b/docs/architecture/004_architecture_impact_map.md
@@ -2056,33 +2056,41 @@ src/auth.ts ──▶ import { ... } from "@/shared/lib/permissions"
 | **student** | `AiStudyPath` | `student/*` | 学习路径推荐 — V2 新增 |

 **依赖关系**：
- `modules/ai` → `shared/lib/ai`（AI SDK 封装，含流式 `createAiChatCompletionStream`）
+- `modules/ai` → `shared/lib/ai`（AI SDK 封装，含流式 `createAiChatCompletionStream`，V3 增加 `withRetry` 重试机制）
 - `modules/ai` → `shared/lib/auth-guard`（权限校验）
 - `modules/ai` → `shared/lib/track-event`（使用量埋点）
 - `modules/ai` → `shared/types/permissions`（权限常量）
 - `modules/ai` → `shared/types/action-state`（返回值类型）
+- `modules/ai` → `modules/textbooks/data-access-graph`（知识图谱+掌握度查询）— V3 新增
 - `app/(dashboard)/layout` → `modules/ai`（全局 Provider + Widget）— V2 新增
+- `app/api/ai/chat/stream` → `modules/ai/services/content-safety`（SSE 端点安全过滤）— V3 加固
 - 业务模块 → `modules/ai/context/ai-client-provider`（通过 Context 注入）
 - 业务模块 → `modules/ai/components/*`（组合 AI 组件）

-**安全机制（V2 新增）**：
+**安全机制（V2 新增 / V3 加固）**：
 - 输入过滤：`filterUserInput` 拦截暴力/自残/色情/PII
 - 输出过滤：`filterAiOutput` 扫描 AI 回复
 - 每日限制：学生 50 次/天，教师 200 次/天，家长 30 次/天
 - 学生 Socratic 模式：system prompt 强制不直接给答案
 - SSE 端点权限校验：`requirePermission(AI_CHAT)`
+- V3 原子配额：`tryConsumeDailyQuota` 解决 TOCTOU 竞态，`refundDailyQuota` 过滤/失败时回退
+- V3 苏格拉底校验：`validateSocraticOutput` 检查问号结尾+连续陈述句限制，`SOCRATIC_TUTOR_SYSTEM_PROMPT` 3 级提示升级
+- V3 SSE 端点加固：Zod 校验 + rate limit + 服务端强制 systemPrompt（忽略客户端）
+- V3 重试机制：`withRetry` 指数退避（429/5xx，2 次重试，1s/2s/4s 延迟）
+- V3 知识图谱集成：`recommendStudyPathAction` 自动从 textbooks 模块获取图谱+掌握度注入 prompt

 **文件清单**：

 | 文件 | 行数 | 职责 |
 |------|------|------|
-| `modules/ai/types.ts` | ~270 | 类型定义（8 个业务场景类型 + AiService/AiClientService） |
-| `modules/ai/schema.ts` | ~205 | Zod 验证 schema（8 个输入 + 8 个输出） |
-| `modules/ai/actions.ts` | ~340 | 9 个 Server Actions（含权限校验） |
-| `modules/ai/services/ai-service.ts` | ~400 | DefaultAiService 实现（8 个方法） |
-| `modules/ai/services/prompt-templates.ts` | ~210 | 8 个系统提示词模板 |
-| `modules/ai/services/usage-tracker.ts` | ~83 | AI 使用量埋点 |
-| `modules/ai/services/content-safety.ts` | ~130 | 内容安全过滤（输入/输出/每日限制）— V2 新增 |
+| `modules/ai/types.ts` | ~290 | 类型定义（8 个业务场景类型 + AiService/AiClientService，V3 新增 knowledgeGraph/textbookId） |
+| `modules/ai/schema.ts` | ~230 | Zod 验证 schema（8 个输入 + 8 个输出，V3 新增 knowledgeGraph/textbookId） |
+| `modules/ai/actions.ts` | ~400 | 9 个 Server Actions（含权限校验，V3 新增知识图谱获取） |
+| `modules/ai/data-access.ts` | ~138 | AI 事件存储+统计聚合 — V3 新增 |
+| `modules/ai/services/ai-service.ts` | ~430 | DefaultAiService 实现（8 个方法，V3 新增知识图谱上下文注入） |
+| `modules/ai/services/prompt-templates.ts` | ~250 | 9 个系统提示词模板（V3 新增 SOCRATIC_TUTOR_SYSTEM_PROMPT） |
+| `modules/ai/services/usage-tracker.ts` | ~90 | AI 使用量埋点（V3 新增 child_summary/study_path 能力） |
+| `modules/ai/services/content-safety.ts` | ~290 | 内容安全过滤（输入/输出/每日限制/原子配额/苏格拉底校验）— V2 新增 / V3 加固 |
 | `modules/ai/context/ai-client-provider.tsx` | ~62 | React Context Provider + Hooks |
 | `modules/ai/components/ai-assistant-widget.tsx` | ~170 | 全局 AI 助手悬浮按钮 — V2 新增 |
 | `modules/ai/components/ai-chat-panel.tsx` | ~305 | AI 对话面板（流式 + Markdown）— V2 增强 |
@@ -2097,7 +2105,8 @@ src/auth.ts ──▶ import { ... } from "@/shared/lib/permissions"
 | `modules/ai/components/ai-error-boundary.tsx` | ~88 | AI 错误边界 |
 | `modules/ai/components/ai-skeleton.tsx` | ~47 | AI 骨架屏 |
 | `modules/ai/components/ai-provider-selector.tsx` | ~129 | 服务商选择器 |
-| `modules/ai/hooks/use-ai-chat-stream.ts` | ~170 | 流式 AI 对话 Hook — V2 新增 |
+| `modules/ai/hooks/use-ai-chat-stream.ts` | ~107 | 流式 AI 对话 Hook（V3 拆分至 ≤80 行函数体）— V2 新增 / V3 重构 |
+| `modules/ai/hooks/stream-utils.ts` | ~120 | SSE 流解析工具（consumeSseStream/getStreamErrorKey 等）— V3 新增 |
 | `modules/ai/hooks/use-ai-chat.ts` | ~57 | 非流式 AI 对话 Hook |
 | `modules/ai/hooks/use-ai-suggestion.ts` | ~72 | AI 建议 Hook |
 | `app/api/ai/chat/stream/route.ts` | ~160 | SSE 流式端点 — V2 新增 |
--- a/docs/architecture/005_architecture_data.json
+++ b/docs/architecture/005_architecture_data.json
@@ -16341,7 +16341,8 @@
    },
    "ai": {
      "dependsOn": [
-        "shared"
+        "shared",
+        "textbooks"
      ],
      "uses": {
        "shared": [
@@ -16353,6 +16354,10 @@
          "types.action-state",
          "lib.track-event.trackEvent",
          "i18n.messages"
+        ],
+        "textbooks": [
+          "data-access-graph.getKnowledgePointsWithRelations",
+          "data-access-graph.getStudentKpMastery"
        ]
      },
      "exports": {
@@ -16422,11 +16427,27 @@
          "filterAiOutput",
          "checkDailyLimit",
          "incrementDailyUsage",
-          "getDailyLimit"
+          "getDailyLimit",
+          "tryConsumeDailyQuota",
+          "refundDailyQuota",
+          "validateSocraticOutput"
+        ],
+        "dataAccess": [
+          "recordAiEvent",
+          "getAiUsageStats"
+        ],
+        "streamUtils": [
+          "consumeSseStream",
+          "getStreamErrorKey",
+          "extractErrorMessage",
+          "removeTrailingEmptyAssistant",
+          "appendTokenToLastAssistant"
        ],
        "promptTemplates": [
          "CHILD_SUMMARY_SYSTEM_PROMPT",
-          "STUDY_PATH_SYSTEM_PROMPT"
+          "STUDY_PATH_SYSTEM_PROMPT",
+          "SOCRATIC_TUTOR_SYSTEM_PROMPT",
+          "CHAT_SYSTEM_PROMPT"
        ]
      },
      "integrations": {
@@ -16841,6 +16862,12 @@
        "type": "normal",
        "description": "使用 lib/ai.createAiChatCompletion、auth-guard.requirePermission、types.permissions、types.action-state"
      },
+      {
+        "from": "ai",
+        "to": "textbooks",
+        "type": "normal",
+        "description": "V3 新增：使用 data-access-graph.getKnowledgePointsWithRelations/getStudentKpMastery 获取知识图谱与掌握度，注入学习路径推荐"
+      },
      {
        "from": "homework",
        "to": "ai",
--- a/src/app/(dashboard)/layout.tsx
+++ b/src/app/(dashboard)/layout.tsx
@@ -34,7 +34,7 @@ export default function DashboardLayout({
  children,
 }: {
  children: React.ReactNode
-}) {
+}): React.ReactNode {
  return (
    <AiClientProvider service={aiClientService}>
      <SidebarProvider sidebar={<AppSidebar />}>
--- a/src/app/api/ai/chat/route.ts
+++ b/src/app/api/ai/chat/route.ts
@@ -6,14 +6,14 @@ import { Permissions } from "@/shared/types/permissions"

 export const dynamic = "force-dynamic"

-const getStatusFromError = (message: string) => {
+const getStatusFromError = (message: string): number => {
  if (message === "Invalid payload" || message === "Messages are required") return 400
  if (message === "AI API key missing") return 500
  if (message === "Empty response") return 502
  return 502
 }

-export async function POST(req: Request) {
+export async function POST(req: Request): Promise<NextResponse> {
  try {
    const ctx = await requirePermission(Permissions.AI_CHAT)
    const userId = ctx.userId
--- a/src/app/api/ai/chat/stream/route.ts
+++ b/src/app/api/ai/chat/stream/route.ts
@@ -6,15 +6,22 @@ import { requirePermission, PermissionDeniedError } from "@/shared/lib/auth-guar
 import { createAiChatCompletionStream } from "@/shared/lib/ai/client"
 import { getAiErrorMessage } from "@/shared/lib/ai"
 import { trackEvent } from "@/shared/lib/track-event"
+import {
+  rateLimit,
+  rateLimitKey,
+  RATE_LIMIT_RULES,
+} from "@/shared/lib/rate-limit"
 import { env } from "@/env.mjs"

-import { CHAT_SYSTEM_PROMPT } from "@/modules/ai/services/prompt-templates"
+import { CHAT_SYSTEM_PROMPT, SOCRATIC_TUTOR_SYSTEM_PROMPT } from "@/modules/ai/services/prompt-templates"
 import {
  filterUserInput,
  filterAiOutput,
-  checkDailyLimit,
-  incrementDailyUsage,
+  tryConsumeDailyQuota,
+  refundDailyQuota,
+  validateSocraticOutput,
 } from "@/modules/ai/services/content-safety"
+import { AiChatInputSchema } from "@/modules/ai/schema"
 import type { AiChatMessage } from "@/modules/ai/types"

 /**
@@ -23,11 +30,14 @@ import type { AiChatMessage } from "@/modules/ai/types"
 * 使用 Server-Sent Events 逐 token 推送 AI 回复，
 * 降低用户感知延迟。
 *
- * 安全：
+ * 安全策略（与非流式端点一致）：
 * - requirePermission(AI_CHAT) 权限校验
+ * - rateLimit 每分钟限流（防高频滥用）
+ * - Zod 校验输入（AiChatInputSchema，限制消息数/长度）
+ * - tryConsumeDailyQuota 原子化每日限额（防 TOCTOU 竞态）
 * - 输入/输出内容安全过滤
- * - 每日交互限制
- * - 学生侧 Socratic 模式
+ * - 学生侧 Socratic 模式（服务端强制，忽略客户端 systemPrompt）
+ * - 过滤/失败时 refundDailyQuota（不惩罚用户）
 */

 const formatEvent = (data: unknown): string => {
@@ -50,34 +60,45 @@ export async function POST(request: NextRequest): Promise<Response> {
    const userRole = session?.user?.role ?? "student"
    const isStudent = userRole === "student"

-    // 2. 每日限制
-    const limitCheck = checkDailyLimit(ctx.userId, userRole)
-    if (limitCheck.blocked) {
+    // 2. Rate limit（与非流式端点一致）
+    const limitResult = rateLimit({
+      key: rateLimitKey("ai-chat-stream", ctx.userId),
+      ...RATE_LIMIT_RULES.AI_CHAT,
+    })
+    if (!limitResult.success) {
+      return new Response(formatError("Rate limit exceeded"), {
+        status: 429,
+        headers: { "Content-Type": "text/event-stream" },
+      })
+    }
+
+    // 3. Zod 校验输入（防止超大 message 导致 token 爆炸）
+    const rawBody = await request.json().catch(() => null)
+    const parseResult = AiChatInputSchema.safeParse(rawBody)
+    if (!parseResult.success) {
+      return new Response(formatError("Invalid input"), {
+        status: 400,
+        headers: { "Content-Type": "text/event-stream" },
+      })
+    }
+    const body = parseResult.data
+
+    // 4. 原子化每日限额检查（防 TOCTOU 竞态）
+    const quotaResult = tryConsumeDailyQuota(ctx.userId, userRole)
+    if (quotaResult.blocked) {
      return new Response(formatError("Daily limit reached"), {
        status: 429,
        headers: { "Content-Type": "text/event-stream" },
      })
    }

-    // 3. 解析请求
-    const body = (await request.json()) as {
-      messages?: AiChatMessage[]
-      providerId?: string
-      systemPrompt?: string
-    }
-
-    if (!body.messages || !Array.isArray(body.messages) || body.messages.length === 0) {
-      return new Response(formatError("Messages are required"), {
-        status: 400,
-        headers: { "Content-Type": "text/event-stream" },
-      })
-    }
-
-    // 4. 输入安全过滤
+    // 5. 输入安全过滤
    for (const msg of body.messages) {
      if (msg.role === "user") {
        const filterResult = filterUserInput(msg.content, { isStudent })
        if (filterResult.blocked) {
+          // 输入被过滤，回退配额
+          refundDailyQuota(ctx.userId)
          return new Response(formatError("Input blocked by safety filter"), {
            status: 400,
            headers: { "Content-Type": "text/event-stream" },
@@ -86,24 +107,35 @@ export async function POST(request: NextRequest): Promise<Response> {
      }
    }

-    // 5. 构建 system prompt（学生侧 Socratic 模式）
-    const baseSystemPrompt = body.systemPrompt ?? CHAT_SYSTEM_PROMPT
-    const studentSystemPrompt = isStudent
-      ? `${baseSystemPrompt}\n\nIMPORTANT: You are in student mode. Use the Socratic method. Do NOT give direct answers. Guide the student to find the answer themselves through questions and hints.`
-      : baseSystemPrompt
+    // 6. 构建 system prompt
+    //    安全：服务端强制使用系统提示词，忽略客户端传入的 systemPrompt，
+    //    防止学生绕过 Socratic 模式获取直接答案。
+    const baseSystemPrompt = isStudent
+      ? SOCRATIC_TUTOR_SYSTEM_PROMPT
+      : CHAT_SYSTEM_PROMPT

    const messages: AiChatMessage[] = [
-      { role: "system", content: studentSystemPrompt },
+      { role: "system", content: baseSystemPrompt },
      ...body.messages,
    ]

-    // 6. 流式调用 AI
+    // 7. 流式调用 AI
    const stream = new ReadableStream<Uint8Array>({
      async start(controller) {
        const startTime = Date.now()
        let fullContent = ""
        let success = true
        let errorMessage: string | undefined
+        let wasFiltered = false
+
+        const safeEnqueue = (text: string): void => {
+          // 防止 controller 已关闭时 enqueue 抛错
+          try {
+            controller.enqueue(encoder.encode(text))
+          } catch {
+            // controller 已关闭，忽略
+          }
+        }

        try {
          const aiStream = createAiChatCompletionStream({
@@ -119,27 +151,49 @@ export async function POST(request: NextRequest): Promise<Response> {
            // 输出安全过滤（逐 chunk 检查关键词）
            const outputFilter = filterAiOutput(chunk, { isStudent })
            if (outputFilter.blocked) {
-              controller.enqueue(encoder.encode(formatEvent({
+              safeEnqueue(formatEvent({
                type: "filtered",
                message: "Content filtered for safety",
-              })))
+              }))
              success = false
+              wasFiltered = true
              break
            }

-            controller.enqueue(encoder.encode(formatEvent({ type: "token", content: chunk })))
+            safeEnqueue(formatEvent({ type: "token", content: chunk }))
          }

-          // 增加每日使用计数
-          incrementDailyUsage(ctx.userId)
+          // 仅在成功完成时保留配额计数
+          // 过滤或失败时回退配额（不惩罚用户）
+          if (!success || wasFiltered) {
+            refundDailyQuota(ctx.userId)
+          }

-          controller.enqueue(encoder.encode(FORMAT_DONE))
+          // 学生侧：苏格拉底式辅导输出校验
+          // 流式完成后检查完整回复是否符合引导式教学原则
+          if (success && !wasFiltered && isStudent && fullContent) {
+            const socraticResult = validateSocraticOutput(fullContent)
+            if (!socraticResult.valid) {
+              safeEnqueue(formatEvent({
+                type: "socratic_warning",
+                message: "AI response did not fully follow Socratic tutoring principles.",
+              }))
+            }
+          }
+
+          safeEnqueue(FORMAT_DONE)
        } catch (error) {
          success = false
          errorMessage = error instanceof Error ? error.message : String(error)
-          controller.enqueue(encoder.encode(formatError(getAiErrorMessage(error))))
+          // AI 调用失败，回退配额
+          refundDailyQuota(ctx.userId)
+          safeEnqueue(formatError(getAiErrorMessage(error)))
        } finally {
+          try {
            controller.close()
+          } catch {
+            // 已关闭，忽略
+          }

          // 埋点
          void trackEvent({
@@ -152,6 +206,7 @@ export async function POST(request: NextRequest): Promise<Response> {
              tokenCount: fullContent.length / 4,
              errorMessage,
              isStudent,
+              filtered: wasFiltered,
            },
          }).catch(() => {
            // 静默失败
--- a/src/modules/ai/actions.ts
+++ b/src/modules/ai/actions.ts
@@ -4,9 +4,11 @@ import { getTranslations } from "next-intl/server"

 import type { ActionState } from "@/shared/types/action-state"
 import { requirePermission, PermissionDeniedError } from "@/shared/lib/auth-guard"
-import { Permissions } from "@/shared/types/permissions"
+import { Permissions, type Permission } from "@/shared/types/permissions"

 import { createAiService, safeAiCall } from "./services/ai-service"
+import { getAiUsageStats } from "./data-access"
+import { getKnowledgePointsWithRelations, getStudentKpMastery } from "@/modules/textbooks/data-access-graph"
 import {
  AiChatInputSchema,
  GradingInputSchema,
@@ -42,10 +44,10 @@ import type {
 // ---------------------------------------------------------------------------

 const requireAiPermission = async (
-  ...permissions: readonly string[]
+  ...permissions: readonly Permission[]
 ): Promise<{ userId: string }> => {
  const results = await Promise.all(
-    permissions.map((p) => requirePermission(p as never))
+    permissions.map((p) => requirePermission(p))
  )
  return { userId: results[0].userId }
 }
@@ -294,8 +296,56 @@ export async function recommendStudyPathAction(
      return { success: false, message: t("error.invalidInput") }
    }

+    const serviceInput: StudyPathInput = { ...parsed.data }
+
+    // V3：知识图谱集成（对标 Squirrel AI 纳米级知识图谱）
+    // 若传入 textbookId 但未传入 knowledgeGraph，自动从 textbooks 模块获取
+    if (parsed.data.textbookId && !parsed.data.knowledgeGraph) {
+      try {
+        const [kps, masteryMap] = await Promise.all([
+          getKnowledgePointsWithRelations(parsed.data.textbookId),
+          getStudentKpMastery(parsed.data.studentId, parsed.data.textbookId),
+        ])
+
+        if (kps.length > 0) {
+          serviceInput.knowledgeGraph = {
+            nodes: kps.map((kp) => ({
+              id: kp.id,
+              name: kp.name,
+              level: kp.level,
+              masteryLevel: masteryMap.get(kp.id)?.masteryLevel,
+            })),
+            edges: kps
+              .flatMap((kp) =>
+                (kp.prerequisiteIds ?? []).map((prereqId) => ({
+                  from: prereqId,
+                  to: kp.id,
+                  type: "prerequisite" as const,
+                }))
+              ),
+          }
+
+          // 同步填充 currentMastery（若未传入）
+          if (!serviceInput.currentMastery || serviceInput.currentMastery.length === 0) {
+            serviceInput.currentMastery = kps
+              .filter((kp) => masteryMap.has(kp.id))
+              .map((kp) => {
+                const m = masteryMap.get(kp.id)!
+                return {
+                  knowledgePoint: kp.name,
+                  masteryLevel: Math.round((m.masteryLevel / 100) * 5),
+                  errorCount: m.totalQuestions - m.correctQuestions,
+                }
+              })
+          }
+        }
+      } catch {
+        // 知识图谱获取失败不阻断主流程，降级为无图谱模式
+      }
+    }
+
    const service = createAiService(ctx.userId)
-    const result = await safeAiCall(() => service.recommendStudyPath(parsed.data))
+    const result = await safeAiCall(() => service.recommendStudyPath(serviceInput))
    if (!result.ok) {
      return { success: false, message: result.message }
    }
@@ -317,20 +367,9 @@ export async function getAiUsageStatsAction(): Promise<ActionState<AiUsageStats>
  try {
    await requirePermission(Permissions.AI_CONFIGURE)

-    // 当前从 trackEvent 的内存数据返回统计
-    // 生产环境应查询数据库或 Redis 聚合
-    const stats: AiUsageStats = {
-      totalCalls: 0,
-      callsToday: 0,
-      callsThisWeek: 0,
-      activeUsers: 0,
-      errorRate: 0,
-      avgDurationMs: 0,
-      byCapability: [],
-      byRole: [],
-      topUsers: [],
-      recentActivity: [],
-    }
+    // 从 data-access 层获取真实聚合统计
+    // 当前为内存事件存储（单实例），生产环境应查询 DB 或 Redis
+    const stats = await getAiUsageStats()

    return { success: true, data: stats }
  } catch (error) {
--- a/src/modules/ai/components/ai-assistant-widget.tsx
+++ b/src/modules/ai/components/ai-assistant-widget.tsx
@@ -72,7 +72,7 @@ export function AiAssistantWidget(): React.ReactNode {
          </span>
        </Button>
      </SheetTrigger>
-      <SheetContent className="w-full sm:max-w-[440px] overflow-y-auto p-0">
+      <SheetContent className="w-full sm:max-w-md overflow-y-auto p-0">
        <SheetHeader className="px-4 py-3 border-b">
          <div className="flex items-center justify-between">
            <SheetTitle className="flex items-center gap-2">
--- a/src/modules/ai/components/ai-chat-panel.tsx
+++ b/src/modules/ai/components/ai-chat-panel.tsx
@@ -5,6 +5,7 @@ import { useTranslations } from "next-intl"
 import { Send, Bot, User, Square, Trash2, Sparkles } from "lucide-react"
 import { toast } from "sonner"

+import { cn } from "@/shared/lib/utils"
 import { Button } from "@/shared/components/ui/button"
 import { Card, CardContent, CardHeader, CardTitle } from "@/shared/components/ui/card"
 import { Textarea } from "@/shared/components/ui/textarea"
@@ -31,8 +32,6 @@ type AiChatPanelProps = {
  title?: string
  /** 最大消息数 */
  maxMessages?: number
-  /** 是否启用流式响应（默认 true） */
-  streaming?: boolean
  /** 建议提示词列表（空状态展示） */
  suggestedPrompts?: string[]
 }
@@ -49,7 +48,7 @@ type AiChatPanelProps = {
 * - 清除对话按钮
 * - 建议提示词
 * - aria-live 无障碍
- * - 对话历史持久化（localStorage）
+ * - 对话历史持久化（localStorage，防抖写入）
 */
 export function AiChatPanel({
  systemPrompt,
@@ -57,7 +56,6 @@ export function AiChatPanel({
  placeholder,
  title,
  maxMessages = 50,
-  streaming: _streamingEnabled = true,
  suggestedPrompts,
 }: AiChatPanelProps): React.ReactNode {
  const t = useTranslations("ai")
@@ -65,33 +63,31 @@ export function AiChatPanel({
  const [input, setInput] = useState("")
  const scrollRef = useRef<HTMLDivElement>(null)
  const storageKey = "ai-chat-history"
+  const persistTimerRef = useRef<ReturnType<typeof setTimeout> | null>(null)

-  // 从 localStorage 恢复对话历史
+  // 持久化对话历史（防抖 500ms，避免流式过程中频繁写入）
  useEffect(() => {
-    try {
-      const stored = localStorage.getItem(storageKey)
-      if (stored) {
-        const parsed = JSON.parse(stored) as AiChatMessage[]
-        if (Array.isArray(parsed) && parsed.length > 0) {
-          // 通过 send 不合适，直接设置 messages 不支持
-          // 这里只是恢复显示，不重新发送
-        }
-      }
-    } catch {
-      // 忽略解析错误
-    }
-  }, [])
+    if (messages.length === 0) return
+    // 流式过程中不写入，流结束后再写
+    if (streaming) return

-  // 持久化对话历史
-  useEffect(() => {
+    if (persistTimerRef.current) {
+      clearTimeout(persistTimerRef.current)
+    }
+    persistTimerRef.current = setTimeout(() => {
      try {
-      if (messages.length > 0) {
        localStorage.setItem(storageKey, JSON.stringify(messages.slice(-20)))
-      }
      } catch {
        // 忽略写入错误
      }
-  }, [messages])
+    }, 500)
+
+    return () => {
+      if (persistTimerRef.current) {
+        clearTimeout(persistTimerRef.current)
+      }
+    }
+  }, [messages, streaming])

  // 自动滚动到底部
  useEffect(() => {
@@ -194,7 +190,7 @@ export function AiChatPanel({

        {messages.length > 0 ? (
          <ScrollArea
-            className="h-[300px] w-full rounded-md border p-3"
+            className="h-72 w-full rounded-md border p-3"
            aria-live="polite"
            aria-relevant="additions text"
          >
@@ -202,7 +198,10 @@ export function AiChatPanel({
              {messages.map((message, index) => (
                <div
                  key={index}
-                  className={`flex gap-2 ${message.role === "user" ? "justify-end" : "justify-start"}`}
+                  className={cn(
+                    "flex gap-2",
+                    message.role === "user" ? "justify-end" : "justify-start"
+                  )}
                >
                  {message.role === "assistant" ? (
                    <Bot className="h-5 w-5 shrink-0 text-primary mt-0.5" />
@@ -210,11 +209,13 @@ export function AiChatPanel({
                    <User className="h-5 w-5 shrink-0 text-muted-foreground mt-0.5" />
                  )}
                  <div
-                    className={`rounded-md px-3 py-2 text-sm max-w-[80%] ${
+                    className={cn(
+                      // 任意值：聊天气泡需按容器宽度百分比限制，固定 max-w-* 无法适应不同面板宽度
+                      "rounded-md px-3 py-2 text-sm max-w-[80%]",
                      message.role === "user"
                        ? "bg-primary text-primary-foreground"
                        : "bg-muted"
-                    }`}
+                    )}
                  >
                    {message.role === "assistant" ? (
                      <AiMarkdownRenderer content={message.content} />
@@ -266,7 +267,7 @@ export function AiChatPanel({
            onChange={(e) => setInput(e.target.value)}
            onKeyDown={handleKeyDown}
            placeholder={placeholder ?? t("chat.placeholder")}
-            className="min-h-[60px] resize-none"
+            className="min-h-16 resize-none"
            disabled={streaming || messages.length >= maxMessages}
            aria-label={t("chat.inputLabel")}
          />
--- a/src/modules/ai/components/ai-lesson-content-generator.tsx
+++ b/src/modules/ai/components/ai-lesson-content-generator.tsx
@@ -128,7 +128,7 @@ export function AiLessonContentGenerator({
              value={additionalContext}
              onChange={(e) => setAdditionalContext(e.target.value)}
              placeholder={t("lessonPrep.additionalContextPlaceholder")}
-              className="min-h-[60px] text-sm"
+              className="min-h-16 text-sm"
              maxLength={500}
            />
          </div>
--- a/src/modules/ai/components/ai-question-variant-generator.tsx
+++ b/src/modules/ai/components/ai-question-variant-generator.tsx
@@ -22,6 +22,12 @@ import type { QuestionVariantResult } from "@/modules/ai/types"

 type VariantType = "same_knowledge_point" | "different_difficulty" | "different_format"

+const VARIANT_TYPES: readonly VariantType[] = ["same_knowledge_point", "different_difficulty", "different_format"]
+
+const isVariantType = (value: string): value is VariantType => {
+  return VARIANT_TYPES.includes(value as VariantType)
+}
+
 type AiQuestionVariantGeneratorProps = {
  /** 原始题目 */
  originalQuestion: {
@@ -107,7 +113,11 @@ export function AiQuestionVariantGenerator({
            </label>
            <Select
              value={variantType}
-              onValueChange={(value) => setVariantType(value as VariantType)}
+              onValueChange={(value) => {
+                if (isVariantType(value)) {
+                  setVariantType(value)
+                }
+              }}
            >
              <SelectTrigger id="variant-type" className="w-full">
                <SelectValue placeholder={t("exam.generate")} />
--- a/src/modules/ai/components/ai-study-path.tsx
+++ b/src/modules/ai/components/ai-study-path.tsx
@@ -18,6 +18,8 @@ type AiStudyPathProps = {
  subject?: string
  currentMastery?: StudyPathInput["currentMastery"]
  learningGoal?: string
+  /** 教材 ID（传入后自动获取知识图谱注入路径推荐，对标 Squirrel AI） */
+  textbookId?: string
  onStartLearning?: (step: StudyPathResult["learningPath"][number]) => void
 }

@@ -30,12 +32,16 @@ type AiStudyPathProps = {
 * - 分步骤学习路径（含状态、建议、预计时间）
 * - 学习总结
 * - 鼓励语
+ *
+ * V3：支持 textbookId 参数，自动获取知识图谱与掌握度数据，
+ * 使 AI 沿前置依赖链推荐学习路径。
 */
 export function AiStudyPath({
  studentId,
  subject,
  currentMastery,
  learningGoal,
+  textbookId,
  onStartLearning,
 }: AiStudyPathProps): React.ReactNode {
  const t = useTranslations("ai")
@@ -55,6 +61,7 @@ export function AiStudyPath({
        subject,
        currentMastery,
        learningGoal,
+        textbookId,
      })
      if (result.success && result.data) {
        setPath(result.data)
--- a/src/modules/ai/data-access.ts
+++ b/src/modules/ai/data-access.ts
@@ -0,0 +1,138 @@
+import "server-only"
+
+import type { AiUsageStats } from "./types"
+
+/**
+ * AI 模块数据访问层
+ *
+ * 当前为内存实现（单实例），生产环境应替换为数据库或 Redis 查询。
+ *
+ * 职责：
+ * 1. AI 使用统计聚合（供管理员仪表盘）
+ * 2. AI 使用日志查询
+ *
+ * 注意：trackEvent 当前仅输出到 console（见 shared/lib/track-event.ts），
+ * 生产环境接入真实 sink（PostHog/Sentry/DB 事件表）后，
+ * 此处的聚合函数需改为查询真实数据源。
+ */
+
+// ---------------------------------------------------------------------------
+// 内存事件存储（生产环境替换为 DB/Redis）
+// ---------------------------------------------------------------------------
+
+interface StoredAiEvent {
+  userId: string
+  capability: string
+  success: boolean
+  durationMs: number
+  timestamp: number
+  errorMessage?: string
+}
+
+// 单实例内存存储，最多保留最近 10000 条事件
+const MAX_EVENTS = 10000
+const eventStore: StoredAiEvent[] = []
+
+/**
+ * 记录 AI 事件到内存存储
+ *
+ * 由 usage-tracker.ts 调用（非导出，仅模块内部使用）。
+ * 生产环境替换为 DB 异步写入。
+ */
+export function recordAiEvent(event: StoredAiEvent): void {
+  eventStore.push(event)
+  // 超出上限时丢弃最旧的事件
+  if (eventStore.length > MAX_EVENTS) {
+    eventStore.splice(0, eventStore.length - MAX_EVENTS)
+  }
+}
+
+// ---------------------------------------------------------------------------
+// 使用统计聚合
+// ---------------------------------------------------------------------------
+
+/**
+ * 获取 AI 使用统计
+ *
+ * 从内存事件存储聚合统计数据。
+ * 生产环境应改为查询 DB 或 Redis。
+ *
+ * @returns AiUsageStats — 聚合后的统计数据
+ */
+export async function getAiUsageStats(): Promise<AiUsageStats> {
+  const now = Date.now()
+  const todayStart = new Date()
+  todayStart.setHours(0, 0, 0, 0)
+  const todayMs = todayStart.getTime()
+  const weekMs = now - 7 * 24 * 60 * 60 * 1000
+
+  const totalCalls = eventStore.length
+  const callsToday = eventStore.filter((e) => e.timestamp >= todayMs).length
+  const callsThisWeek = eventStore.filter((e) => e.timestamp >= weekMs).length
+  const failedCalls = eventStore.filter((e) => !e.success).length
+  const errorRate = totalCalls > 0 ? failedCalls / totalCalls : 0
+
+  const durations = eventStore.map((e) => e.durationMs)
+  const avgDurationMs = durations.length > 0
+    ? Math.round(durations.reduce((a, b) => a + b, 0) / durations.length)
+    : 0
+
+  // 活跃用户（今日有调用的去重用户数）
+  const activeUserIds = new Set(
+    eventStore
+      .filter((e) => e.timestamp >= todayMs)
+      .map((e) => e.userId)
+  )
+
+  // 按 capability 聚合
+  const capabilityMap = new Map<string, number>()
+  for (const e of eventStore) {
+    capabilityMap.set(e.capability, (capabilityMap.get(e.capability) ?? 0) + 1)
+  }
+  const byCapability = Array.from(capabilityMap.entries())
+    .map(([capability, count]) => ({
+      capability,
+      count,
+      percentage: totalCalls > 0 ? Math.round((count / totalCalls) * 100) : 0,
+    }))
+    .sort((a, b) => b.count - a.count)
+
+  // 按角色聚合（当前无法从事件中获取角色，返回空数组）
+  // 生产环境若需按角色聚合，需在 StoredAiEvent 中增加 role 字段
+  const byRole: Array<{ role: string; count: number }> = []
+
+  // Top 用户
+  const userCallMap = new Map<string, number>()
+  for (const e of eventStore) {
+    userCallMap.set(e.userId, (userCallMap.get(e.userId) ?? 0) + 1)
+  }
+  const topUsers = Array.from(userCallMap.entries())
+    .map(([userId, count]) => ({ userId, count }))
+    .sort((a, b) => b.count - a.count)
+    .slice(0, 10)
+
+  // 最近活动
+  const recentActivity = eventStore
+    .slice(-20)
+    .reverse()
+    .map((e) => ({
+      userId: e.userId,
+      capability: e.capability,
+      success: e.success,
+      durationMs: e.durationMs,
+      timestamp: new Date(e.timestamp).toISOString(),
+    }))
+
+  return {
+    totalCalls,
+    callsToday,
+    callsThisWeek,
+    activeUsers: activeUserIds.size,
+    errorRate: Math.round(errorRate * 100) / 100,
+    avgDurationMs,
+    byCapability,
+    byRole,
+    topUsers,
+    recentActivity,
+  }
+}
--- a/src/modules/ai/hooks/stream-utils.ts
+++ b/src/modules/ai/hooks/stream-utils.ts
@@ -0,0 +1,135 @@
+/**
+ * SSE 流式响应解析与处理工具
+ *
+ * 从 use-ai-chat-stream hook 中抽取的纯函数，便于单独测试与复用。
+ */
+
+import type { AiChatMessage } from "../types"
+
+/** SSE 事件类型 */
+export type SseEvent =
+  | { type: "token"; content: string }
+  | { type: "error"; message: string }
+  | { type: "filtered" }
+  | { type: "socratic_warning"; message: string }
+
+/**
+ * 从 Response 中读取并解析 SSE 事件流
+ *
+ * @param response - fetch 返回的 Response 对象
+ * @param onEvent - 每个解析出的事件回调
+ */
+export async function consumeSseStream(
+  response: Response,
+  onEvent: (event: SseEvent) => void
+): Promise<void> {
+  const reader = response.body?.getReader()
+  if (!reader) return
+
+  const decoder = new TextDecoder()
+  let buffer = ""
+
+  while (true) {
+    const { done, value } = await reader.read()
+    if (done) break
+
+    buffer += decoder.decode(value, { stream: true })
+    const lines = buffer.split("\n")
+    buffer = lines.pop() ?? ""
+
+    for (const line of lines) {
+      if (!line.startsWith("data: ")) continue
+      const data = line.slice(6).trim()
+      if (data === "[DONE]") continue
+
+      try {
+        const parsed = JSON.parse(data) as {
+          type: "token" | "error" | "filtered" | "socratic_warning"
+          content?: string
+          message?: string
+        }
+
+        if (parsed.type === "token" && parsed.content) {
+          onEvent({ type: "token", content: parsed.content })
+        } else if (parsed.type === "error") {
+          onEvent({ type: "error", message: parsed.message ?? "Unknown error" })
+        } else if (parsed.type === "filtered") {
+          onEvent({ type: "filtered" })
+        } else if (parsed.type === "socratic_warning") {
+          onEvent({ type: "socratic_warning", message: parsed.message ?? "Socratic warning" })
+        }
+      } catch {
+        // 忽略解析错误
+      }
+    }
+  }
+}
+
+/** 流式错误对应的 i18n key（相对于 "ai" 命名空间） */
+export type StreamErrorKey =
+  | "safety.dailyLimit"
+  | "error.unauthorized"
+  | "safety.blocked"
+  | "error.chatFailed"
+
+/**
+ * 根据 HTTP 状态码映射错误消息 i18n key
+ *
+ * 返回值用于 next-intl 的 t() 调用（相对于 "ai" 命名空间）。
+ */
+export function getStreamErrorKey(status: number): StreamErrorKey {
+  if (status === 429) return "safety.dailyLimit"
+  if (status === 403) return "error.unauthorized"
+  if (status === 400) return "safety.blocked"
+  return "error.chatFailed"
+}
+
+/**
+ * 从错误响应体中提取错误消息
+ */
+export async function extractErrorMessage(
+  response: Response,
+  fallback: string
+): Promise<string> {
+  try {
+    const errorText = await response.text()
+    const errorData = JSON.parse(errorText) as { message?: string }
+    return errorData.message ?? fallback
+  } catch {
+    return fallback
+  }
+}
+
+/**
+ * 移除消息列表末尾空的 assistant 消息
+ *
+ * 用于流式失败/中止时清理占位消息。
+ */
+export function removeTrailingEmptyAssistant(
+  messages: AiChatMessage[]
+): AiChatMessage[] {
+  const filtered = [...messages]
+  const last = filtered[filtered.length - 1]
+  if (last && last.role === "assistant" && last.content === "") {
+    filtered.pop()
+  }
+  return filtered
+}
+
+/**
+ * 向消息列表末尾追加 token 到 assistant 消息
+ */
+export function appendTokenToLastAssistant(
+  messages: AiChatMessage[],
+  token: string
+): AiChatMessage[] {
+  const updated = [...messages]
+  const last = updated[updated.length - 1]
+  if (last && last.role === "assistant") {
+    updated[updated.length - 1] = {
+      ...last,
+      content: last.content + token,
+    }
+  }
+  return updated
+}
--- a/src/modules/ai/hooks/use-ai-chat-stream.ts
+++ b/src/modules/ai/hooks/use-ai-chat-stream.ts
@@ -3,24 +3,24 @@
 import { useState, useCallback, useRef } from "react"
 import { useTranslations } from "next-intl"
 import type { AiChatMessage } from "../types"
+import {
+  consumeSseStream,
+  getStreamErrorKey,
+  extractErrorMessage,
+  removeTrailingEmptyAssistant,
+  appendTokenToLastAssistant,
+} from "./stream-utils"

 /**
 * AI 流式聊天 Hook
 *
 * 通过 SSE 端点消费流式 AI 回复。
- * 支持：
- * - 逐 token 渲染
- * - 停止生成（AbortController）
- * - 错误处理
+ * 支持：逐 token 渲染、停止生成（AbortController）、错误处理。
 */
-
-type StreamState = {
+type UseAiChatStreamReturn = {
  messages: AiChatMessage[]
  streaming: boolean
  error: string | null
-}
-
-type UseAiChatStreamReturn = StreamState & {
  send: (messages: AiChatMessage[], options?: { systemPrompt?: string; providerId?: string }) => Promise<void>
  stop: () => void
  clear: () => void
@@ -46,8 +46,6 @@ export function useAiChatStream(): UseAiChatStreamReturn {
      if (userMessage && userMessage.role === "user") {
        setMessages((prev) => [...prev, userMessage])
      }
-
-      // 添加空的 assistant 消息，用于流式更新
      setMessages((prev) => [...prev, { role: "assistant", content: "" }])

      const controller = new AbortController()
@@ -66,119 +64,42 @@ export function useAiChatStream(): UseAiChatStreamReturn {
        })

        if (!response.ok) {
-          const errorText = await response.text()
-          let errorMessage = t("error.chatFailed")
-          try {
-            const errorData = JSON.parse(errorText) as { message?: string }
-            errorMessage = errorData.message ?? errorMessage
-          } catch {
-            // 使用默认错误消息
-          }
-          if (response.status === 429) {
-            errorMessage = t("safety.dailyLimit")
-          } else if (response.status === 403) {
-            errorMessage = t("error.unauthorized")
-          } else if (response.status === 400) {
-            errorMessage = t("safety.blocked")
-          }
+          const fallback = t(getStreamErrorKey(response.status))
+          const errorMessage = await extractErrorMessage(response, fallback)
          setError(errorMessage)
-          // 移除空的 assistant 消息
-          setMessages((prev) => {
-            const filtered = [...prev]
-            const last = filtered[filtered.length - 1]
-            if (last && last.role === "assistant" && last.content === "") {
-              filtered.pop()
-            }
-            return filtered
-          })
+          setMessages((prev) => removeTrailingEmptyAssistant(prev))
          return
        }

-        const reader = response.body?.getReader()
-        if (!reader) {
-          setError(t("error.chatFailed"))
-          return
-        }
-
-        const decoder = new TextDecoder()
-        let buffer = ""
-
-        while (true) {
-          const { done, value } = await reader.read()
-          if (done) break
-
-          buffer += decoder.decode(value, { stream: true })
-          const lines = buffer.split("\n")
-          buffer = lines.pop() ?? ""
-
-          for (const line of lines) {
-            if (!line.startsWith("data: ")) continue
-            const data = line.slice(6).trim()
-            if (data === "[DONE]") continue
-
-            try {
-              const parsed = JSON.parse(data) as {
-                type: "token" | "error" | "filtered"
-                content?: string
-                message?: string
-              }
-
-              if (parsed.type === "token" && parsed.content) {
-                setMessages((prev) => {
-                  const updated = [...prev]
-                  const last = updated[updated.length - 1]
-                  if (last && last.role === "assistant") {
-                    updated[updated.length - 1] = {
-                      ...last,
-                      content: last.content + parsed.content,
-                    }
-                  }
-                  return updated
-                })
-              } else if (parsed.type === "error") {
-                setError(parsed.message ?? t("error.chatFailed"))
-                setMessages((prev) => {
-                  const filtered = [...prev]
-                  const last = filtered[filtered.length - 1]
-                  if (last && last.role === "assistant" && last.content === "") {
-                    filtered.pop()
-                  }
-                  return filtered
-                })
-              } else if (parsed.type === "filtered") {
+        await consumeSseStream(response, (event) => {
+          if (event.type === "token") {
+            setMessages((prev) => appendTokenToLastAssistant(prev, event.content))
+          } else if (event.type === "error") {
+            setError(event.message)
+            setMessages((prev) => removeTrailingEmptyAssistant(prev))
+          } else if (event.type === "filtered") {
            setError(t("safety.contentFiltered"))
+          } else if (event.type === "socratic_warning") {
+            // 苏格拉底式辅导警告：不阻断，仅提示
+            // 可通过 toast 或 UI 标记展示，此处暂存 error 供组件判断
+            setError(event.message)
          }
-            } catch {
-              // 忽略解析错误
-            }
-          }
-        }
+        })
      } catch (err) {
-        if (err instanceof DOMException && err.name === "AbortError") {
-          // 用户主动停止，不显示错误
-        } else {
+        if (!(err instanceof DOMException && err.name === "AbortError")) {
          setError(err instanceof Error ? err.message : String(err))
        }
      } finally {
        setStreaming(false)
        abortControllerRef.current = null
-        // 清理空的 assistant 消息
-        setMessages((prev) => {
-          const last = prev[prev.length - 1]
-          if (last && last.role === "assistant" && last.content === "") {
-            return prev.slice(0, -1)
-          }
-          return prev
-        })
+        setMessages((prev) => removeTrailingEmptyAssistant(prev))
      }
    },
    [streaming, t]
  )

  const stop = useCallback((): void => {
-    if (abortControllerRef.current) {
-      abortControllerRef.current.abort()
-    }
+    abortControllerRef.current?.abort()
  }, [])

  const clear = useCallback((): void => {
--- a/src/modules/ai/schema.ts
+++ b/src/modules/ai/schema.ts
@@ -187,6 +187,28 @@ export const StudyPathInputSchema = z.object({
    )
    .optional(),
  learningGoal: z.string().optional(),
+  /** 教材 ID（传入后 action 层自动获取知识图谱注入） */
+  textbookId: z.string().optional(),
+  /** 知识图谱（可直接传入，优先于 textbookId 自动获取） */
+  knowledgeGraph: z
+    .object({
+      nodes: z.array(
+        z.object({
+          id: z.string().min(1),
+          name: z.string().min(1),
+          level: z.number().int().min(0),
+          masteryLevel: z.number().optional(),
+        })
+      ),
+      edges: z.array(
+        z.object({
+          from: z.string().min(1),
+          to: z.string().min(1),
+          type: z.literal("prerequisite"),
+        })
+      ),
+    })
+    .optional(),
 })

 export const StudyPathResultSchema = z.object({
--- a/src/modules/ai/services/ai-service.ts
+++ b/src/modules/ai/services/ai-service.ts
@@ -330,10 +330,10 @@ export class DefaultAiService implements AiService {
  }

  async generateChildSummary(input: ChildSummaryInput): Promise<ChildSummaryResult> {
-    return withAiTracking(this.userId, "weakness_analysis", undefined, async () => {
+    return withAiTracking(this.userId, "child_summary", undefined, async () => {
+      // PII 最小化：不传学生真实姓名，用 ID 替代（COPPA/FERPA 合规）
      const userLines = [
        `Student ID: ${input.studentId}`,
-        input.studentName ? `Student Name: ${input.studentName}` : "",
        input.grade ? `Grade: ${input.grade}` : "",
        input.recentGrades && input.recentGrades.length > 0
          ? `Recent Grades:\n${JSON.stringify(input.recentGrades, null, 2)}`
@@ -370,7 +370,7 @@ export class DefaultAiService implements AiService {
  }

  async recommendStudyPath(input: StudyPathInput): Promise<StudyPathResult> {
-    return withAiTracking(this.userId, "weakness_analysis", undefined, async () => {
+    return withAiTracking(this.userId, "study_path", undefined, async () => {
      const userLines = [
        `Student ID: ${input.studentId}`,
        input.subject ? `Subject: ${input.subject}` : "",
@@ -379,6 +379,21 @@ export class DefaultAiService implements AiService {
          : "",
        input.learningGoal ? `Learning Goal: ${input.learningGoal}` : "",
      ].filter((line) => line.length > 0)
+
+      // 知识图谱上下文注入（V3：对标 Squirrel AI 纳米级知识图谱）
+      if (input.knowledgeGraph && input.knowledgeGraph.nodes.length > 0) {
+        const graphLines = [
+          "Knowledge Graph:",
+          "Nodes (id | name | level | mastery 0-100):",
+          ...input.knowledgeGraph.nodes.map(
+            (n) => `  ${n.id} | ${n.name} | L${n.level} | ${n.masteryLevel ?? "unassessed"}`
+          ),
+          "Prerequisite edges (from -> to, meaning 'from' must be mastered before 'to'):",
+          ...input.knowledgeGraph.edges.map((e) => `  ${e.from} -> ${e.to}`),
+        ]
+        userLines.push(graphLines.join("\n"))
+      }
+
      const { content } = await callAi(
        buildChatMessages(STUDY_PATH_SYSTEM_PROMPT, userLines.join("\n\n")),
        { temperature: 0.5, maxTokens: 2000 }
--- a/src/modules/ai/services/content-safety.ts
+++ b/src/modules/ai/services/content-safety.ts
@@ -6,9 +6,11 @@ import "server-only"
 * 多层防护：
 * 1. 输入过滤：检查用户输入是否包含不当内容
 * 2. 输出过滤：检查 AI 回复是否包含不当内容
- * 3. 每日限制：按用户 + 日期计数
+ * 3. 每日限制：按用户 + 日期计数（原子操作，防 TOCTOU 竞态）
 *
 * 参考 Khanmigo 的多层 moderation 模式。
+ *
+ * 注意：当前为内存实现，多实例部署需替换为 Redis（INCR + EXPIRE）。
 */

 // ---------------------------------------------------------------------------
@@ -34,6 +36,10 @@ const BLOCKED_OUTPUT_PATTERNS: readonly RegExp[] = [
 const STUDENT_BLOCKED_PATTERNS: readonly RegExp[] = [
  // 学生侧额外限制：禁止直接给出作业答案
  /\b(here is the (complete )?answer|the answer is:?)\b/i,
+  // 强化：匹配"答案是 X" / "正确答案是 X" / "final answer: X"
+  /\b(the (correct )?answer is\s*[:：]?\s*[A-Z\d])/i,
+  /\bfinal answer[:：]\s*\S+/i,
+  /\b答案(是|应该为|为)\s*[:：]?\s*[A-F\d]/i,
 ]

 // ---------------------------------------------------------------------------
@@ -61,7 +67,6 @@ export const filterUserInput = (
  }

  if (options?.isStudent) {
-    // 学生侧额外检查
    for (const pattern of STUDENT_BLOCKED_PATTERNS) {
      if (pattern.test(text)) {
        return {
@@ -109,7 +114,7 @@ export const filterAiOutput = (
 }

 // ---------------------------------------------------------------------------
-// 每日限制
+// 每日限制（原子操作，防 TOCTOU 竞态）
 // ---------------------------------------------------------------------------

 const DAILY_LIMITS: Record<string, number> = {
@@ -124,10 +129,10 @@ export const getDailyLimit = (role: string): number => {
 }

 /**
- * 检查用户今日 AI 使用次数
+ * 每日使用计数（内存实现，多实例需替换为 Redis）
 *
- * 生产环境应接入 Redis 或数据库计数器。
- * 当前实现为内存映射（单实例场景），多实例需替换为 Redis。
+ * 注意：当前为单实例内存映射，多实例部署下每个实例独立计数，
+ * 实际可用次数 = 限额 × 实例数。生产环境应接入 Redis INCR + EXPIRE。
 */
 const dailyUsageMap = new Map<string, { date: string; count: number }>()

@@ -171,3 +176,116 @@ export const incrementDailyUsage = (userId: string): void => {
    }
  }
 }
+
+/**
+ * 原子化检查并递增每日使用计数
+ *
+ * 解决 checkDailyLimit + incrementDailyUsage 分离导致的 TOCTOU 竞态：
+ * 并发请求在限额临界点同时通过检查，导致超额。
+ *
+ * 此函数在一次调用内完成「递增 + 判断是否超限」，
+ * 若递增后超过限额，回滚计数并返回 blocked。
+ *
+ * @returns { blocked, currentCount, limit } — blocked 为 true 表示已超限
+ */
+export const tryConsumeDailyQuota = (
+  userId: string,
+  role: string
+): { blocked: boolean; currentCount: number; limit: number } => {
+  const today = new Date().toISOString().slice(0, 10)
+  const key = `${userId}:${today}`
+  const limit = getDailyLimit(role)
+  const current = dailyUsageMap.get(key)
+
+  // 原子递增
+  const newCount = current && current.date === today ? current.count + 1 : 1
+  dailyUsageMap.set(key, { date: today, count: newCount })
+
+  // 清理过期条目
+  if (dailyUsageMap.size > 10000) {
+    for (const [k, v] of dailyUsageMap.entries()) {
+      if (v.date !== today) {
+        dailyUsageMap.delete(k)
+      }
+    }
+  }
+
+  if (newCount > limit) {
+    // 超限，回滚计数（不惩罚用户因竞态多出的尝试）
+    dailyUsageMap.set(key, { date: today, count: limit })
+    return { blocked: true, currentCount: limit, limit }
+  }
+
+  return { blocked: false, currentCount: newCount, limit }
+}
+
+/**
+ * 回退每日使用计数（当 AI 调用失败或内容被过滤时调用）
+ *
+ * 确保用户不会因 AI 输出被过滤或调用失败而损失配额。
+ */
+export const refundDailyQuota = (userId: string): void => {
+  const today = new Date().toISOString().slice(0, 10)
+  const key = `${userId}:${today}`
+  const current = dailyUsageMap.get(key)
+
+  if (current && current.date === today && current.count > 0) {
+    current.count -= 1
+  }
+}
+
+// ---------------------------------------------------------------------------
+// 苏格拉底式辅导输出校验
+// ---------------------------------------------------------------------------
+
+export type SocraticValidationResult = {
+  valid: boolean
+  reason?: string
+}
+
+/**
+ * 校验 AI 回复是否符合苏格拉底式辅导原则
+ *
+ * 规则：
+ * 1. 回复必须以问号结尾（中英文均可）
+ * 2. 不得包含超过 2 句连续陈述句而不提问
+ * 3. 不得直接给出最终答案（复用 STUDENT_BLOCKED_PATTERNS）
+ *
+ * 用于学生侧 AI 对话，强制引导式教学。
+ */
+export const validateSocraticOutput = (content: string): SocraticValidationResult => {
+  const text = String(content ?? "").trim()
+
+  if (!text) {
+    return { valid: false, reason: "Empty response" }
+  }
+
+  // 检查是否直接给出答案
+  for (const pattern of STUDENT_BLOCKED_PATTERNS) {
+    if (pattern.test(text)) {
+      return { valid: false, reason: "Response contains direct answer" }
+    }
+  }
+
+  // 检查是否以问号结尾
+  if (!/[?？]$/.test(text)) {
+    return { valid: false, reason: "Response must end with a question" }
+  }
+
+  // 检查连续陈述句数量（按句号/感叹号分割）
+  const sentences = text.split(/[。！？.!?]/).filter((s) => s.trim().length > 0)
+  let consecutiveStatements = 0
+  for (const sentence of sentences) {
+    // 如果句子本身是疑问句（包含 ? 或 ？），重置计数
+    if (/[?？]/.test(sentence)) {
+      consecutiveStatements = 0
+    } else {
+      consecutiveStatements += 1
+      if (consecutiveStatements > 2) {
+        return { valid: false, reason: "Too many consecutive statements without a question" }
+      }
+    }
+  }
+
+  return { valid: true }
+}
--- a/src/modules/ai/services/prompt-templates.ts
+++ b/src/modules/ai/services/prompt-templates.ts
@@ -126,6 +126,7 @@ export const WEAKNESS_ANALYSIS_SYSTEM_PROMPT = [
  "    {",
  '      "area": "knowledge area name",',
  '      "severity": "high | medium | low",',
+  '      "rootCause": "underlying reason, e.g. missing prerequisite",',
  '      "suggestion": "specific improvement suggestion"',
  "    }",
  "  ],",
@@ -135,6 +136,7 @@ export const WEAKNESS_ANALYSIS_SYSTEM_PROMPT = [
  "Rules:",
  "- Identify 2-5 weak areas based on error frequency and mastery level.",
  "- severity: high = mastery < 2, medium = mastery 2-3, low = mastery 3-4.",
+  "- If prerequisite knowledge is provided and a prerequisite mastery < 2, list the prerequisite as the rootCause.",
  "- Suggestions should be actionable and specific.",
  "- Study plan should be concise (3-5 sentences).",
  "- Recommended resources can be topic names or study strategies.",
@@ -165,6 +167,22 @@ export const CHAT_SYSTEM_PROMPT = [
  "Be concise, accurate, and pedagogically sound.",
 ].join("\n")

+// ---------------------------------------------------------------------------
+// 苏格拉底式辅导（学生专用，强制引导式教学）
+// ---------------------------------------------------------------------------
+
+export const SOCRATIC_TUTOR_SYSTEM_PROMPT = [
+  "You are a Socratic tutor for K12 students.",
+  "STRICT RULES (never violate):",
+  "- NEVER output the final answer directly.",
+  "- NEVER output more than 2 consecutive sentences without asking a question.",
+  "- Use a 3-tier hint escalation: Tier 1 (conceptual question) → Tier 2 (concrete hint) → Tier 3 (worked example without the final step).",
+  "- If the student asks for the answer 3+ times, explain why guided discovery is better for learning.",
+  "- Always end your response with a question that moves the student forward.",
+  "- Track the student's reasoning and point out the exact step where they went wrong.",
+  "- Respond in the student's language (Chinese by default).",
+].join("\n")
+
 // ---------------------------------------------------------------------------
 // 家长学情摘要
 // ---------------------------------------------------------------------------
@@ -196,7 +214,7 @@ export const CHILD_SUMMARY_SYSTEM_PROMPT = [

 export const STUDY_PATH_SYSTEM_PROMPT = [
  "You are an expert K12 adaptive learning path designer.",
-  "Based on the student's current mastery levels, recommend a personalized learning path.",
+  "Based on the student's current mastery levels and knowledge graph, recommend a personalized learning path.",
  "Return JSON only without markdown.",
  "Output schema:",
  "{",
@@ -215,7 +233,8 @@ export const STUDY_PATH_SYSTEM_PROMPT = [
  "}",
  "Rules:",
  "- Order learning path from foundational to advanced.",
-  "- Prioritize weak areas (mastery < 2) first.",
+  "- MUST follow prerequisite chains: if a knowledge point has unmastered prerequisites, list the prerequisites first.",
+  "- Prioritize weak areas (mastery < 2) first, but only after their prerequisites are addressed.",
  "- Include 3-7 steps in the learning path.",
  "- estimatedTime should be realistic (5-30 min per step).",
  "- motivation should be age-appropriate and encouraging.",
--- a/src/modules/ai/services/usage-tracker.ts
+++ b/src/modules/ai/services/usage-tracker.ts
@@ -1,10 +1,11 @@
 import "server-only"

 import { trackEvent, type EventName } from "@/shared/lib/track-event"
+import { recordAiEvent } from "../data-access"

 export type AiUsageEvent = {
  userId: string
-  capability: "chat" | "similar_question" | "grading_assist" | "lesson_content" | "question_variant" | "weakness_analysis"
+  capability: "chat" | "similar_question" | "grading_assist" | "lesson_content" | "question_variant" | "weakness_analysis" | "child_summary" | "study_path"
  providerId?: string
  model?: string
  success: boolean
@@ -20,16 +21,31 @@ const AI_EVENT_MAP: Record<AiUsageEvent["capability"], EventName> = {
  lesson_content: "ai.lesson_content",
  question_variant: "ai.question_variant",
  weakness_analysis: "ai.weakness_analysis",
+  child_summary: "ai.child_summary",
+  study_path: "ai.study_path",
 }

 /**
 * AI 使用埋点
 *
 * 记录每次 AI 调用的元数据，用于监控、成本分析与异常排查。
+ * 同时写入 data-access 层的内存事件存储（供管理员仪表盘聚合查询）。
 * 非阻塞，失败不影响主流程。
 */
 export const trackAiUsage = (event: AiUsageEvent): void => {
  const eventName = AI_EVENT_MAP[event.capability]
+
+  // 写入 data-access 层（供 getAiUsageStats 聚合）
+  recordAiEvent({
+    userId: event.userId,
+    capability: event.capability,
+    success: event.success,
+    durationMs: event.durationMs,
+    timestamp: Date.now(),
+    errorMessage: event.errorMessage,
+  })
+
+  // 写入全局 trackEvent（供外部监控系统）
  void trackEvent({
    event: eventName,
    userId: event.userId,
--- a/src/modules/ai/types.ts
+++ b/src/modules/ai/types.ts
@@ -168,6 +168,23 @@ export type StudyPathInput = {
    errorCount: number
  }>
  learningGoal?: string
+  /** 教材 ID（传入后 action 层自动获取知识图谱注入） */
+  textbookId?: string
+  /**
+   * 知识图谱上下文（V3 新增，对标 Squirrel AI 纳米级知识图谱）
+   *
+   * 由 actions 层从 textbooks/data-access-graph 获取后注入，
+   * 使 AI 能沿前置依赖链推荐学习路径。
+   */
+  knowledgeGraph?: {
+    nodes: Array<{
+      id: string
+      name: string
+      level: number
+      masteryLevel?: number
+    }>
+    edges: Array<{ from: string; to: string; type: "prerequisite" }>
+  }
 }

 /** 学习路径推荐结果 */
--- a/src/shared/lib/ai/client.ts
+++ b/src/shared/lib/ai/client.ts
@@ -5,15 +5,68 @@ import OpenAI from "openai"
 import { extractMessageContent, type AiChatRequest } from "./payload-parser"
 import { getAiProviderConfig } from "./provider-config"

-const getAiClient = async (config: { apiKey: string; baseUrl?: string }) => {
+/** AI 请求超时（毫秒） */
+const AI_TIMEOUT_MS = 30000
+
+/** 可重试的 HTTP 状态码（429 限流 + 5xx 服务端错误） */
+const RETRYABLE_STATUS_CODES = new Set([429, 500, 502, 503, 504])
+
+/** 最大重试次数 */
+const MAX_RETRIES = 2
+
+/** 基础重试延迟（毫秒），实际延迟 = base * 2^attempt */
+const RETRY_BASE_DELAY_MS = 1000
+
+const getAiClient = async (config: { apiKey: string; baseUrl?: string }): Promise<OpenAI> => {
  const baseUrl = String(config.baseUrl ?? "https://api.openai.com").replace(/\/+$/, "")
  return new OpenAI({
    apiKey: config.apiKey,
    baseURL: baseUrl.length ? baseUrl : undefined,
+    timeout: AI_TIMEOUT_MS,
+    maxRetries: 0, // 由业务层控制重试
  })
 }

-export const testAiProviderConfig = async (input: { apiKey: string; baseUrl?: string; model: string }) => {
+/**
+ * 判断错误是否可重试（429 限流或 5xx 服务端错误）
+ */
+const isRetryableError = (error: unknown): boolean => {
+  if (error instanceof OpenAI.APIError) {
+    return RETRYABLE_STATUS_CODES.has(error.status ?? 0)
+  }
+  // 网络超时/连接错误也可重试
+  if (error instanceof Error) {
+    const msg = error.message.toLowerCase()
+    return msg.includes("timeout") || msg.includes("econnreset") || msg.includes("socket hang up")
+  }
+  return false
+}
+
+/**
+ * 指数退避重试包装器
+ *
+ * 仅对 429（限流）和 5xx（服务端错误）重试，4xx 不重试。
+ * 最大重试 2 次，延迟 1s/2s/4s。
+ */
+async function withRetry<T>(fn: () => Promise<T>): Promise<T> {
+  let lastError: unknown
+  for (let attempt = 0; attempt <= MAX_RETRIES; attempt++) {
+    try {
+      return await fn()
+    } catch (error) {
+      lastError = error
+      if (attempt < MAX_RETRIES && isRetryableError(error)) {
+        const delay = RETRY_BASE_DELAY_MS * Math.pow(2, attempt)
+        await new Promise((resolve) => setTimeout(resolve, delay))
+        continue
+      }
+      throw error
+    }
+  }
+  throw lastError
+}
+
+export const testAiProviderConfig = async (input: { apiKey: string; baseUrl?: string; model: string }): Promise<boolean> => {
  const client = await getAiClient({ apiKey: input.apiKey, baseUrl: input.baseUrl })
  const result = await client.chat.completions.create({
    model: input.model,
@@ -29,7 +82,7 @@ export const testAiProviderConfig = async (input: { apiKey: string; baseUrl?: st
 export const testAiProviderById = async (
  providerId: string,
  overrides?: { baseUrl?: string; model?: string }
-) => {
+): Promise<boolean> => {
  const config = await getAiProviderConfig(providerId)
  const client = await getAiClient({ apiKey: config.apiKey, baseUrl: overrides?.baseUrl ?? config.baseUrl })
  const result = await client.chat.completions.create({
@@ -43,18 +96,19 @@ export const testAiProviderById = async (
  return true
 }

-export const createAiChatCompletion = async (input: AiChatRequest) => {
+export const createAiChatCompletion = async (input: AiChatRequest): Promise<{ content: string; usage: unknown }> => {
  const config = await getAiProviderConfig(input.providerId)
  const client = await getAiClient(config)
-  const result = (await client.chat.completions.create({
+
+  const result = await withRetry(() =>
+    client.chat.completions.create({
      model: config.model || input.model,
      messages: input.messages,
      temperature: input.temperature,
      ...(typeof input.maxTokens === "number" ? { max_tokens: input.maxTokens } : {}),
      ...(input.thinking ? { thinking: input.thinking } : {}),
-  } as Parameters<typeof client.chat.completions.create>[0])) as Awaited<
-    ReturnType<typeof client.chat.completions.create>
-  >
+    } as Parameters<typeof client.chat.completions.create>[0])
+  ) as Awaited<ReturnType<typeof client.chat.completions.create>>

  const hasChoices = "choices" in result && Array.isArray(result.choices) && result.choices.length > 0
  if (!hasChoices) throw new Error("Empty response from provider. Check API URL, model, and API key.")
@@ -71,6 +125,9 @@ export const createAiChatCompletion = async (input: AiChatRequest) => {
 *
 * 返回 AsyncGenerator，逐 token 产出内容。
 * 用于 SSE 流式响应，降低用户感知延迟。
+ *
+ * 注意：流式调用不使用 withRetry，因为流一旦开始无法重试。
+ * 超时由 OpenAI SDK 的 timeout 配置控制。
 */
 export async function* createAiChatCompletionStream(
  input: AiChatRequest
--- a/src/shared/lib/track-event.ts
+++ b/src/shared/lib/track-event.ts
@@ -27,6 +27,7 @@ export type EventName =
  | "notification.marked_all_read"
  | "notification.sent"
  | "notification.send_failed"
+  | "notification.archived"
  | "attendance.recorded"
  | "attendance.batch_recorded"
  | "attendance.updated"
@@ -66,6 +67,8 @@ export type EventName =
  | "ai.lesson_content"
  | "ai.question_variant"
  | "ai.weakness_analysis"
+  | "ai.child_summary"
+  | "ai.study_path"

 /** 埋点事件负载 */
 export interface TrackEventPayload {