feat(llm-trace): 提示词追踪增强与缺陷修复

新增追踪字段与落库链路：
- FAQ 命中详情（faqId/faqQuestion/faqMatchType/faqScore）
- 错误分类（errorType/errorMessage）与 Token 用量三档
- MCP 工具调用事件（toolCallsJson）、多轮历史（historyMessagesJson/historyTurns）
- RAG 命中片段（ragHitsJson）、检索模式（searchMode）、账户/API Key
- 数据库迁移新增 13 列 + 3 索引，同步 init-database.sql

前端增强：
- 筛选新增 API Key / 错误类型下拉，列表新增 RAG 列与刷新按钮
- 详情 Tab 扩展 FAQ / 工具调用 / 对话上下文，Prompt 关键词搜索高亮
- 新增会话时间线抽屉、调用趋势图（Chart.js 双轴折线）

缺陷修复：
- MCP 工具事件采集改用 ToolContext 显式收集器，修复 Reactor 流式跨线程丢 ThreadLocal 导致 toolCallsJson 为空
- 趋势统计 GROUP BY 别名改为 date_trunc 表达式，修复 PostgreSQL 语法错误
- 详情表格移除重复 row-key，新增 /error-types 与 /trend 接口

CLAUDE.md

@@ -214,6 +214,7 @@ catch (e) { toast('操作失败', 'error') }
 - FAQ 管理: `/faq/*`（`FaqController`）
 - SDK 认证: `/open-api/auth/*`（`AuthController`，Token 换取）
 - API Key 角色绑定: `/api-key/{id}/roles`（`ApiKeyController`，admin 角色）
+- LLM 调用追踪: `/llm-trace/*`（`LlmCallTraceController`，admin 角色）
 
 ### Filter 优先级
 
@@ -263,6 +264,15 @@ catch (e) { toast('操作失败', 'error') }
 - **vector_store 全文检索**: 新增 `content_tsvector` 列 + GIN 索引 + PostgreSQL 触发器自动维护
 - **前端**: `DocSearch.js` 增加检索模式下拉选择（向量/关键词/混合），结果标注来源模式
 
+## LLM 调用追踪（提示词调试）
+
+- **llm_call_trace 表**：append-only，无逻辑删除。记录每次 LLM 调用的完整现场——最终 `system_prompt` + `global_prompt`/`role_prompt`/`rag_context` 分段快照、用户消息、AI 回复（截断）、模型参数（model_name/provider/temperature/max_tokens）、耗时、意图（CHAT/CHITCHAT/FAQ/RAG）、状态（COMPLETE/ERROR/CANCEL/FAQ/BYPASS）。
+- **埋点位置**：`AssistantApp.chatWithEvents`/`chatStream` 顶部埋点，覆盖 FAQ 命中、熔断降级、流式断连/异常等路径；流式用 `doOnNext` 聚合分片 + `doFinally` 按终止信号落库（绝不在方法返回处计时）。
+- **异步写入**：`LlmCallTraceService.recordAsync` 走 `@Async("traceExecutor")`（有界线程池见 `AsyncExecutorConfig`，替代默认 `SimpleAsyncTaskExecutor` 线程爆炸隐患；`RagHitLogService` 已一并切换）。
+- **查询面板**：前端「系统设置 → 提示词追踪」（`PromptTracePanel.vue`，admin），支持筛选/搜索/详情（来源标注）/聚合统计/并排对比/跳转编辑（deep-link `?roleId=`、`?key=ai_system_prompt`）。
+- **安全声明**：`system_prompt`/`user_message`/`ai_response` 属敏感数据（提示词为商业秘密），当前**明文存储**、仅 admin 可查，落库前做敏感词脱敏（`ContentSafetyService.mask`）+ AI 回复截断；保留期限由 `system_config.llm_trace_retention_days` 控制（默认 30 天，`@Scheduled` 每日 2 点自动清理）。字段级加密留二期。
+- **预留**：`account_id`/`api_key_id` 列预留租户隔离与个人信息删除权，本期不写入；旁路 LLM 调用（意图识别/查询重写/建议生成）追踪留二期（表加 `trace_type`/`source` 字段）。
+
 ## 已知 TODO
 
 - `DocumentService.updateDocumentMetadata()`: Spring AI 无直接更新 vector_store metadata 的 API，向量元数据同步留后续

frontend/components.d.ts

@@ -9,6 +9,7 @@ declare module 'vue' {
   export interface GlobalComponents {
     FormDialog: typeof import('./src/components/FormDialog.vue')['default']
     MessageSources: typeof import('./src/components/MessageSources.vue')['default']
+    PromptSearchHighlight: typeof import('./src/components/PromptSearchHighlight.vue')['default']
     RouterLink: typeof import('vue-router')['RouterLink']
     RouterView: typeof import('vue-router')['RouterView']
     TButton: typeof import('tdesign-vue-next')['Button']
@@ -21,6 +22,7 @@ declare module 'vue' {
     TDatePicker: typeof import('tdesign-vue-next')['DatePicker']
     TDialog: typeof import('tdesign-vue-next')['Dialog']
     TDivider: typeof import('tdesign-vue-next')['Divider']
+    TDrawer: typeof import('tdesign-vue-next')['Drawer']
     TEmpty: typeof import('tdesign-vue-next')['Empty']
     TForm: typeof import('tdesign-vue-next')['Form']
     TFormItem: typeof import('tdesign-vue-next')['FormItem']
@@ -42,6 +44,8 @@ declare module 'vue' {
     TTabs: typeof import('tdesign-vue-next')['Tabs']
     TTag: typeof import('tdesign-vue-next')['Tag']
     TTextarea: typeof import('tdesign-vue-next')['Textarea']
+    TTimeline: typeof import('tdesign-vue-next')['Timeline']
+    TTimelineItem: typeof import('tdesign-vue-next')['TimelineItem']
     TUpload: typeof import('tdesign-vue-next')['Upload']
   }
 }

frontend/src/api/api-key.ts

@@ -4,6 +4,10 @@ import type { ApiResponse } from '@/types/api'
 export function listApiKeys(page = 1, size = 20): Promise<ApiResponse> {
   return request.get(`/api-key/list?page=${page}&size=${size}`).then(r => r.data)
 }
+/** 拉取全部 API Key（供筛选下拉，最多 1000 条） */
+export function listAllApiKeys(): Promise<ApiResponse> {
+  return request.get(`/api-key/list?page=1&size=1000`).then(r => r.data)
+}
 export function createApiKey(data: any): Promise<ApiResponse> { return request.post('/api-key', data).then(r => r.data) }
 export function revokeApiKey(id: string): Promise<ApiResponse> { return request.put(`/api-key/${id}/revoke`).then(r => r.data) }
 export function enableApiKey(id: string): Promise<ApiResponse> { return request.put(`/api-key/${id}/enable`).then(r => r.data) }

frontend/src/api/llm-trace.ts

@@ -0,0 +1,72 @@
+import request from './request'
+import type { ApiResponse } from '@/types/api'
+
+/** LLM 调用追踪列表查询参数 */
+export interface LlmTraceQuery {
+  page?: number
+  size?: number
+  roleId?: string
+  conversationId?: string
+  intent?: string
+  startTime?: string
+  endTime?: string
+  keyword?: string
+  apiKeyId?: string
+  errorType?: string
+}
+
+/** 分页查询调用记录 */
+export function listLlmTraces(query: LlmTraceQuery = {}): Promise<ApiResponse> {
+  const params = new URLSearchParams()
+  params.set('page', String(query.page ?? 1))
+  params.set('size', String(query.size ?? 20))
+  if (query.roleId) params.set('roleId', query.roleId)
+  if (query.conversationId) params.set('conversationId', query.conversationId)
+  if (query.intent) params.set('intent', query.intent)
+  if (query.startTime) params.set('startTime', query.startTime)
+  if (query.endTime) params.set('endTime', query.endTime)
+  if (query.keyword) params.set('keyword', query.keyword)
+  if (query.apiKeyId) params.set('apiKeyId', query.apiKeyId)
+  if (query.errorType) params.set('errorType', query.errorType)
+  return request.get(`/llm-trace/list?${params.toString()}`).then(r => r.data)
+}
+
+/** 单条详情（含完整 system prompt） */
+export function getLlmTrace(id: string): Promise<ApiResponse> {
+  return request.get(`/llm-trace/${id}`).then(r => r.data)
+}
+
+/** 清理 N 天前的记录（后端返回 {success, message, deleted}，无 data 字段） */
+export interface CleanLlmTracesResult { success: boolean; message?: string; deleted?: number }
+export function cleanLlmTraces(keepDays: number): Promise<CleanLlmTracesResult> {
+  return request.post('/llm-trace/clean', { keepDays }).then(r => r.data)
+}
+
+/** 聚合统计（按角色或模型分组） */
+export function getLlmTraceStats(groupBy: 'role' | 'model' = 'role'): Promise<ApiResponse> {
+  return request.get(`/llm-trace/stats?groupBy=${groupBy}`).then(r => r.data)
+}
+
+/** 可选错误类型列表 */
+export function getLlmTraceErrorTypes(): Promise<ApiResponse> {
+  return request.get('/llm-trace/error-types').then(r => r.data)
+}
+
+/** 单会话调用记录（会话时间线，复用 list 接口按 conversationId 过滤） */
+export function getLlmTraceConversation(conversationId: string): Promise<ApiResponse> {
+  return request.get(`/llm-trace/list?conversationId=${encodeURIComponent(conversationId)}&size=100`).then(r => r.data)
+}
+
+/** 时间维度趋势统计 */
+export interface LlmTraceTrendQuery {
+  groupBy?: 'HOUR' | 'DAY'
+  startTime?: string
+  endTime?: string
+}
+export function getLlmTraceTrend(query: LlmTraceTrendQuery = {}): Promise<ApiResponse> {
+  const params = new URLSearchParams()
+  params.set('groupBy', query.groupBy ?? 'HOUR')
+  if (query.startTime) params.set('startTime', query.startTime)
+  if (query.endTime) params.set('endTime', query.endTime)
+  return request.get(`/llm-trace/trend?${params.toString()}`).then(r => r.data)
+}

frontend/src/components/PromptSearchHighlight.vue

@@ -0,0 +1,66 @@
+<template>
+  <div class="prompt-highlight">
+    <div class="ph-toolbar">
+      <t-input v-model="keyword" placeholder="在提示词中搜索高亮" clearable size="small" style="width: 240px;">
+        <template #prefix-icon><SearchIcon /></template>
+      </t-input>
+      <span v-if="keyword" class="ph-count">{{ matchCount }} 处匹配</span>
+    </div>
+    <div class="ph-body" v-html="highlighted"></div>
+  </div>
+</template>
+
+<script setup lang="ts">
+import { ref, computed } from 'vue'
+import { SearchIcon } from 'tdesign-icons-vue-next'
+
+const props = defineProps<{ text: string }>()
+
+const keyword = ref('')
+
+/** 匹配数量（基于原始文本，忽略大小写） */
+const matchCount = computed(() => {
+  if (!keyword.value.trim() || !props.text) return 0
+  const regex = new RegExp(escapeRegex(keyword.value.trim()), 'gi')
+  return (props.text.match(regex) || []).length
+})
+
+/** 高亮后的 HTML（原始文本转义 + 关键词包裹 mark，防止 XSS） */
+const highlighted = computed(() => {
+  const escaped = escapeHtml(props.text || '')
+  if (!keyword.value.trim()) return escaped
+  const kw = escapeHtml(keyword.value.trim())
+  const regex = new RegExp(`(${escapeRegex(kw)})`, 'gi')
+  return escaped.replace(regex, '<mark class="ph-mark">$1</mark>')
+})
+
+function escapeHtml(s: string): string {
+  return s
+    .replace(/&/g, '&amp;')
+    .replace(/</g, '&lt;')
+    .replace(/>/g, '&gt;')
+    .replace(/"/g, '&quot;')
+}
+
+function escapeRegex(s: string): string {
+  return s.replace(/[.*+?^${}()|[\]\\]/g, '\\$&')
+}
+</script>
+
+<style scoped>
+.prompt-highlight { display: flex; flex-direction: column; gap: 8px; }
+.ph-toolbar { display: flex; align-items: center; gap: 10px; }
+.ph-count { font-size: 12px; color: var(--td-text-color-secondary); }
+.ph-body {
+  white-space: pre-wrap;
+  word-break: break-word;
+  font-size: 13px;
+  line-height: 1.6;
+  background: var(--td-bg-color-secondarycontainer, #f7f7f7);
+  padding: 12px;
+  border-radius: 6px;
+  max-height: 480px;
+  overflow: auto;
+}
+:deep(.ph-mark) { background: var(--td-warning-color-1, #fff3e0); color: var(--td-text-color-primary); padding: 0 1px; }
+</style>

frontend/src/router/index.ts

@@ -36,6 +36,7 @@ const routes: RouteRecordRaw[] = [
   { path: '/settings/pipeline-flow', name: 'PipelineFlow', component: () => import('@/views/PipelineFlow.vue') },
   { path: '/settings/system-config', name: 'SystemConfig', component: () => import('@/views/SystemConfigManager.vue') },
   { path: '/settings/log-viewer', name: 'LogViewer', component: () => import('@/views/LogViewer.vue') },
+  { path: '/settings/prompt-trace', name: 'PromptTrace', component: () => import('@/views/PromptTracePanel.vue') },
   // ==================== 兜底 ====================
   { path: '/:pathMatch(.*)*', redirect: '/chat' },
 ]

frontend/src/stores/navigation.ts

@@ -33,6 +33,7 @@ export const MENU_ITEMS = [
       { id: 'pipeline-flow', label: 'AI 执行链', icon: '🔀', path: '/settings/pipeline-flow' },
       { id: 'system-config', label: '系统配置', icon: '🔧', path: '/settings/system-config', roles: ['admin'] },
       { id: 'log-viewer', label: '系统日志', icon: '📜', path: '/settings/log-viewer', roles: ['admin'] },
+      { id: 'prompt-trace', label: '提示词追踪', icon: '🔬', path: '/settings/prompt-trace', roles: ['admin'] },
     ],
   },
 ]

frontend/src/views/PromptTracePanel.vue

@@ -0,0 +1,705 @@
+<template>
+  <t-card title="提示词追踪" :bordered="false">
+    <p class="desc-text">记录每次 LLM 调用的完整现场（system prompt / 模型参数 / 耗时 / 意图），用于提示词优化调试。记录自动保留 N 天后清理。</p>
+
+    <!-- 聚合统计 -->
+    <div class="stats-bar">
+      <t-radio-group v-model="groupBy" variant="default-filled" size="small" @change="loadStats">
+        <t-radio-button value="role">按角色</t-radio-button>
+        <t-radio-button value="model">按模型</t-radio-button>
+      </t-radio-group>
+      <div class="stat-card"><span class="stat-num">{{ statSummary.count }}</span><span class="stat-label">总调用</span></div>
+      <div class="stat-card"><span class="stat-num">{{ statSummary.avgLatency }}</span><span class="stat-label">平均耗时(ms)</span></div>
+      <div class="stat-card"><span class="stat-num">{{ statSummary.faqRate }}</span><span class="stat-label">FAQ 命中率</span></div>
+    </div>
+    <t-table :data="stats" :columns="statsColumns" row-key="key" size="small" :loading="statsLoading" :pagination="false" style="margin-bottom:16px;" />
+
+    <!-- 筛选栏 -->
+    <div class="toolbar">
+      <t-select v-model="filterRoleId" :options="roleOptions" placeholder="全部角色" clearable size="small" style="width:150px;" @change="onFilterChange" />
+      <t-input v-model="filterConversationId" placeholder="会话 ID" clearable size="small" style="width:180px;" @enter="onSearch" />
+      <t-select v-model="filterIntent" :options="intentOptions" placeholder="全部意图" clearable size="small" style="width:120px;" @change="onFilterChange" />
+      <t-select v-model="filterApiKeyId" :options="apiKeyOptions" placeholder="全部 API Key" clearable size="small" style="width:160px;" @change="onFilterChange" />
+      <t-select v-model="filterErrorType" :options="errorTypeOptions" placeholder="全部错误类型" clearable size="small" style="width:140px;" @change="onFilterChange" />
+      <t-date-picker v-model="filterStart" placeholder="开始日期" clearable size="small" style="width:140px;" @change="onFilterChange" />
+      <t-date-picker v-model="filterEnd" placeholder="结束日期" clearable size="small" style="width:140px;" @change="onFilterChange" />
+      <t-input v-model="filterKeyword" placeholder="搜索用户消息 / AI 回复" clearable size="small" style="width:200px;" @enter="onSearch" />
+      <div style="flex:1;" />
+      <t-button size="small" variant="outline" @click="refreshAll">刷新</t-button>
+      <t-button size="small" :disabled="selectedRowKeys.length !== 2" @click="openCompare">对比（{{ selectedRowKeys.length }}/2）</t-button>
+      <t-button size="small" variant="outline" @click="openTrend">趋势</t-button>
+      <t-button size="small" variant="outline" @click="openClean">清理</t-button>
+    </div>
+
+    <!-- 调用列表 -->
+    <t-table :data="traces" :columns="columns" row-key="id" :loading="loading"
+      :selected-row-keys="selectedRowKeys"
+      :pagination="{ current: page, total: total, pageSize: pageSize, showJumper: true }"
+      @page-change="onPageChange" @select-change="onSelectChange">
+      <template #roleName="{ row }">
+        <span v-if="row.roleName">{{ row.roleName }}</span>
+        <span v-else class="muted">(无角色)</span>
+      </template>
+      <template #intent="{ row }">
+        <t-tag size="small" variant="light" :theme="intentTheme(row.intent)">{{ intentLabel(row.intent) }}</t-tag>
+      </template>
+      <template #faqHit="{ row }">
+        <t-tag size="small" variant="light" :theme="row.faqHit ? 'success' : 'default'">{{ row.faqHit ? '命中' : '未命中' }}</t-tag>
+      </template>
+      <template #enableRag="{ row }">
+        <t-tag size="small" variant="light" :theme="row.enableRag ? 'primary' : 'default'">{{ row.enableRag ? '是' : '否' }}</t-tag>
+      </template>
+      <template #status="{ row }">
+        <t-tag size="small" variant="light" :theme="statusTheme(row.status)">{{ statusLabel(row.status) }}</t-tag>
+      </template>
+      <template #latencyMs="{ row }">
+        <span>{{ row.latencyMs }}ms</span>
+      </template>
+      <template #op="{ row }">
+        <t-button size="small" variant="text" @click="openDetail(row)">详情</t-button>
+        <t-button size="small" variant="text" @click="openTimeline(row)">时间线</t-button>
+      </template>
+    </t-table>
+
+    <!-- 详情抽屉 -->
+    <t-drawer v-model:visible="detailVisible" size="large" header="调用详情" :footer="false">
+      <t-loading :loading="detailLoading" show-overlay>
+        <div v-if="detail">
+          <div class="detail-meta">
+            <t-tag size="small" variant="light" :theme="intentTheme(detail.intent)">{{ intentLabel(detail.intent) }}</t-tag>
+            <t-tag size="small" variant="light" :theme="statusTheme(detail.status)">{{ statusLabel(detail.status) }}</t-tag>
+            <t-tag v-if="detail.searchMode" size="small" variant="outline">{{ detail.searchMode }}</t-tag>
+            <t-tag v-if="detail.errorType" size="small" variant="light" theme="danger">{{ detail.errorType }}</t-tag>
+            <span class="muted">{{ detail.roleName || '(无角色)' }} · {{ detail.modelName || '未知模型' }} · {{ detail.provider }} · {{ detail.latencyMs }}ms</span>
+          </div>
+          <div v-if="detail.errorMessage" class="error-box">{{ detail.errorMessage }}</div>
+          <t-tabs v-model="detailTab">
+            <t-tab-panel value="system" label="系统提示词">
+              <div class="prompt-segments">
+                <div class="seg" v-if="detail.globalPrompt"><span class="seg-tag">全局</span><pre class="seg-text">{{ detail.globalPrompt }}</pre></div>
+                <div class="seg" v-if="detail.rolePrompt"><span class="seg-tag">角色</span><pre class="seg-text">{{ detail.rolePrompt }}</pre></div>
+                <div class="seg" v-if="detail.ragContext"><span class="seg-tag">RAG 资料</span><pre class="seg-text">{{ detail.ragContext }}</pre></div>
+              </div>
+              <div class="block-title">最终完整 system prompt</div>
+              <PromptSearchHighlight :text="detail.systemPrompt || '(空)'" />
+              <div class="detail-actions">
+                <t-button size="small" @click="copyPrompt">复制完整 Prompt</t-button>
+                <t-button size="small" variant="outline" :disabled="!detail.roleId" @click="jumpEditRole">编辑角色提示词</t-button>
+                <t-button size="small" variant="outline" @click="jumpEditGlobal">编辑全局提示词</t-button>
+              </div>
+            </t-tab-panel>
+            <t-tab-panel value="user" label="用户消息"><pre class="full-prompt">{{ detail.userMessage }}</pre></t-tab-panel>
+            <t-tab-panel value="ai" label="AI 回复">
+              <div v-if="detail.totalTokens || detail.promptTokens || detail.completionTokens" class="muted" style="margin-bottom:8px;">
+                Token：prompt {{ detail.promptTokens ?? '-' }} / completion {{ detail.completionTokens ?? '-' }} / 合计 {{ detail.totalTokens ?? '-' }}
+              </div>
+              <pre class="full-prompt">{{ detail.aiResponse }}<span v-if="detail.aiResponseTruncated" class="muted">（已截断）</span></pre>
+            </t-tab-panel>
+            <t-tab-panel value="rag" label="RAG 资料">
+              <div class="muted" style="margin-bottom:8px;">检索模式：{{ detail.searchMode || '-' }} · 命中：{{ detail.hitCount ?? '-' }} 条</div>
+              <t-table v-if="detailRagHits.length" :data="detailRagHits" :columns="ragHitsColumns" size="small" :pagination="false" style="margin-bottom:12px;" />
+              <div class="block-title">注入的 RAG 上下文</div>
+              <pre class="full-prompt">{{ detail.ragContext || '(无 RAG 资料)' }}</pre>
+            </t-tab-panel>
+            <t-tab-panel value="faq" label="FAQ">
+              <t-empty v-if="!detail.faqQuestion && !detail.faqId" description="本次调用未命中 FAQ" />
+              <div v-else class="faq-box">
+                <div class="faq-row"><span class="faq-label">FAQ ID</span><span>{{ detail.faqId ?? '-' }}</span></div>
+                <div class="faq-row"><span class="faq-label">问题</span><span>{{ detail.faqQuestion ?? '-' }}</span></div>
+                <div class="faq-row"><span class="faq-label">匹配方式</span><span>{{ detail.faqMatchType || '-' }}</span></div>
+                <div class="faq-row"><span class="faq-label">相似度</span><span>{{ formatScore(detail.faqScore) }}</span></div>
+              </div>
+            </t-tab-panel>
+            <t-tab-panel value="tools" label="工具调用">
+              <t-table v-if="detailToolCalls.length" :data="detailToolCalls" :columns="toolCallsColumns" size="small" :pagination="false">
+                <template #latencyMs="{ row }"><span>{{ row.latencyMs ?? '-' }}ms</span></template>
+                <template #error="{ row }">
+                  <t-tag v-if="row.error" size="small" variant="light" theme="danger">失败</t-tag>
+                  <span v-else class="muted">-</span>
+                </template>
+              </t-table>
+              <t-empty v-else description="本次调用未触发 MCP 工具" />
+            </t-tab-panel>
+            <t-tab-panel value="context" label="对话上下文">
+              <div v-if="detail.historyTurns" class="muted" style="margin-bottom:8px;">历史轮数：{{ detail.historyTurns }}</div>
+              <t-table v-if="detailHistory.length" :data="detailHistory" :columns="historyColumns" size="small" :pagination="false" />
+              <t-empty v-else description="无多轮历史记录" />
+            </t-tab-panel>
+          </t-tabs>
+        </div>
+        <t-empty v-else description="记录不存在或已被清理" />
+      </t-loading>
+    </t-drawer>
+
+    <!-- 对比抽屉 -->
+    <t-drawer v-model:visible="compareVisible" size="large" header="并排对比" :footer="false">
+      <div v-if="compareA && compareB" class="compare-grid">
+        <div v-for="(c, idx) in [compareA, compareB]" :key="idx" class="compare-col">
+          <div class="compare-head">{{ c.roleName || '(无角色)' }} · {{ c.intent }} · {{ formatDate(c.createTime) }}</div>
+          <div class="compare-meta">
+            <t-tag size="small" variant="light" :theme="c.enableRag ? 'primary' : 'default'">RAG {{ c.enableRag ? '开' : '关' }}</t-tag>
+            <t-tag v-if="c.searchMode" size="small" variant="outline">{{ c.searchMode }}</t-tag>
+            <t-tag v-if="c.errorType" size="small" variant="light" theme="danger">{{ c.errorType }}</t-tag>
+            <span v-if="c.faqMatchType" class="muted">FAQ：{{ c.faqMatchType }}</span>
+            <span v-if="c.totalTokens" class="muted">Token：{{ c.totalTokens }}</span>
+          </div>
+          <div class="block-title">系统提示词</div>
+          <pre class="full-prompt small">{{ c.systemPrompt || '(空)' }}</pre>
+          <div class="block-title">用户消息</div>
+          <pre class="full-prompt small">{{ c.userMessage }}</pre>
+          <div class="block-title">AI 回复</div>
+          <pre class="full-prompt small">{{ c.aiResponse }}</pre>
+        </div>
+      </div>
+      <t-empty v-else description="请先在列表中勾选 2 条记录" />
+    </t-drawer>
+
+    <!-- 会话时间线抽屉 -->
+    <t-drawer v-model:visible="timelineVisible" size="medium" header="会话时间线" :footer="false">
+      <t-loading :loading="timelineLoading" show-overlay>
+        <div class="muted" style="margin-bottom:12px;">会话 ID：{{ timelineConversationId }}</div>
+        <t-empty v-if="!timelineLoading && !timelineRows.length" description="该会话无调用记录" />
+        <t-timeline v-else>
+          <t-timeline-item v-for="row in timelineRows" :key="row.id">
+            <div class="tl-head">
+              <t-tag size="small" variant="light" :theme="intentTheme(row.intent)">{{ intentLabel(row.intent) }}</t-tag>
+              <t-tag size="small" variant="light" :theme="statusTheme(row.status)">{{ statusLabel(row.status) }}</t-tag>
+              <span class="muted">{{ formatDate(row.createTime) }} · {{ row.latencyMs }}ms</span>
+            </div>
+            <div class="tl-user">{{ row.userMessage }}</div>
+            <div class="tl-ai">{{ row.aiResponse }}</div>
+          </t-timeline-item>
+        </t-timeline>
+      </t-loading>
+    </t-drawer>
+
+    <!-- 趋势图抽屉 -->
+    <t-drawer v-model:visible="trendVisible" size="large" header="调用趋势" :footer="false" @opened="loadTrend" @closed="destroyTrendChart">
+      <div class="trend-toolbar">
+        <t-radio-group v-model="trendRange" variant="default-filled" size="small" @change="loadTrend">
+          <t-radio-button value="24h">近 24 小时</t-radio-button>
+          <t-radio-button value="7d">近 7 天</t-radio-button>
+        </t-radio-group>
+        <t-radio-group v-model="trendGroupBy" variant="default-filled" size="small" @change="loadTrend">
+          <t-radio-button value="HOUR">按小时</t-radio-button>
+          <t-radio-button value="DAY">按天</t-radio-button>
+        </t-radio-group>
+        <div style="flex:1;" />
+        <t-button size="small" variant="outline" :loading="trendLoading" @click="loadTrend">刷新</t-button>
+      </div>
+      <div class="trend-chart-wrap">
+        <canvas ref="trendChartRef"></canvas>
+      </div>
+    </t-drawer>
+
+    <!-- 清理弹窗 -->
+    <t-dialog v-model:visible="cleanVisible" header="清理调用记录" width="420px" :footer="false">
+      <t-form label-align="top">
+        <t-form-item label="保留最近 N 天（删除 N 天前的记录）">
+          <t-input-number v-model="cleanKeepDays" :min="1" :max="365" style="width:100%;" />
+        </t-form-item>
+      </t-form>
+      <p class="muted" style="font-size:12px;">最小保留 1 天，清理动作不可恢复，请谨慎操作。</p>
+      <div class="dialog-footer">
+        <t-button variant="outline" @click="cleanVisible = false">取消</t-button>
+        <t-button theme="danger" :loading="cleaning" @click="doClean">确认清理</t-button>
+      </div>
+    </t-dialog>
+  </t-card>
+</template>
+
+<script setup lang="ts">
+import { ref, computed, onMounted, onBeforeUnmount, nextTick } from 'vue'
+import { useRouter } from 'vue-router'
+import { listLlmTraces, getLlmTrace, cleanLlmTraces, getLlmTraceStats, getLlmTraceErrorTypes, getLlmTraceConversation, getLlmTraceTrend } from '@/api/llm-trace'
+import { getAllRoles } from '@/api/role'
+import { listAllApiKeys } from '@/api/api-key'
+import PromptSearchHighlight from '@/components/PromptSearchHighlight.vue'
+import { toast } from '@/utils/toast'
+import { formatDate } from '@/utils/format'
+import { useDebounce } from '@/composables/useDebounce'
+import { useConfirm } from '@/composables/useConfirm'
+import { palette } from '@/utils/palette'
+
+const router = useRouter()
+const { confirm } = useConfirm()
+const { debounce } = useDebounce()
+
+// ===== 意图 / 状态映射 =====
+const INTENT_LABEL: Record<string, string> = { CHAT: '普通对话', CHITCHAT: '闲聊', FAQ: 'FAQ', RAG: 'RAG' }
+const STATUS_LABEL: Record<string, string> = { COMPLETE: '完成', ERROR: '失败', CANCEL: '断连', FAQ: 'FAQ', BYPASS: '熔断降级' }
+function intentLabel(v: string) { return INTENT_LABEL[v] || v || '-' }
+function intentTheme(v: string) { return ({ CHITCHAT: 'warning', FAQ: 'success', RAG: 'primary', CHAT: 'default' } as Record<string, string>)[v] || 'default' }
+function statusLabel(v: string) { return STATUS_LABEL[v] || v || '-' }
+function statusTheme(v: string) { return ({ COMPLETE: 'success', ERROR: 'danger', CANCEL: 'warning', FAQ: 'success', BYPASS: 'warning' } as Record<string, string>)[v] || 'default' }
+
+const intentOptions = [
+  { label: '普通对话', value: 'CHAT' },
+  { label: '闲聊', value: 'CHITCHAT' },
+  { label: 'FAQ', value: 'FAQ' },
+  { label: 'RAG', value: 'RAG' },
+]
+
+// ===== 筛选状态 =====
+const filterRoleId = ref('')
+const filterConversationId = ref('')
+const filterIntent = ref('')
+const filterStart = ref('')
+const filterEnd = ref('')
+const filterKeyword = ref('')
+const filterApiKeyId = ref('')
+const filterErrorType = ref('')
+const roleOptions = ref<{ label: string; value: string }[]>([])
+const apiKeyOptions = ref<{ label: string; value: string }[]>([])
+const errorTypeOptions = ref<{ label: string; value: string }[]>([])
+
+// ===== 列表状态 =====
+const traces = ref<any[]>([])
+const loading = ref(false)
+const page = ref(1)
+const pageSize = ref(20)
+const total = ref(0)
+const selectedRowKeys = ref<string[]>([])
+
+const columns = [
+  { colKey: 'row-select', type: 'multiple', width: 40 },
+  { colKey: 'createTime', title: '时间', width: 160, cell: (_: any, { row }: any) => formatDate(row.createTime) },
+  { colKey: 'roleName', title: '角色', width: 120 },
+  { colKey: 'intent', title: '意图', width: 100 },
+  { colKey: 'enableRag', title: 'RAG', width: 70 },
+  { colKey: 'modelName', title: '模型', width: 140, ellipsis: true },
+  { colKey: 'latencyMs', title: '耗时', width: 90 },
+  { colKey: 'faqHit', title: 'FAQ', width: 80 },
+  { colKey: 'status', title: '状态', width: 100 },
+  { colKey: 'userMessage', title: '用户消息', ellipsis: true },
+  { colKey: 'op', title: '操作', width: 150, fixed: 'right' },
+]
+
+// ===== 聚合统计状态 =====
+const groupBy = ref<'role' | 'model'>('role')
+const stats = ref<any[]>([])
+const statsLoading = ref(false)
+const statsColumns = [
+  { colKey: 'key', title: '分组', ellipsis: true },
+  { colKey: 'count', title: '调用数', width: 100 },
+  { colKey: 'avgLatencyMs', title: '平均耗时(ms)', width: 120 },
+  { colKey: 'faqHitRate', title: 'FAQ 命中率', width: 110, cell: (_: any, { row }: any) => `${((row.faqHitRate || 0) * 100).toFixed(1)}%` },
+]
+
+// ===== 详情状态 =====
+const detailVisible = ref(false)
+const detailLoading = ref(false)
+const detail = ref<any>(null)
+const detailTab = ref('system')
+
+const detailRagHits = computed(() => parseJsonArray(detail.value?.ragHitsJson))
+const detailToolCalls = computed(() => parseJsonArray(detail.value?.toolCallsJson))
+const detailHistory = computed(() => parseJsonArray(detail.value?.historyMessagesJson))
+
+const ragHitsColumns = [
+  { colKey: 'title', title: '标题', ellipsis: true },
+  { colKey: 'documentId', title: '文档 ID', width: 150, ellipsis: true },
+  { colKey: 'chunkIndex', title: '分块', width: 70 },
+  { colKey: 'score', title: '相似度', width: 90, cell: (_: any, { row }: any) => formatScore(row.score) },
+  { colKey: 'searchMode', title: '检索模式', width: 100 },
+]
+
+const toolCallsColumns = [
+  { colKey: 'tool', title: '工具', width: 150, ellipsis: true },
+  { colKey: 'input', title: '输入', ellipsis: true },
+  { colKey: 'result', title: '结果', ellipsis: true },
+  { colKey: 'latencyMs', title: '耗时', width: 90 },
+  { colKey: 'error', title: '错误', width: 70 },
+]
+
+const historyColumns = [
+  { colKey: 'role', title: '角色', width: 100 },
+  { colKey: 'content', title: '内容', ellipsis: true },
+]
+
+function formatScore(v: any): string {
+  if (v === null || v === undefined || v === '') return '-'
+  const n = Number(v)
+  if (Number.isFinite(n)) return n.toFixed(4)
+  return String(v)
+}
+
+// ===== 对比状态 =====
+const compareVisible = ref(false)
+const compareA = ref<any>(null)
+const compareB = ref<any>(null)
+
+// ===== 清理状态 =====
+const cleanVisible = ref(false)
+const cleanKeepDays = ref(30)
+const cleaning = ref(false)
+
+// ===== 会话时间线状态 =====
+const timelineVisible = ref(false)
+const timelineLoading = ref(false)
+const timelineConversationId = ref('')
+const timelineRows = ref<any[]>([])
+
+// ===== 趋势图状态 =====
+const trendVisible = ref(false)
+const trendGroupBy = ref<'HOUR' | 'DAY'>('HOUR')
+const trendRange = ref('24h')
+const trendLoading = ref(false)
+const trendRows = ref<any[]>([])
+const trendChartRef = ref<HTMLCanvasElement | null>(null)
+let trendChart: any = null
+
+onMounted(async () => {
+  await loadRoles()
+  loadApiKeys()
+  loadErrorTypes()
+  loadList()
+  loadStats()
+})
+
+async function loadRoles() {
+  try {
+    const r = await getAllRoles()
+    if (r.success) {
+      roleOptions.value = (r.data || []).map((ro: any) => ({ label: ro.name, value: String(ro.id) }))
+    }
+  } catch (e: any) { console.warn('角色下拉加载失败：' + e.message) }
+}
+
+async function loadApiKeys() {
+  try {
+    const r = await listAllApiKeys()
+    if (r.success) {
+      apiKeyOptions.value = (r.data || []).map((k: any) => ({ label: k.name || ('API Key ' + k.id), value: String(k.id) }))
+    }
+  } catch (e: any) { console.warn('API Key 下拉加载失败：' + e.message) }
+}
+
+async function loadErrorTypes() {
+  try {
+    const r = await getLlmTraceErrorTypes()
+    if (r.success) {
+      errorTypeOptions.value = (r.data || []).map((t: any) => ({ label: t.label, value: t.value }))
+    }
+  } catch (e: any) { console.warn('错误类型下拉加载失败：' + e.message) }
+}
+
+// ===== 聚合统计 =====
+const statSummary = computed(() => {
+  let count = 0, latencySum = 0, faqHits = 0
+  for (const s of stats.value) {
+    count += s.count || 0
+    latencySum += (s.count || 0) * (s.avgLatencyMs || 0)
+    faqHits += Math.round((s.count || 0) * (s.faqHitRate || 0))
+  }
+  const avgLatency = count > 0 ? Math.round(latencySum / count) : '-'
+  const faqRate = count > 0 ? ((faqHits / count) * 100).toFixed(1) + '%' : '-'
+  return { count, avgLatency, faqRate }
+})
+
+async function loadStats() {
+  statsLoading.value = true
+  try {
+    const r = await getLlmTraceStats(groupBy.value)
+    if (r.success) {
+      const rows = (r.data || []).map((s: any) => {
+        const key = groupBy.value === 'model'
+          ? `${s.modelName || '-'}|${s.provider || '-'}`
+          : `${s.roleId ?? s.roleName ?? '(无角色)'}`
+        const faqRate = s.count > 0 ? ((s.faqHitCount || 0) / s.count) : 0
+        return { ...s, key, faqHitRate: faqRate }
+      })
+      stats.value = rows
+    } else {
+      toast(r.message || '加载统计失败', 'error')
+    }
+  } catch (e: any) {
+    toast('加载统计失败：' + e.message, 'error')
+  } finally {
+    statsLoading.value = false
+  }
+}
+
+// ===== 列表 =====
+async function loadList() {
+  loading.value = true
+  try {
+    const r = await listLlmTraces({
+      page: page.value, size: pageSize.value,
+      roleId: filterRoleId.value || undefined,
+      conversationId: filterConversationId.value || undefined,
+      intent: filterIntent.value || undefined,
+      startTime: filterStart.value ? filterStart.value + ' 00:00:00' : undefined,
+      endTime: filterEnd.value ? filterEnd.value + ' 23:59:59' : undefined,
+      keyword: filterKeyword.value || undefined,
+      apiKeyId: filterApiKeyId.value || undefined,
+      errorType: filterErrorType.value || undefined,
+    })
+    if (r.success) {
+      traces.value = r.data?.records || r.data || []
+      total.value = r.data?.total || r.total || 0
+    } else {
+      toast(r.message || '加载失败', 'error')
+    }
+  } catch (e: any) {
+    toast('加载失败：' + e.message, 'error')
+  } finally {
+    loading.value = false
+  }
+}
+
+function onPageChange(info: { current: number; pageSize: number }) {
+  page.value = info.current
+  pageSize.value = info.pageSize
+  selectedRowKeys.value = []
+  loadList()
+}
+
+function onFilterChange() {
+  page.value = 1
+  selectedRowKeys.value = []
+  loadList()
+}
+
+function refreshAll() {
+  loadList()
+  loadStats()
+}
+
+const onSearch = debounce(() => { page.value = 1; selectedRowKeys.value = []; loadList() })
+
+function onSelectChange(keys: (string | number)[], options?: { currentRowKey?: string | number }) {
+  const normalized = keys.map(String)
+  if (normalized.length > 2) {
+    toast('最多选择 2 条进行对比', 'warning')
+    // 保留最近勾选的一条 + 已选的第一条，避免勾选第 3 条时静默丢弃新勾选项
+    const current = String(options?.currentRowKey ?? normalized[normalized.length - 1])
+    const others = normalized.filter(k => k !== current)
+    selectedRowKeys.value = [current, ...others].slice(0, 2)
+  } else {
+    selectedRowKeys.value = normalized
+  }
+}
+
+// ===== 详情 =====
+async function openDetail(row: any) {
+  detailVisible.value = true
+  detailLoading.value = true
+  detail.value = null
+  detailTab.value = 'system'
+  try {
+    const r = await getLlmTrace(row.id)
+    if (r.success) {
+      detail.value = r.data
+    } else {
+      detail.value = null
+    }
+  } catch (e: any) {
+    toast('加载详情失败：' + e.message, 'error')
+  } finally {
+    detailLoading.value = false
+  }
+}
+
+async function copyPrompt() {
+  if (!detail.value?.systemPrompt) return
+  try {
+    await navigator.clipboard.writeText(detail.value.systemPrompt)
+    toast('已复制完整 Prompt', 'success')
+  } catch {
+    toast('复制失败', 'error')
+  }
+}
+
+function jumpEditRole() {
+  if (detail.value?.roleId) {
+    router.push({ path: '/settings/role', query: { roleId: String(detail.value.roleId) } })
+  }
+}
+function jumpEditGlobal() {
+  router.push({ path: '/settings/system-config', query: { key: 'ai_system_prompt' } })
+}
+
+// ===== 对比 =====
+async function openCompare() {
+  if (selectedRowKeys.value.length !== 2) return
+  const ids = selectedRowKeys.value
+  try {
+    const [a, b] = await Promise.all([getLlmTrace(ids[0]), getLlmTrace(ids[1])])
+    if (!a.success || !b.success) {
+      toast(a.message || b.message || '加载对比数据失败', 'error')
+      return
+    }
+    compareA.value = a.data
+    compareB.value = b.data
+    compareVisible.value = true
+  } catch (e: any) {
+    toast('加载对比数据失败：' + e.message, 'error')
+  }
+}
+
+// ===== 清理 =====
+function openClean() {
+  cleanKeepDays.value = 30
+  cleanVisible.value = true
+}
+
+async function doClean() {
+  if (!(await confirm('确认清理 ' + cleanKeepDays.value + ' 天前的调用记录？'))) return
+  cleaning.value = true
+  try {
+    const r = await cleanLlmTraces(cleanKeepDays.value)
+    if (r.success) {
+      toast('清理完成，删除 ' + (r.deleted ?? 0) + ' 条', 'success')
+      cleanVisible.value = false
+      loadList()
+      loadStats()
+    } else {
+      toast(r.message || '清理失败', 'error')
+    }
+  } catch (e: any) {
+    toast('清理失败：' + e.message, 'error')
+  } finally {
+    cleaning.value = false
+  }
+}
+
+// ===== 会话时间线 =====
+async function openTimeline(row: any) {
+  if (!row.conversationId) { toast('该记录无会话 ID', 'warning'); return }
+  timelineVisible.value = true
+  timelineConversationId.value = row.conversationId
+  timelineLoading.value = true
+  timelineRows.value = []
+  try {
+    const r = await getLlmTraceConversation(row.conversationId)
+    if (r.success) {
+      timelineRows.value = r.data || []
+    } else {
+      toast(r.message || '加载会话时间线失败', 'error')
+    }
+  } catch (e: any) {
+    toast('加载会话时间线失败：' + e.message, 'error')
+  } finally {
+    timelineLoading.value = false
+  }
+}
+
+// ===== 趋势图 =====
+function openTrend() {
+  trendVisible.value = true
+}
+
+function trendTimeRange(): { startTime?: string; endTime?: string } {
+  const end = new Date()
+  if (trendRange.value === '24h') {
+    const start = new Date(end.getTime() - 24 * 3600 * 1000)
+    return { startTime: fmtDateTime(start), endTime: fmtDateTime(end) }
+  }
+  if (trendRange.value === '7d') {
+    const start = new Date(end.getTime() - 7 * 24 * 3600 * 1000)
+    return { startTime: fmtDateTime(start), endTime: fmtDateTime(end) }
+  }
+  return {}
+}
+
+function fmtDateTime(d: Date): string {
+  const p = (n: number) => String(n).padStart(2, '0')
+  return `${d.getFullYear()}-${p(d.getMonth() + 1)}-${p(d.getDate())} ${p(d.getHours())}:${p(d.getMinutes())}:${p(d.getSeconds())}`
+}
+
+async function loadTrend() {
+  trendLoading.value = true
+  try {
+    const range = trendTimeRange()
+    const r = await getLlmTraceTrend({ groupBy: trendGroupBy.value, ...range })
+    if (r.success) {
+      trendRows.value = r.data || []
+      await nextTick()
+      renderTrendChart()
+    } else {
+      toast(r.message || '加载趋势失败', 'error')
+    }
+  } catch (e: any) {
+    toast('加载趋势失败：' + e.message, 'error')
+  } finally {
+    trendLoading.value = false
+  }
+}
+
+async function renderTrendChart() {
+  const { Chart, LineController, LineElement, PointElement, LinearScale, CategoryScale, Tooltip, Legend, Filler } = await import('chart.js')
+  Chart.register(LineController, LineElement, PointElement, LinearScale, CategoryScale, Tooltip, Legend, Filler)
+  const labels = trendRows.value.map((s: any) => s.timeBucket || '')
+  if (trendChart) { trendChart.destroy(); trendChart = null }
+  if (!trendChartRef.value) return
+  trendChart = new Chart(trendChartRef.value, {
+    type: 'line',
+    data: { labels, datasets: [
+      { label: '调用量', data: trendRows.value.map((s: any) => s.callCount ?? 0), borderColor: palette.blue, backgroundColor: `rgba(${palette.blueRgb},0.1)`, fill: true, tension: 0.3, pointRadius: 3, yAxisID: 'y' },
+      { label: '平均耗时(ms)', data: trendRows.value.map((s: any) => s.avgLatencyMs ?? 0), borderColor: palette.purple, backgroundColor: `rgba(${palette.purpleRgb},0.1)`, fill: false, tension: 0.3, pointRadius: 3, yAxisID: 'y1' },
+      { label: '错误率(%)', data: trendRows.value.map((s: any) => s.errorRate ?? 0), borderColor: palette.red, backgroundColor: `rgba(${palette.redRgb},0.1)`, fill: false, tension: 0.3, pointRadius: 3, yAxisID: 'y1' },
+      { label: 'FAQ 命中率(%)', data: trendRows.value.map((s: any) => s.faqHitRate ?? 0), borderColor: palette.green, backgroundColor: `rgba(${palette.greenRgb},0.1)`, fill: false, tension: 0.3, pointRadius: 3, yAxisID: 'y1' },
+      { label: '总 Token', data: trendRows.value.map((s: any) => s.totalTokens ?? 0), borderColor: palette.orange, backgroundColor: `rgba(${palette.orangeRgb},0.1)`, fill: false, tension: 0.3, pointRadius: 3, yAxisID: 'y1' },
+    ] },
+    options: {
+      responsive: true, maintainAspectRatio: false,
+      interaction: { mode: 'index', intersect: false },
+      plugins: { legend: { position: 'bottom', labels: { boxWidth: 12, font: { size: 12 } } } },
+      scales: {
+        y: { type: 'linear', display: true, position: 'left', beginAtZero: true, title: { display: true, text: '调用量' } },
+        y1: { type: 'linear', display: true, position: 'right', beginAtZero: true, grid: { drawOnChartArea: false }, title: { display: true, text: '耗时/率/Token' } },
+        x: { grid: { display: false } },
+      },
+    },
+  })
+}
+
+function destroyTrendChart() { if (trendChart) { trendChart.destroy(); trendChart = null } }
+
+onBeforeUnmount(() => destroyTrendChart())
+
+// ===== JSON 解析辅助 =====
+function parseJsonArray(text: string | null | undefined): any[] {
+  if (!text) return []
+  try {
+    const arr = JSON.parse(text)
+    return Array.isArray(arr) ? arr : []
+  } catch {
+    return []
+  }
+}
+</script>
+
+<style scoped>
+.desc-text { color: var(--color-text-secondary); font-size: 13px; margin: 0 0 16px; }
+.toolbar { display: flex; align-items: center; gap: 8px; margin: 12px 0; flex-wrap: wrap; }
+.stats-bar { display: flex; align-items: center; gap: 16px; margin-bottom: 8px; flex-wrap: wrap; }
+.stat-card { display: flex; flex-direction: column; align-items: center; min-width: 80px; padding: 4px 12px; border-radius: 6px; background: var(--color-bg-secondary, #f3f3f3); }
+.stat-num { font-size: 18px; font-weight: 600; color: var(--color-text-primary); }
+.stat-label { font-size: 12px; color: var(--color-text-secondary); }
+.muted { color: var(--color-text-secondary); }
+.detail-meta { display: flex; align-items: center; gap: 8px; margin-bottom: 12px; flex-wrap: wrap; }
+.block-title { font-size: 13px; font-weight: 600; margin: 12px 0 6px; }
+.prompt-segments { display: flex; flex-direction: column; gap: 8px; }
+.seg { border: 1px solid var(--color-border, #e7e7e7); border-radius: 6px; padding: 8px; }
+.seg-tag { display: inline-block; font-size: 12px; color: var(--color-text-secondary); margin-bottom: 4px; }
+.seg-text, .full-prompt { white-space: pre-wrap; word-break: break-word; font-family: inherit; margin: 0; font-size: 13px; line-height: 1.6; }
+.full-prompt { background: var(--color-bg-secondary, #f7f7f7); padding: 12px; border-radius: 6px; max-height: 480px; overflow: auto; }
+.full-prompt.small { max-height: 300px; font-size: 12px; }
+.detail-actions { margin-top: 12px; display: flex; gap: 8px; flex-wrap: wrap; }
+.dialog-footer { display: flex; justify-content: flex-end; gap: 8px; margin-top: 16px; }
+.compare-grid { display: grid; grid-template-columns: 1fr 1fr; gap: 16px; }
+.compare-col { min-width: 0; }
+.compare-head { font-weight: 600; margin-bottom: 8px; }
+.compare-meta { display: flex; align-items: center; gap: 8px; flex-wrap: wrap; margin-bottom: 8px; }
+.error-box { background: var(--td-error-color-1, #fdecec); color: var(--td-error-color, #d54941); padding: 8px 12px; border-radius: 6px; font-size: 13px; margin-bottom: 12px; word-break: break-word; white-space: pre-wrap; }
+.faq-box { border: 1px solid var(--color-border, #e7e7e7); border-radius: 6px; padding: 12px; display: flex; flex-direction: column; gap: 10px; }
+.faq-row { display: flex; gap: 16px; font-size: 13px; }
+.faq-label { flex-shrink: 0; width: 80px; color: var(--color-text-secondary); }
+.tl-head { display: flex; align-items: center; gap: 8px; margin-bottom: 6px; }
+.tl-user { font-size: 13px; color: var(--color-text-secondary); margin-bottom: 4px; white-space: pre-wrap; word-break: break-word; }
+.tl-ai { font-size: 13px; background: var(--color-bg-secondary, #f7f7f7); padding: 8px 12px; border-radius: 6px; white-space: pre-wrap; word-break: break-word; }
+.trend-toolbar { display: flex; align-items: center; gap: 12px; margin-bottom: 16px; flex-wrap: wrap; }
+.trend-chart-wrap { position: relative; height: 420px; }
+</style>

frontend/src/views/RoleManager.vue

@@ -184,6 +184,7 @@
 
 <script setup lang="ts">
 import { ref, reactive, computed, onMounted, watch } from 'vue'
+import { useRoute } from 'vue-router'
 import { useCategoryStore } from '@/stores/category'
 import { getAllRoles, createRole, updateRole, deleteRole, updateRoleCategories, updateRoleMcpTools } from '@/api/role'
 import { listAvailableMcpTools } from '@/api/mcp-server'
@@ -192,6 +193,7 @@ import { useConfirm } from '@/composables/useConfirm'
 
 const { confirm } = useConfirm()
 const categoryStore = useCategoryStore()
+const route = useRoute()
 
 // ---- 数据状态 ----
 const roles = ref<any[]>([])
@@ -334,10 +336,16 @@ function toggleTool(name: string, val: boolean): void {
 }
 
 // ---- 初始化 ----
-onMounted(() => {
-  reload()
+onMounted(async () => {
+  await reload()
   categoryStore.loadCategories()
   loadMcpTools()
+  // deep-link：从 URL 选中指定角色（如 /settings/role?roleId=xxx）
+  const roleId = route.query.roleId as string | undefined
+  if (roleId) {
+    const target = roles.value.find((r: any) => r.id === roleId)
+    if (target) await selectRole(target)
+  }
 })
 
 // ---- 监听 selectedRole 变化：被他人删除时自动回退 ----

frontend/src/views/SystemConfigManager.vue

@@ -174,6 +174,7 @@
 
 <script setup lang="ts">
 import { ref, reactive, computed, onMounted, watch } from 'vue'
+import { useRoute } from 'vue-router'
 import { SettingIcon, AddIcon, RefreshIcon, SearchIcon, DeleteIcon, CheckIcon } from 'tdesign-icons-vue-next'
 import DOMPurify from 'dompurify'
 import { listSystemConfigs, updateSystemConfig, deleteSystemConfig } from '@/api/system-config'
@@ -181,6 +182,7 @@ import { toast } from '@/utils/toast'
 import { useConfirm } from '@/composables/useConfirm'
 
 const { confirm } = useConfirm()
+const route = useRoute()
 
 // ==================== 类型预设目录 ====================
 type ConfigType = 'html' | 'text' | 'boolean'
@@ -291,7 +293,16 @@ function displayName(item: any): string {
 }
 
 // ==================== 初始化 ====================
-onMounted(() => load())
+onMounted(async () => {
+  await load()
+  // deep-link：从 URL 选中指定配置（如 /settings/system-config?key=ai_system_prompt）
+  const key = route.query.key as string | undefined
+  if (key) {
+    const target = configs.value.find((c: any) => c.configKey === key)
+    if (target) await selectConfig(target)
+    else toast('未找到配置项 ' + key + '，请先新建', 'warning')
+  }
+})
 
 // 监听 selectedConfig 变化：被他人删除时自动回退
 watch(selectedConfig, (cfg) => {

src/main/java/com/wok/supportbot/app/AssistantApp.java

@@ -5,13 +5,23 @@ import com.wok.supportbot.advisor.MyLoggerAdvisor;
 import com.wok.supportbot.chatmemory.DatabaseChatMemory;
 import com.wok.supportbot.config.ChatModelFactory;
 import com.wok.supportbot.config.SimpleCircuitBreaker;
+import com.wok.supportbot.entity.AiModelConfig;
+import com.wok.supportbot.entity.LlmCallTrace;
 import com.wok.supportbot.mcp.McpToolCallback;
+import com.wok.supportbot.mcp.McpToolCallback.ToolCallEvent;
 import com.wok.supportbot.mcp.McpToolCallbackAdapter;
+import com.wok.supportbot.service.AiModelConfigService;
+import com.wok.supportbot.service.ContentSafetyService;
+import com.wok.supportbot.service.LlmCallTraceService;
+import com.fasterxml.jackson.databind.ObjectMapper;
 import jakarta.annotation.Resource;
 import lombok.extern.slf4j.Slf4j;
 import org.springframework.ai.chat.client.ChatClient;
 import org.springframework.ai.chat.client.advisor.MessageChatMemoryAdvisor;
+import org.springframework.ai.chat.messages.Message;
+import org.springframework.ai.chat.metadata.Usage;
 import org.springframework.ai.chat.model.ChatModel;
+import org.springframework.ai.chat.model.ChatResponse;
 import org.springframework.ai.document.Document;
 import org.springframework.ai.tool.ToolCallback;
 import org.springframework.ai.vectorstore.VectorStore;
@@ -21,12 +31,16 @@ import org.springframework.util.StringUtils;
 import reactor.core.Disposable;
 import reactor.core.publisher.Flux;
 import reactor.core.publisher.FluxSink;
+import reactor.core.publisher.SignalType;
 
 import java.util.ArrayList;
 import java.util.Collections;
 import java.util.LinkedHashMap;
 import java.util.List;
 import java.util.Map;
+import java.util.concurrent.CopyOnWriteArrayList;
+import java.util.concurrent.atomic.AtomicInteger;
+import java.util.concurrent.atomic.AtomicReference;
 
 import static org.springframework.ai.chat.memory.ChatMemory.CONVERSATION_ID;
 
@@ -67,6 +81,15 @@ public class AssistantApp {
     @Resource
     private ChatPipeline chatPipeline;
 
+    @Resource
+    private AiModelConfigService aiModelConfigService;
+
+    @Resource
+    private LlmCallTraceService llmCallTraceService;
+
+    @Resource
+    private ContentSafetyService contentSafetyService;
+
     /** MCP 工具开关，默认启用，可通过 application.yml 的 chat.mcp.enabled 关闭 */
     @Value("${chat.mcp.enabled:true}")
     private boolean enableMcpTools;
@@ -102,6 +125,46 @@ public class AssistantApp {
     /** 尾部空白缓冲上限：超过后强制发出，避免纯空白输出导致 buffer 无界增长 */
     private static final int MAX_TRAILING_WHITESPACE_BUFFER = 256;
 
+    /** AI 回复落库的最大字符数，超过则保留头部 + 尾部 */
+    private static final int MAX_AI_RESPONSE_CHARS = 2000;
+
+    /** 截断时保留的头部字符数 */
+    private static final int AI_RESPONSE_HEAD_CHARS = 1500;
+
+    /** 截断时保留的尾部字符数 */
+    private static final int AI_RESPONSE_TAIL_CHARS = 500;
+
+    /** 埋点 JSON 序列化器（构建 ragHitsJson / toolCallsJson / historyMessagesJson） */
+    private static final ObjectMapper OBJECT_MAPPER = new ObjectMapper();
+
+    /** 历史消息每轮内容截断长度（避免 trace 行过大） */
+    private static final int HISTORY_MESSAGE_MAX_CHARS = 200;
+
+    /** 历史消息记录上限（条数） */
+    private static final int HISTORY_MESSAGE_MAX_COUNT = 10;
+
+    /** 错误消息截断长度 */
+    private static final int ERROR_MESSAGE_MAX_CHARS = 500;
+
+    /** MCP 工具调用结果落库截断长度（避免返回数据过大撑爆 trace 行） */
+    private static final int TOOL_CALL_RESULT_MAX_CHARS = 2000;
+
+    /** MCP 工具调用入参落库截断长度 */
+    private static final int TOOL_CALL_INPUT_MAX_CHARS = 500;
+
+    /**
+     * 埋点附加元信息：承载错误分类/消息、token 用量与 MCP 工具调用事件，
+     * 避免 recordTrace 参数过多。
+     */
+    private record TraceMeta(
+            String errorType,
+            String errorMessage,
+            Integer promptTokens,
+            Integer completionTokens,
+            Integer totalTokens,
+            List<ToolCallEvent> mcpEvents
+    ) {}
+
     /**
      * 初始化 ChatClient
      *
@@ -192,17 +255,24 @@ public class AssistantApp {
      * @return 回答文本 + MCP 事件
      */
     public ChatResult chatWithEvents(ChatContext ctx) {
-        // 熔断：全局 AI 调用处于熔断状态，直接返回降级提示
+        long startNanos = System.nanoTime();
+        // 熔断：全局 AI 调用处于熔断状态，直接返回降级提示（不做 buildRequest，避免熔断期间仍走意图路由/检索）
         if (aiCircuitBreaker.isOpen(AI_CIRCUIT_KEY)) {
             log.warn("AI 调用熔断中，返回降级提示");
+            recordTrace(ctx, null, CIRCUIT_OPEN_MESSAGE, 0, "BYPASS",
+                    new TraceMeta("CIRCUIT_BREAK", "AI 服务熔断降级", null, null, null, null));
             return new ChatResult(CIRCUIT_OPEN_MESSAGE, List.of());
         }
         ChatRequest req = chatPipeline.buildRequest(ctx);
         if (req.faqHit()) {
-            return new ChatResult(req.faqAnswer().get(), List.of());
+            String faqAnswer = req.faqAnswer().get();
+            recordTrace(ctx, req, faqAnswer, 0, "FAQ",
+                    new TraceMeta(null, null, null, null, null, null));
+            return new ChatResult(faqAnswer, List.of());
         }
-        McpToolCallback.resetEvents();
-        McpToolCallback.resetCallRounds();
+        // 显式事件收集器 + 轮次计数器，通过 toolContext 传给 McpToolCallback，规避 Reactor 跨线程丢 ThreadLocal 的问题
+        List<ToolCallEvent> events = new CopyOnWriteArrayList<>();
+        AtomicInteger rounds = new AtomicInteger(0);
         try {
             ChatClient.ChatClientRequestSpec spec = getChatClient(ctx.appType(), ctx.allowedMcpTools())
                     .prompt()
@@ -211,15 +281,29 @@ public class AssistantApp {
             if (StringUtils.hasText(req.finalSystemPrompt())) {
                 spec = spec.system(req.finalSystemPrompt());
             }
-            String text = spec.call().chatResponse().getResult().getOutput().getText();
+            spec = spec.toolContext(Map.of(
+                    McpToolCallback.MCP_EVENTS_KEY, events,
+                    McpToolCallback.MCP_ROUNDS_KEY, rounds));
+            ChatResponse response = spec.call().chatResponse();
+            String text = response.getResult().getOutput().getText();
+            Usage usage = response.getMetadata() != null ? response.getMetadata().getUsage() : null;
             aiCircuitBreaker.recordSuccess(AI_CIRCUIT_KEY);
+            recordTrace(ctx, req, text, elapsedMillis(startNanos), "COMPLETE",
+                    new TraceMeta(null, null,
+                            usage != null ? usage.getPromptTokens() : null,
+                            usage != null ? usage.getCompletionTokens() : null,
+                            usage != null ? usage.getTotalTokens() : null,
+                            events));
 
             // 推荐问题已不再由主回复同步生成，改由 SuggestionGenerator 异步按需生成
-            return new ChatResult(text, McpToolCallback.drainEvents(), List.of());
+            return new ChatResult(text, events, List.of());
         } catch (Exception e) {
             aiCircuitBreaker.recordFailure(AI_CIRCUIT_KEY);
             log.error("AI 同步调用失败: chatId={}, error={}", ctx.chatId(), e.getMessage());
-            return new ChatResult("抱歉，AI 服务调用失败：" + e.getMessage(), List.of());
+            String fallback = "抱歉，AI 服务调用失败：" + e.getMessage();
+            recordTrace(ctx, req, fallback, elapsedMillis(startNanos), "ERROR",
+                    new TraceMeta(classifyError(e), maskError(e.getMessage()), null, null, null, events));
+            return new ChatResult(fallback, List.of());
         }
     }
 
@@ -233,19 +317,26 @@ public class AssistantApp {
      * @return 纯文本流式回答（每个元素为一段自然语言文本）
      */
     public Flux<String> chatStream(ChatContext ctx) {
-        // 熔断：全局 AI 调用处于熔断状态
+        long startNanos = System.nanoTime();
+        // 熔断：全局 AI 调用处于熔断状态（不做 buildRequest，避免熔断期间仍走意图路由/检索）
         if (aiCircuitBreaker.isOpen(AI_CIRCUIT_KEY)) {
             log.warn("AI 调用熔断中（流式），返回降级提示");
+            recordTrace(ctx, null, CIRCUIT_OPEN_MESSAGE, 0, "BYPASS",
+                    new TraceMeta("CIRCUIT_BREAK", "AI 服务熔断降级", null, null, null, null));
             return Flux.just(CIRCUIT_OPEN_MESSAGE);
         }
         ChatRequest req = chatPipeline.buildRequest(ctx);
         if (req.faqHit()) {
             // FAQ 命中：整段答案原样输出，由 SSE 编码器处理内部换行，
             // 后端不做任何格式增删（不拆行、不加换行、不补空格）。
-            return Flux.just(req.faqAnswer().get());
+            String faqAnswer = req.faqAnswer().get();
+            recordTrace(ctx, req, faqAnswer, 0, "FAQ",
+                    new TraceMeta(null, null, null, null, null, null));
+            return Flux.just(faqAnswer);
         }
-        McpToolCallback.resetEvents();
-        McpToolCallback.resetCallRounds();
+        // 显式事件收集器 + 轮次计数器，通过 toolContext 传给 McpToolCallback，规避 Reactor 跨线程丢 ThreadLocal 的问题
+        List<ToolCallEvent> events = new CopyOnWriteArrayList<>();
+        AtomicInteger rounds = new AtomicInteger(0);
         ChatClient.ChatClientRequestSpec spec = getChatClient(ctx.appType(), ctx.allowedMcpTools())
                 .prompt()
                 .user(req.finalMessage())
@@ -253,18 +344,47 @@ public class AssistantApp {
         if (StringUtils.hasText(req.finalSystemPrompt())) {
             spec = spec.system(req.finalSystemPrompt());
         }
-        // 原始文本流；推荐问题已不再由主回复同步生成，改由 SuggestionGenerator 异步按需生成
-        Flux<String> rawStream = spec.stream().content();
+        spec = spec.toolContext(Map.of(
+                McpToolCallback.MCP_EVENTS_KEY, events,
+                McpToolCallback.MCP_ROUNDS_KEY, rounds));
+        // 改为 chatResponse 流以采集 token 用量，再映射回纯文本流
+        AtomicReference<Usage> usageRef = new AtomicReference<>();
+        AtomicReference<String> errorTypeRef = new AtomicReference<>();
+        AtomicReference<String> errorMessageRef = new AtomicReference<>();
+        Flux<ChatResponse> responseFlux = spec.stream().chatResponse();
+        Flux<String> rawStream = responseFlux
+                .doOnNext(r -> {
+                    if (r != null && r.getMetadata() != null && r.getMetadata().getUsage() != null) {
+                        usageRef.set(r.getMetadata().getUsage());
+                    }
+                })
+                .map(r -> {
+                    String out = r != null && r.getResult() != null && r.getResult().getOutput() != null
+                            ? r.getResult().getOutput().getText() : "";
+                    return out != null ? out : "";
+                });
+        // 聚合所有分片用于埋点（在 doFinally 时取完整回复文本）
+        StringBuilder aggregated = new StringBuilder();
         return preserveTrailingWhitespace(rawStream)
+                .doOnNext(aggregated::append)
                 .doOnComplete(() -> aiCircuitBreaker.recordSuccess(AI_CIRCUIT_KEY))
                 .doOnError(e -> {
                     aiCircuitBreaker.recordFailure(AI_CIRCUIT_KEY);
+                    errorTypeRef.set(classifyError(e));
+                    errorMessageRef.set(maskError(e.getMessage()));
                     log.error("AI 流式调用失败: chatId={}, error={}", ctx.chatId(), e.getMessage());
                 })
                 .doFinally(signalType -> {
-                    // 确保 ThreadLocal 清理，防止线程池复用时数据残留
-                    McpToolCallback.resetEvents();
-                    McpToolCallback.resetCallRounds();
+                    // 流式埋点：按终止信号区分状态，断连/异常也落库（events 由 toolContext 显式收集，跨线程安全）
+                    String status = signalType == SignalType.ON_COMPLETE ? "COMPLETE"
+                            : signalType == SignalType.ON_ERROR ? "ERROR" : "CANCEL";
+                    Usage usage = usageRef.get();
+                    recordTrace(ctx, req, aggregated.toString(), elapsedMillis(startNanos), status,
+                            new TraceMeta(errorTypeRef.get(), errorMessageRef.get(),
+                                    usage != null ? usage.getPromptTokens() : null,
+                                    usage != null ? usage.getCompletionTokens() : null,
+                                    usage != null ? usage.getTotalTokens() : null,
+                                    events));
                 })
                 .onErrorResume(e -> Flux.just("抱歉，AI 服务调用失败：" + e.getMessage()));
     }
@@ -325,6 +445,238 @@ public class AssistantApp {
         }, FluxSink.OverflowStrategy.BUFFER);
     }
 
+    /**
+     * 组装并异步写入一条 LLM 调用追踪记录。
+     * <p>
+     * 所有字段均在调用现场（非异步线程）组装为实体，避免把 ChatRequest/ChatContext
+     * 全量对象塞进异步队列。落库前对用户消息与 AI 回复做敏感词脱敏与长度截断。
+     *
+     * @param ctx          对话上下文
+     * @param req          编排决策（熔断早退时为 null）
+     * @param responseText AI 回复文本
+     * @param latencyMs    耗时（毫秒）
+     * @param status       状态：COMPLETE / ERROR / CANCEL / FAQ / BYPASS
+     */
+    private void recordTrace(ChatContext ctx, ChatRequest req, String responseText, long latencyMs,
+                             String status, TraceMeta meta) {
+        try {
+            // 走缓存的活跃配置（仅取模型元信息，不落 apiKey），避免每次对话在 Reactor 线程同步查库
+            AiModelConfig cfg = aiModelConfigService.getActiveConfigWithFullKey(ctx.appType());
+            // AI 回复脱敏 + 截断（保留头部 + 尾部，避免丢失末尾的推荐问题块）
+            String maskedResponse = contentSafetyService.mask(responseText);
+            String aiResponse = maskedResponse;
+            boolean truncated = false;
+            if (maskedResponse != null && maskedResponse.length() > MAX_AI_RESPONSE_CHARS) {
+                aiResponse = maskedResponse.substring(0, AI_RESPONSE_HEAD_CHARS)
+                        + "\n\n…（内容过长已截断）…\n\n"
+                        + maskedResponse.substring(maskedResponse.length() - AI_RESPONSE_TAIL_CHARS);
+                truncated = true;
+            }
+
+            // FAQ 命中详情（faqMatchResult 来自编排决策，未命中为 null）
+            Long faqId = null;
+            String faqQuestion = null;
+            String faqMatchType = null;
+            Double faqScore = null;
+            if (req != null && req.faqMatchResult() != null && req.faqMatchResult().getFaq() != null) {
+                faqId = req.faqMatchResult().getFaq().getId();
+                faqQuestion = req.faqMatchResult().getFaq().getQuestion();
+                faqMatchType = req.faqMatchResult().getMatchType();
+                faqScore = req.faqMatchResult().getScore();
+            }
+
+            // 历史消息（一次读取，同时用于 JSON 与条数）
+            List<Message> history = safeGetHistory(ctx.chatId());
+
+            LlmCallTrace trace = LlmCallTrace.builder()
+                    .conversationId(ctx.chatId())
+                    .roleId(ctx.roleId())
+                    .roleName(ctx.roleName())
+                    .accountId(ctx.accountId())
+                    .apiKeyId(ctx.apiKeyId())
+                    .intent(req != null ? req.intent() : null)
+                    .enableRag(ctx.enableRag())
+                    .systemPrompt(req != null ? contentSafetyService.mask(req.finalSystemPrompt()) : null)
+                    .globalPrompt(req != null ? contentSafetyService.mask(req.globalPrompt()) : null)
+                    .rolePrompt(contentSafetyService.mask(ctx.systemPrompt()))
+                    .userMessage(contentSafetyService.mask(ctx.message()))
+                    .aiResponse(aiResponse)
+                    .aiResponseTruncated(truncated)
+                    .ragContext(req != null ? contentSafetyService.mask(req.ragContextText()) : null)
+                    .ragHitsJson(buildRagHitsJson(req))
+                    .faqHit(req != null ? req.faqHit() : false)
+                    .faqId(faqId)
+                    .faqQuestion(faqQuestion)
+                    .faqMatchType(faqMatchType)
+                    .faqScore(faqScore)
+                    .searchMode(req != null ? req.searchMode() : null)
+                    .hitCount(req != null ? req.hitCount() : null)
+                    .toolCallsJson(buildToolCallsJson(meta != null ? meta.mcpEvents() : null))
+                    .historyMessagesJson(buildHistoryJson(history))
+                    .historyTurns(history != null ? history.size() : null)
+                    .modelName(cfg != null ? cfg.getModelName() : null)
+                    .provider(cfg != null ? cfg.getProvider() : null)
+                    .temperature(cfg != null ? cfg.getTemperature() : null)
+                    .maxTokens(cfg != null ? cfg.getMaxTokens() : null)
+                    .promptTokens(meta != null ? meta.promptTokens() : null)
+                    .completionTokens(meta != null ? meta.completionTokens() : null)
+                    .totalTokens(meta != null ? meta.totalTokens() : null)
+                    .errorType(meta != null ? meta.errorType() : null)
+                    .errorMessage(meta != null ? contentSafetyService.mask(meta.errorMessage()) : null)
+                    .latencyMs((int) latencyMs)
+                    .status(status)
+                    .build();
+            llmCallTraceService.recordAsync(trace);
+        } catch (Exception e) {
+            log.warn("构造 LLM 调用追踪失败（不影响主流程）: chatId={}, error={}", ctx.chatId(), e.getMessage());
+        }
+    }
+
+    /**
+     * 序列化 RAG 命中文档片段为 JSON（含 documentId/title/chunkIndex/sourceName/score/searchMode）。
+     * 无命中返回 null。
+     */
+    private String buildRagHitsJson(ChatRequest req) {
+        if (req == null || req.hitDocuments() == null || req.hitDocuments().isEmpty()) {
+            return null;
+        }
+        try {
+            List<Map<String, Object>> items = new ArrayList<>();
+            for (Document doc : req.hitDocuments()) {
+                Map<String, Object> meta = doc.getMetadata();
+                Map<String, Object> item = new LinkedHashMap<>();
+                item.put("documentId", meta.get("documentId"));
+                item.put("title", meta.get("title"));
+                item.put("chunkIndex", meta.get("chunkIndex"));
+                item.put("sourceName", meta.get("sourceName"));
+                // 距离字段在不同检索实现下可能是 distance 或 score，二者取一
+                Object score = meta.get("distance") != null ? meta.get("distance") : meta.get("score");
+                item.put("score", score);
+                item.put("searchMode", req.searchMode());
+                items.add(item);
+            }
+            return OBJECT_MAPPER.writeValueAsString(items);
+        } catch (Exception e) {
+            log.warn("序列化 RAG 命中片段失败: {}", e.getMessage());
+            return null;
+        }
+    }
+
+    /**
+     * 序列化 MCP 工具调用事件为 JSON（input/result 先脱敏再截断，避免返回数据过大撑爆 trace 行）。
+     * 无事件返回 null。
+     */
+    private String buildToolCallsJson(List<ToolCallEvent> events) {
+        if (events == null || events.isEmpty()) {
+            return null;
+        }
+        try {
+            List<Map<String, Object>> items = new ArrayList<>();
+            for (ToolCallEvent e : events) {
+                Map<String, Object> item = new LinkedHashMap<>();
+                item.put("tool", e.tool());
+                item.put("input", truncateToolCallText(contentSafetyService.mask(e.input()), TOOL_CALL_INPUT_MAX_CHARS));
+                item.put("result", truncateToolCallText(contentSafetyService.mask(e.result()), TOOL_CALL_RESULT_MAX_CHARS));
+                item.put("latencyMs", e.latencyMs());
+                item.put("error", e.error());
+                items.add(item);
+            }
+            return OBJECT_MAPPER.writeValueAsString(items);
+        } catch (Exception ex) {
+            log.warn("序列化 MCP 工具调用事件失败: {}", ex.getMessage());
+            return null;
+        }
+    }
+
+    /**
+     * 对工具调用 input/result 做长度截断（保留头部 + 截断标记）。
+     *
+     * @param text     脱敏后的文本
+     * @param maxChars 最大保留字符数
+     * @return 未超限返回原文本；超限返回前 maxChars 个字符 + 截断标记
+     */
+    private String truncateToolCallText(String text, int maxChars) {
+        if (text == null || text.length() <= maxChars) {
+            return text;
+        }
+        int total = text.length();
+        return text.substring(0, maxChars) + "…（共 " + total + " 字符，已截断）";
+    }
+
+    /** 从会话记忆读取最近若干条历史消息（失败返回 null）。 */
+    private List<Message> safeGetHistory(String chatId) {
+        try {
+            return chatMemory.get(chatId, HISTORY_MESSAGE_MAX_COUNT);
+        } catch (Exception e) {
+            return null;
+        }
+    }
+
+    /** 将历史消息序列化为 JSON（每轮内容截断）。无历史返回 null。 */
+    private String buildHistoryJson(List<Message> history) {
+        if (history == null || history.isEmpty()) {
+            return null;
+        }
+        try {
+            List<Map<String, String>> items = new ArrayList<>();
+            for (Message m : history) {
+                String role = m.getMessageType() != null ? m.getMessageType().name() : "unknown";
+                String content = m.getText() != null ? m.getText() : "";
+                if (content.length() > HISTORY_MESSAGE_MAX_CHARS) {
+                    content = content.substring(0, HISTORY_MESSAGE_MAX_CHARS) + "…";
+                }
+                items.add(Map.of("role", role, "content", content));
+            }
+            return OBJECT_MAPPER.writeValueAsString(items);
+        } catch (Exception e) {
+            return null;
+        }
+    }
+
+    /**
+     * 按异常类型分类错误，供 trace.error_type 落库。
+     */
+    private String classifyError(Throwable t) {
+        if (t == null) {
+            return "UNKNOWN";
+        }
+        String name = t.getClass().getSimpleName();
+        String msg = t.getMessage() == null ? "" : t.getMessage().toLowerCase();
+        if (name.contains("Mcp") || msg.contains("mcp") || msg.contains("tool")) {
+            return "MCP";
+        }
+        if (name.contains("CircuitBreaker") || msg.contains("circuit")) {
+            return "CIRCUIT_BREAK";
+        }
+        if (name.contains("Validation") || name.contains("IllegalArgument")) {
+            return "VALIDATION";
+        }
+        if (name.contains("Ai") || name.contains("OpenAi") || name.contains("DashScope")
+                || name.contains("Http") || name.contains("Timeout") || msg.contains("timeout")) {
+            return "LLM_API";
+        }
+        return "UNKNOWN";
+    }
+
+    /** 对异常消息做脱敏并截断，供 trace.error_message 落库。 */
+    private String maskError(String message) {
+        if (message == null || message.isBlank()) {
+            return null;
+        }
+        String masked = contentSafetyService.mask(message);
+        if (masked != null && masked.length() > ERROR_MESSAGE_MAX_CHARS) {
+            return masked.substring(0, ERROR_MESSAGE_MAX_CHARS) + "…";
+        }
+        return masked;
+    }
+
+    /**
+     * 计算自 startNanos 起的耗时（毫秒）。
+     */
+    private long elapsedMillis(long startNanos) {
+        return (System.nanoTime() - startNanos) / 1_000_000;
+    }
+
     /**
      * 统一检索引用来源（新入口，委托 {@link ChatPipeline#retrieveSources}）。
      *

src/main/java/com/wok/supportbot/app/ChatContext.java

@@ -20,6 +20,10 @@ import java.util.List;
  * @param rewriteStrategy RAG 查询重写策略（REWRITE / TRANSLATION / COMPRESSION / MULTI_QUERY），可为 null
  * @param enableRag       是否启用 RAG 检索；false=普通对话
  * @param streaming       是否流式输出
+ * @param roleId          客服角色 ID（可空，供调用追踪等使用）
+ * @param roleName        客服角色名（可空，快照用途）
+ * @param accountId       账户 ID（可空，供调用追踪/个人信息删除权使用）
+ * @param apiKeyId        API Key ID（可空，Open API 路径为实际鉴权 Key，供租户隔离与追踪）
  */
 public record ChatContext(
         String message,
@@ -30,7 +34,11 @@ public record ChatContext(
         List<Long> categoryIds,
         String rewriteStrategy,
         boolean enableRag,
-        boolean streaming
+        boolean streaming,
+        Long roleId,
+        String roleName,
+        String accountId,
+        Long apiKeyId
 ) {
 
     /** 默认应用类型 */
@@ -49,30 +57,38 @@ public record ChatContext(
      * 便捷构造：仅指定核心字段，其余取默认值（非 RAG、非流式）。
      */
     public static ChatContext of(String message, String chatId) {
-        return new ChatContext(message, chatId, DEFAULT_APP_TYPE, null, null, null, null, false, false);
+        return new ChatContext(message, chatId, DEFAULT_APP_TYPE, null, null, null, null, false, false, null, null, null, null);
     }
 
     public ChatContext withSystemPrompt(String systemPrompt) {
-        return new ChatContext(message, chatId, appType, systemPrompt, allowedMcpTools, categoryIds, rewriteStrategy, enableRag, streaming);
+        return new ChatContext(message, chatId, appType, systemPrompt, allowedMcpTools, categoryIds, rewriteStrategy, enableRag, streaming, roleId, roleName, accountId, apiKeyId);
     }
 
     public ChatContext withAllowedMcpTools(List<String> allowedMcpTools) {
-        return new ChatContext(message, chatId, appType, systemPrompt, allowedMcpTools, categoryIds, rewriteStrategy, enableRag, streaming);
+        return new ChatContext(message, chatId, appType, systemPrompt, allowedMcpTools, categoryIds, rewriteStrategy, enableRag, streaming, roleId, roleName, accountId, apiKeyId);
     }
 
     public ChatContext withCategoryIds(List<Long> categoryIds) {
-        return new ChatContext(message, chatId, appType, systemPrompt, allowedMcpTools, categoryIds, rewriteStrategy, enableRag, streaming);
+        return new ChatContext(message, chatId, appType, systemPrompt, allowedMcpTools, categoryIds, rewriteStrategy, enableRag, streaming, roleId, roleName, accountId, apiKeyId);
     }
 
     public ChatContext withRewriteStrategy(String rewriteStrategy) {
-        return new ChatContext(message, chatId, appType, systemPrompt, allowedMcpTools, categoryIds, rewriteStrategy, enableRag, streaming);
+        return new ChatContext(message, chatId, appType, systemPrompt, allowedMcpTools, categoryIds, rewriteStrategy, enableRag, streaming, roleId, roleName, accountId, apiKeyId);
     }
 
     public ChatContext withEnableRag(boolean enableRag) {
-        return new ChatContext(message, chatId, appType, systemPrompt, allowedMcpTools, categoryIds, rewriteStrategy, enableRag, streaming);
+        return new ChatContext(message, chatId, appType, systemPrompt, allowedMcpTools, categoryIds, rewriteStrategy, enableRag, streaming, roleId, roleName, accountId, apiKeyId);
     }
 
     public ChatContext withStreaming(boolean streaming) {
-        return new ChatContext(message, chatId, appType, systemPrompt, allowedMcpTools, categoryIds, rewriteStrategy, enableRag, streaming);
+        return new ChatContext(message, chatId, appType, systemPrompt, allowedMcpTools, categoryIds, rewriteStrategy, enableRag, streaming, roleId, roleName, accountId, apiKeyId);
+    }
+
+    public ChatContext withAccountId(String accountId) {
+        return new ChatContext(message, chatId, appType, systemPrompt, allowedMcpTools, categoryIds, rewriteStrategy, enableRag, streaming, roleId, roleName, accountId, apiKeyId);
+    }
+
+    public ChatContext withApiKeyId(Long apiKeyId) {
+        return new ChatContext(message, chatId, appType, systemPrompt, allowedMcpTools, categoryIds, rewriteStrategy, enableRag, streaming, roleId, roleName, accountId, apiKeyId);
     }
 }

src/main/java/com/wok/supportbot/app/ChatPipeline.java

@@ -2,6 +2,7 @@ package com.wok.supportbot.app;
 
 import com.wok.supportbot.rag.RagContext;
 import com.wok.supportbot.rag.RagPipeline;
+import com.wok.supportbot.service.FaqMatchEngine.FaqMatchResult;
 import com.wok.supportbot.service.IntentRouter;
 import com.wok.supportbot.service.SystemConfigService;
 import com.wok.supportbot.service.RagHitLogService;
@@ -72,11 +73,13 @@ public class ChatPipeline {
      * @return 执行决策
      */
     public ChatRequest buildRequest(ChatContext ctx) {
-        String baseSystem = effectiveSystem(ctx.systemPrompt());
+        String globalPrompt = systemConfigService.getValueByKey("ai_system_prompt");
+        String baseSystem = effectiveSystem(ctx.systemPrompt(), globalPrompt);
 
         // 普通对话：enableRag=false（含 Controller 层 isKbDenied 强制置 false 的情况）
         if (!ctx.enableRag()) {
-            return new ChatRequest(ctx, ctx.message(), baseSystem, Optional.empty());
+            return new ChatRequest(ctx, ctx.message(), baseSystem, Optional.empty(),
+                    globalPrompt, null, null, "CHAT", null, null, null);
         }
 
         // 意图路由：先用 IntentRouter 做细粒度分类
@@ -85,10 +88,12 @@ public class ChatPipeline {
         // FAQ 高置信度：优先匹配标准答案；未命中时降级到 RAG 检索，避免知识库中已有答案却返回兜底提示
         if (intent != null && "FAQ".equals(intent.getIntent())
                 && intent.getConfidence() >= FAQ_HIGH_CONFIDENCE_THRESHOLD) {
-            Optional<String> faqAnswer = ragPipeline.tryFaqMatch(ctx.message());
-            if (faqAnswer.isPresent()) {
+            Optional<FaqMatchResult> faqMatch = ragPipeline.tryFaqMatchResult(ctx.message());
+            if (faqMatch.isPresent()) {
                 log.info("FAQ 高置信({})，命中标准答案: chatId={}", intent.getConfidence(), ctx.chatId());
-                return new ChatRequest(ctx, ctx.message(), baseSystem, faqAnswer);
+                Optional<String> faqAnswer = Optional.ofNullable(faqMatch.get().getFaq().getAnswer());
+                return new ChatRequest(ctx, ctx.message(), baseSystem, faqAnswer,
+                        globalPrompt, null, null, "FAQ", null, null, faqMatch.get());
             }
             log.info("FAQ 高置信({}) 未命中标准答案，降级到 RAG 检索: chatId={}", intent.getConfidence(), ctx.chatId());
         }
@@ -96,7 +101,8 @@ public class ChatPipeline {
         // 寒暄/闲聊：IntentRouter 判定 CHITCHAT 高置信，跳过 KB 检索
         if (intent != null && "CHITCHAT".equals(intent.getIntent())
                 && intent.getConfidence() >= CHITCHAT_CONFIDENCE_THRESHOLD) {
-            return new ChatRequest(ctx, ctx.message(), baseSystem, Optional.empty());
+            return new ChatRequest(ctx, ctx.message(), baseSystem, Optional.empty(),
+                    globalPrompt, null, null, "CHITCHAT", null, null, null);
         }
 
         // RAG 检索（含 FAQ 优先匹配）
@@ -118,12 +124,15 @@ public class ChatPipeline {
             ragHitLogService.recordMiss(ctx.chatId(), ctx.message(), searchMode);
         }
         if (rag.faqHit()) {
-            return new ChatRequest(ctx, ctx.message(), baseSystem, rag.faqAnswer());
+            return new ChatRequest(ctx, ctx.message(), baseSystem, rag.faqAnswer(),
+                    globalPrompt, null, null, "FAQ", null, null, rag.faqMatchResult());
         }
 
         // RAG 生成：资料块注入 system，重写后查询作为 user 消息
         String finalSystem = baseSystem + ragPipeline.buildRagContextBlock(rag.contextText());
-        return new ChatRequest(ctx, rag.rewrittenQuery(), finalSystem, Optional.empty());
+        return new ChatRequest(ctx, rag.rewrittenQuery(), finalSystem, Optional.empty(),
+                globalPrompt, rag.contextText(), rag.documents() != null ? rag.documents().size() : 0, "RAG",
+                rag.searchMode(), rag.documents(), null);
     }
 
     /**
@@ -193,11 +202,10 @@ public class ChatPipeline {
      * 注意：推荐问题（suggest-message-list）已不再由主回复生成，
      * 改由 {@link SuggestionGenerator} 在 AI 回复结束后按需异步生成。
      */
-    private String effectiveSystem(String rolePrompt) {
+    private String effectiveSystem(String rolePrompt, String globalPrompt) {
         StringBuilder sb = new StringBuilder();
 
-        // 全局系统提示词（从 DB system_config 表动态读取）
-        String globalPrompt = systemConfigService.getValueByKey("ai_system_prompt");
+        // 全局系统提示词（由 buildRequest 读取后传入，避免重复查询）
         if (StringUtils.hasText(globalPrompt)) {
             sb.append(globalPrompt);
         }

src/main/java/com/wok/supportbot/app/ChatRequest.java

@@ -1,5 +1,9 @@
 package com.wok.supportbot.app;
 
+import com.wok.supportbot.service.FaqMatchEngine.FaqMatchResult;
+import org.springframework.ai.document.Document;
+
+import java.util.List;
 import java.util.Optional;
 
 /**
@@ -13,18 +17,39 @@ import java.util.Optional;
  *   <li>{@link #faqAnswer()}：FAQ 命中时直接返回标准答案，跳过 ChatClient 调用</li>
  *   <li>{@link #finalMessage()}：传给模型的用户消息（RAG 场景为重写后的查询，否则为原始 message）</li>
  *   <li>{@link #finalSystemPrompt()}：传给模型的系统提示词（角色人设 + RAG 资料块），可为空</li>
+ *   <li>{@link #globalPrompt()}：全局提示词快照（来自 system_config.ai_system_prompt），可为空</li>
+ *   <li>{@link #ragContextText()}：RAG 资料块文本（未注入时为空/null）</li>
+ *   <li>{@link #hitCount()}：命中的知识库片段数（非 RAG 为 null）</li>
+ *   <li>{@link #intent()}：本次决策命中的意图：CHAT / CHITCHAT / FAQ / RAG</li>
+ *   <li>{@link #searchMode()}：真实检索模式 VECTOR / KEYWORD / HYBRID（非 RAG 为 null）</li>
+ *   <li>{@link #hitDocuments()}：命中的知识库片段（RAG 场景，非 RAG 为 null）</li>
+ *   <li>{@link #faqMatchResult()}：FAQ 命中详情（含 matchType/score，未命中为 null）</li>
  * </ul>
  *
  * @param ctx              原始上下文
  * @param finalMessage     传给模型的用户消息
  * @param finalSystemPrompt 传给模型的系统提示词，可为 null/空
  * @param faqAnswer        FAQ 命中答案，未命中为 {@link Optional#empty()}
+ * @param globalPrompt     全局提示词快照，可为 null
+ * @param ragContextText   RAG 资料块文本，可为 null
+ * @param hitCount         命中文档数，可为 null
+ * @param intent           意图：CHAT / CHITCHAT / FAQ / RAG
+ * @param searchMode       真实检索模式，可为 null
+ * @param hitDocuments     命中的知识库片段，可为 null
+ * @param faqMatchResult   FAQ 命中详情，可为 null
  */
 public record ChatRequest(
         ChatContext ctx,
         String finalMessage,
         String finalSystemPrompt,
-        Optional<String> faqAnswer
+        Optional<String> faqAnswer,
+        String globalPrompt,
+        String ragContextText,
+        Integer hitCount,
+        String intent,
+        String searchMode,
+        List<Document> hitDocuments,
+        FaqMatchResult faqMatchResult
 ) {
 
     /** FAQ 是否命中 */

src/main/java/com/wok/supportbot/config/AsyncExecutorConfig.java

@@ -0,0 +1,36 @@
+package com.wok.supportbot.config;
+
+import org.springframework.context.annotation.Bean;
+import org.springframework.context.annotation.Configuration;
+import org.springframework.scheduling.concurrent.ThreadPoolTaskExecutor;
+
+import java.util.concurrent.ThreadPoolExecutor;
+
+/**
+ * 异步执行器配置。
+ * <p>
+ * 为日志/追踪类异步写入提供有界线程池，避免 Spring 默认 {@code SimpleAsyncTaskExecutor}
+ * 每任务新建一线程导致的高并发线程爆炸与无背压问题。
+ */
+@Configuration
+public class AsyncExecutorConfig {
+
+    /**
+     * LLM 追踪等日志类异步写入线程池。
+     * 有界队列 + CallerRunsPolicy：队列满时由调用线程执行，保证不丢数据（背压优先于丢弃）。
+     */
+    @Bean("traceExecutor")
+    public ThreadPoolTaskExecutor traceExecutor() {
+        ThreadPoolTaskExecutor executor = new ThreadPoolTaskExecutor();
+        executor.setCorePoolSize(4);
+        executor.setMaxPoolSize(8);
+        executor.setQueueCapacity(1000);
+        executor.setThreadNamePrefix("llm-trace-");
+        executor.setRejectedExecutionHandler(new ThreadPoolExecutor.CallerRunsPolicy());
+        // 优雅停机时等待队列中的追踪/日志任务完成，避免丢记录
+        executor.setWaitForTasksToCompleteOnShutdown(true);
+        executor.setAwaitTerminationSeconds(10);
+        executor.initialize();
+        return executor;
+    }
+}

src/main/java/com/wok/supportbot/config/DatabaseInitConfig.java

@@ -185,6 +185,12 @@ public class DatabaseInitConfig {
                 createDashboardSnapshotTable();
             }
         });
+        safeInit("创建 LLM 调用追踪表 llm_call_trace", () -> {
+            if (!checkTableExists("llm_call_trace")) {
+                createLlmCallTraceTable();
+            }
+        });
+        safeInit("迁移 llm_call_trace 扩展列", this::addLlmCallTraceExtensionColumns);
 
         // P1-003: API 开放平台
         safeInit("创建 API Key 表 api_key", () -> {
@@ -260,7 +266,8 @@ public class DatabaseInitConfig {
                 "rag_hit_log", "dashboard_snapshot",
                 "api_key", "webhook_config",
                 "mcp_server_config",
-                "system_config"
+                "system_config",
+                "llm_call_trace"
         };
 
         java.util.List<String> missingTables = new java.util.ArrayList<>();
@@ -1091,6 +1098,80 @@ public class DatabaseInitConfig {
         jdbcTemplate.execute("CREATE INDEX IF NOT EXISTS idx_rag_hit_log_document ON rag_hit_log (document_id)");
     }
 
+    private void createLlmCallTraceTable() {
+        String sql = """
+            CREATE TABLE IF NOT EXISTS llm_call_trace (
+                id                    BIGINT          PRIMARY KEY,
+                conversation_id       VARCHAR(64),
+                role_id               BIGINT,
+                role_name             VARCHAR(100),
+                account_id            VARCHAR(64),
+                api_key_id            BIGINT,
+                intent                VARCHAR(16),
+                enable_rag            BOOLEAN,
+                system_prompt         TEXT,
+                global_prompt         TEXT,
+                role_prompt           TEXT,
+                user_message          TEXT,
+                ai_response           TEXT,
+                ai_response_truncated BOOLEAN,
+                tool_calls_json       TEXT,
+                history_messages_json TEXT,
+                history_turns         INTEGER,
+                rag_context           TEXT,
+                faq_hit               BOOLEAN,
+                faq_id                BIGINT,
+                faq_question          TEXT,
+                faq_match_type        VARCHAR(16),
+                faq_score             DOUBLE PRECISION,
+                search_mode           VARCHAR(20),
+                hit_count             INTEGER,
+                rag_hits_json         TEXT,
+                model_name            VARCHAR(128),
+                provider              VARCHAR(64),
+                temperature           DOUBLE PRECISION,
+                max_tokens            INTEGER,
+                prompt_tokens         INTEGER,
+                completion_tokens     INTEGER,
+                total_tokens          INTEGER,
+                error_type            VARCHAR(32),
+                error_message         TEXT,
+                latency_ms            INTEGER,
+                status                VARCHAR(16),
+                create_time           TIMESTAMP       DEFAULT CURRENT_TIMESTAMP NOT NULL
+            )
+            """;
+        jdbcTemplate.execute(sql);
+        jdbcTemplate.execute("CREATE INDEX IF NOT EXISTS idx_llm_trace_created ON llm_call_trace (create_time DESC, id DESC)");
+        jdbcTemplate.execute("CREATE INDEX IF NOT EXISTS idx_llm_trace_role_created ON llm_call_trace (role_id, create_time DESC)");
+        jdbcTemplate.execute("CREATE INDEX IF NOT EXISTS idx_llm_trace_conv_created ON llm_call_trace (conversation_id, create_time DESC)");
+        jdbcTemplate.execute("CREATE INDEX IF NOT EXISTS idx_llm_trace_error_type ON llm_call_trace (error_type, create_time DESC)");
+        jdbcTemplate.execute("CREATE INDEX IF NOT EXISTS idx_llm_trace_api_key ON llm_call_trace (api_key_id, create_time DESC)");
+        jdbcTemplate.execute("CREATE INDEX IF NOT EXISTS idx_llm_trace_search_mode ON llm_call_trace (search_mode, create_time DESC)");
+    }
+
+    /**
+     * 为已存在的 llm_call_trace 表补加扩展列（幂等，供旧部署环境迁移）。
+     */
+    private void addLlmCallTraceExtensionColumns() {
+        jdbcTemplate.execute("ALTER TABLE llm_call_trace ADD COLUMN IF NOT EXISTS tool_calls_json TEXT");
+        jdbcTemplate.execute("ALTER TABLE llm_call_trace ADD COLUMN IF NOT EXISTS history_messages_json TEXT");
+        jdbcTemplate.execute("ALTER TABLE llm_call_trace ADD COLUMN IF NOT EXISTS history_turns INTEGER");
+        jdbcTemplate.execute("ALTER TABLE llm_call_trace ADD COLUMN IF NOT EXISTS faq_id BIGINT");
+        jdbcTemplate.execute("ALTER TABLE llm_call_trace ADD COLUMN IF NOT EXISTS faq_question TEXT");
+        jdbcTemplate.execute("ALTER TABLE llm_call_trace ADD COLUMN IF NOT EXISTS faq_match_type VARCHAR(16)");
+        jdbcTemplate.execute("ALTER TABLE llm_call_trace ADD COLUMN IF NOT EXISTS faq_score DOUBLE PRECISION");
+        jdbcTemplate.execute("ALTER TABLE llm_call_trace ADD COLUMN IF NOT EXISTS rag_hits_json TEXT");
+        jdbcTemplate.execute("ALTER TABLE llm_call_trace ADD COLUMN IF NOT EXISTS prompt_tokens INTEGER");
+        jdbcTemplate.execute("ALTER TABLE llm_call_trace ADD COLUMN IF NOT EXISTS completion_tokens INTEGER");
+        jdbcTemplate.execute("ALTER TABLE llm_call_trace ADD COLUMN IF NOT EXISTS total_tokens INTEGER");
+        jdbcTemplate.execute("ALTER TABLE llm_call_trace ADD COLUMN IF NOT EXISTS error_type VARCHAR(32)");
+        jdbcTemplate.execute("ALTER TABLE llm_call_trace ADD COLUMN IF NOT EXISTS error_message TEXT");
+        jdbcTemplate.execute("CREATE INDEX IF NOT EXISTS idx_llm_trace_error_type ON llm_call_trace (error_type, create_time DESC)");
+        jdbcTemplate.execute("CREATE INDEX IF NOT EXISTS idx_llm_trace_api_key ON llm_call_trace (api_key_id, create_time DESC)");
+        jdbcTemplate.execute("CREATE INDEX IF NOT EXISTS idx_llm_trace_search_mode ON llm_call_trace (search_mode, create_time DESC)");
+    }
+
     private void createDashboardSnapshotTable() {
         String sql = """
             CREATE TABLE IF NOT EXISTS dashboard_snapshot (
@@ -1255,6 +1336,13 @@ public class DatabaseInitConfig {
             VALUES (?, ?, ?)
             ON CONFLICT (config_key) DO NOTHING
             """, "suggestion_prompt", defaultSuggestionPrompt, "AI 推荐问题 Prompt 模板（仅在 suggestion_enabled=true 时生效）");
+
+        // LLM 调用追踪记录保留天数（自动清理，默认 30 天）
+        jdbcTemplate.update("""
+            INSERT INTO system_config (config_key, config_value, description)
+            VALUES (?, ?, ?)
+            ON CONFLICT (config_key) DO NOTHING
+            """, "llm_trace_retention_days", "30", "LLM 调用追踪记录保留天数（自动清理，默认 30 天）");
     }
 
     /**
@@ -1505,6 +1593,34 @@ public class DatabaseInitConfig {
             // ===== api_key.role_ids =====
             executeComment("COLUMN api_key.role_ids", "绑定的客服角色 ID 列表（JSONB 数组，空数组=不限制，返回所有启用角色）");
 
+            // ===== llm_call_trace =====
+            executeComment("TABLE llm_call_trace", "LLM 调用追踪表（记录每次 LLM 调用的 system prompt/回复/模型参数/耗时，用于提示词优化调试，append-only）");
+            executeComment("COLUMN llm_call_trace.id", "主键（雪花算法生成）");
+            executeComment("COLUMN llm_call_trace.conversation_id", "会话 ID");
+            executeComment("COLUMN llm_call_trace.role_id", "客服角色 ID（可空）");
+            executeComment("COLUMN llm_call_trace.role_name", "角色名称快照（角色改名后仍可追溯）");
+            executeComment("COLUMN llm_call_trace.account_id", "账户 ID（可空，预留租户/个人信息删除权）");
+            executeComment("COLUMN llm_call_trace.api_key_id", "API Key ID（可空，预留租户隔离）");
+            executeComment("COLUMN llm_call_trace.intent", "意图：CHAT / CHITCHAT / FAQ / RAG");
+            executeComment("COLUMN llm_call_trace.enable_rag", "是否启用 RAG 增强");
+            executeComment("COLUMN llm_call_trace.system_prompt", "最终注入 LLM 的完整 system prompt");
+            executeComment("COLUMN llm_call_trace.global_prompt", "全局提示词快照（可空）");
+            executeComment("COLUMN llm_call_trace.role_prompt", "角色提示词快照（可空）");
+            executeComment("COLUMN llm_call_trace.user_message", "用户原始消息（脱敏后）");
+            executeComment("COLUMN llm_call_trace.ai_response", "AI 回复（截断，保留头部+尾部）");
+            executeComment("COLUMN llm_call_trace.ai_response_truncated", "AI 回复是否被截断");
+            executeComment("COLUMN llm_call_trace.rag_context", "RAG 资料块（可空）");
+            executeComment("COLUMN llm_call_trace.faq_hit", "是否 FAQ 命中");
+            executeComment("COLUMN llm_call_trace.search_mode", "检索模式：VECTOR / KEYWORD / HYBRID（可空）");
+            executeComment("COLUMN llm_call_trace.hit_count", "命中文档数（可空）");
+            executeComment("COLUMN llm_call_trace.model_name", "模型名称");
+            executeComment("COLUMN llm_call_trace.provider", "提供商");
+            executeComment("COLUMN llm_call_trace.temperature", "温度参数");
+            executeComment("COLUMN llm_call_trace.max_tokens", "最大 Token");
+            executeComment("COLUMN llm_call_trace.latency_ms", "调用耗时（毫秒）");
+            executeComment("COLUMN llm_call_trace.status", "状态：COMPLETE / ERROR / CANCEL / FAQ / BYPASS");
+            executeComment("COLUMN llm_call_trace.create_time", "创建时间");
+
             log.info("数据库表注释已应用");
         } catch (Exception e) {
             log.warn("应用数据库表注释时出错", e);

src/main/java/com/wok/supportbot/controller/AiController.java

@@ -118,7 +118,7 @@ public class AiController {
         try {
             List<Long> cats = resolveCategoryIds(scope, categoryId, categoryIds);
             ChatContext ctx = new ChatContext(message, chatId, "CHAT", null, null, cats,
-                    normalizeStrategy(rewriteStrategy), true, false);
+                    normalizeStrategy(rewriteStrategy), true, false, context.roleId(), scope.name(), context.accountId(), null);
             List<Document> docs = assistantApp.retrieveSources(ctx);
             List<Map<String, Object>> out = new ArrayList<>();
             for (Document doc : docs) {
@@ -148,7 +148,8 @@ public class AiController {
         bindConversation(chatId, context);
         RoleScope scope = customerServiceRoleService.getRoleScope(context.roleId());
         return new ChatContext(message, chatId, "CHAT", resolveSystemPrompt(scope, systemPrompt),
-                scope.hasRole() ? scope.allowedMcpTools() : null, null, null, false, false);
+                scope.hasRole() ? scope.allowedMcpTools() : null, null, null, false, false,
+                context.roleId(), scope.name(), context.accountId(), null);
     }
 
     /** 构造 RAG 对话的 ChatContext（含严格隔离判断：KbDenied 则 enableRag=false）。 */
@@ -161,7 +162,7 @@ public class AiController {
         boolean enableRag = !isKbDenied(scope);
         List<Long> cats = resolveCategoryIds(scope, categoryId, categoryIds);
         return new ChatContext(message, chatId, "CHAT", sys, scope.hasRole() ? scope.allowedMcpTools() : null, cats,
-                normalizeStrategy(rewriteStrategy), enableRag, false);
+                normalizeStrategy(rewriteStrategy), enableRag, false, context.roleId(), scope.name(), context.accountId(), null);
     }
 
     // ==================== SDK 会话管理接口（带账户归属校验） ====================
@@ -354,7 +355,8 @@ public class AiController {
             ctx = buildChatContext(message, chatId, roleId, accountId, systemPrompt);
         }
         ctx = new ChatContext(ctx.message(), ctx.chatId(), ctx.appType(), ctx.systemPrompt(),
-                ctx.allowedMcpTools(), ctx.categoryIds(), ctx.rewriteStrategy(), ctx.enableRag(), true);
+                ctx.allowedMcpTools(), ctx.categoryIds(), ctx.rewriteStrategy(), ctx.enableRag(), true,
+                ctx.roleId(), ctx.roleName(), ctx.accountId(), ctx.apiKeyId());
         return assistantApp.chatStream(ctx);
     }
 

src/main/java/com/wok/supportbot/controller/LlmCallTraceController.java

@@ -0,0 +1,192 @@
+package com.wok.supportbot.controller;
+
+import com.wok.supportbot.entity.LlmCallTrace;
+import com.wok.supportbot.service.LlmCallTraceService;
+import lombok.extern.slf4j.Slf4j;
+import org.springframework.beans.factory.annotation.Autowired;
+import org.springframework.http.ResponseEntity;
+import org.springframework.security.access.prepost.PreAuthorize;
+import org.springframework.security.core.Authentication;
+import org.springframework.security.core.context.SecurityContextHolder;
+import org.springframework.web.bind.annotation.GetMapping;
+import org.springframework.web.bind.annotation.PathVariable;
+import org.springframework.web.bind.annotation.PostMapping;
+import org.springframework.web.bind.annotation.RequestBody;
+import org.springframework.web.bind.annotation.RequestMapping;
+import org.springframework.web.bind.annotation.RequestParam;
+import org.springframework.web.bind.annotation.RestController;
+
+import java.util.LinkedHashMap;
+import java.util.List;
+import java.util.Map;
+
+/**
+ * LLM 调用追踪接口（仅 admin 可访问）。
+ * <p>
+ * 提供调用记录的分页查询、详情、聚合统计与清理。
+ * 端点仅挂管理路径（/llm-trace），绝不进入 /ai/** 或 /open-api/**（SDK 开放路径），
+ * 防止 conversation_id 跨租户越权。
+ */
+@Slf4j
+@RestController
+@RequestMapping("/llm-trace")
+@PreAuthorize("hasRole('admin')")
+public class LlmCallTraceController {
+
+    @Autowired
+    private LlmCallTraceService llmCallTraceService;
+
+    /** 清理的最小保留天数（防止误清全部） */
+    private static final int MIN_KEEP_DAYS = 1;
+
+    /** 清理的最大保留天数（防止超过 PostgreSQL make_interval 合理范围） */
+    private static final int MAX_KEEP_DAYS = 3650;
+
+    /**
+     * 分页查询调用记录（列表不含大 TEXT 字段）。
+     */
+    @GetMapping("/list")
+    public ResponseEntity<Map<String, Object>> list(
+            @RequestParam(defaultValue = "1") int page,
+            @RequestParam(defaultValue = "20") int size,
+            @RequestParam(required = false) Long roleId,
+            @RequestParam(required = false) String conversationId,
+            @RequestParam(required = false) String intent,
+            @RequestParam(required = false) String startTime,
+            @RequestParam(required = false) String endTime,
+            @RequestParam(required = false) String keyword,
+            @RequestParam(required = false) String errorType) {
+        try {
+            Map<String, Object> result = llmCallTraceService.pageQuery(
+                    page, size, roleId, conversationId, intent, startTime, endTime, keyword, errorType);
+            Map<String, Object> data = new LinkedHashMap<>();
+            data.put("success", true);
+            data.put("data", result.get("records"));
+            data.put("total", result.get("total"));
+            data.put("page", result.get("page"));
+            data.put("size", result.get("size"));
+            data.put("pages", result.get("pages"));
+            return ResponseEntity.ok(data);
+        } catch (Exception e) {
+            log.error("查询 LLM 调用追踪失败", e);
+            return ResponseEntity.status(500).body(Map.of(
+                    "success", false,
+                    "message", "查询失败：" + e.getMessage()
+            ));
+        }
+    }
+
+    /**
+     * 单条详情（含完整 system prompt）。
+     */
+    @GetMapping("/{id}")
+    public ResponseEntity<Map<String, Object>> detail(@PathVariable Long id) {
+        try {
+            LlmCallTrace trace = llmCallTraceService.getDetail(id);
+            if (trace == null) {
+                return ResponseEntity.status(404).body(Map.of(
+                        "success", false,
+                        "message", "调用记录不存在"
+                ));
+            }
+            return ResponseEntity.ok(Map.of("success", true, "data", trace));
+        } catch (Exception e) {
+            log.error("查询 LLM 调用追踪详情失败: id={}", id, e);
+            return ResponseEntity.status(500).body(Map.of(
+                    "success", false,
+                    "message", "查询失败：" + e.getMessage()
+            ));
+        }
+    }
+
+    /**
+     * 清理 N 天前的记录（POST，避免 DELETE + RequestBody 路径冲突）。
+     * 保留天数设下限，防止误清全部。
+     */
+    @PostMapping("/clean")
+    public ResponseEntity<Map<String, Object>> clean(@RequestBody Map<String, Object> body) {
+        try {
+            Object raw = body.get("keepDays");
+            int keepDays = raw instanceof Number n ? n.intValue() : Integer.parseInt(String.valueOf(raw));
+            if (keepDays < MIN_KEEP_DAYS || keepDays > MAX_KEEP_DAYS) {
+                return ResponseEntity.badRequest().body(Map.of(
+                        "success", false,
+                        "message", "保留天数必须在 " + MIN_KEEP_DAYS + " ~ " + MAX_KEEP_DAYS + " 天之间"
+                ));
+            }
+            int deleted = llmCallTraceService.cleanBefore(keepDays);
+            log.info("管理员清理 LLM 调用追踪：操作人={}, 保留 {} 天, 删除 {} 条", currentUser(), keepDays, deleted);
+            return ResponseEntity.ok(Map.of(
+                    "success", true,
+                    "message", "清理完成",
+                    "deleted", deleted
+            ));
+        } catch (Exception e) {
+            log.error("清理 LLM 调用追踪失败", e);
+            return ResponseEntity.status(500).body(Map.of(
+                    "success", false,
+                    "message", "清理失败：" + e.getMessage()
+            ));
+        }
+    }
+
+    /**
+     * 聚合统计（按角色或模型分组）。
+     */
+    @GetMapping("/stats")
+    public ResponseEntity<Map<String, Object>> stats(@RequestParam(defaultValue = "role") String groupBy) {
+        try {
+            List<Map<String, Object>> rows = llmCallTraceService.stats(groupBy);
+            return ResponseEntity.ok(Map.of("success", true, "data", rows));
+        } catch (Exception e) {
+            log.error("查询 LLM 调用追踪统计失败", e);
+            return ResponseEntity.status(500).body(Map.of(
+                    "success", false,
+                    "message", "统计失败：" + e.getMessage()
+            ));
+        }
+    }
+
+    /**
+     * 可选错误类型列表（供前端筛选下拉）。
+     */
+    @GetMapping("/error-types")
+    public ResponseEntity<Map<String, Object>> errorTypes() {
+        List<Map<String, String>> types = List.of(
+                Map.of("label", "LLM 调用失败", "value", "LLM_API"),
+                Map.of("label", "MCP 工具异常", "value", "MCP"),
+                Map.of("label", "熔断降级", "value", "CIRCUIT_BREAK"),
+                Map.of("label", "参数校验", "value", "VALIDATION"),
+                Map.of("label", "未知", "value", "UNKNOWN")
+        );
+        return ResponseEntity.ok(Map.of("success", true, "data", types));
+    }
+
+    /**
+     * 时间维度趋势统计（按小时或天聚合）。
+     */
+    @GetMapping("/trend")
+    public ResponseEntity<Map<String, Object>> trend(
+            @RequestParam(defaultValue = "HOUR") String groupBy,
+            @RequestParam(required = false) String startTime,
+            @RequestParam(required = false) String endTime) {
+        try {
+            List<Map<String, Object>> rows = llmCallTraceService.trend(groupBy, startTime, endTime);
+            return ResponseEntity.ok(Map.of("success", true, "data", rows));
+        } catch (Exception e) {
+            log.error("查询 LLM 调用追踪趋势失败", e);
+            return ResponseEntity.status(500).body(Map.of(
+                    "success", false,
+                    "message", "统计失败：" + e.getMessage()
+            ));
+        }
+    }
+
+    /**
+     * 获取当前登录用户名（用于审计日志，best-effort）。
+     */
+    private String currentUser() {
+        Authentication auth = SecurityContextHolder.getContext().getAuthentication();
+        return auth != null ? auth.getName() : "unknown";
+    }
+}

src/main/java/com/wok/supportbot/controller/OpenApiController.java

@@ -225,7 +225,7 @@ public class OpenApiController {
 
         return new ChatContext(message, chatId, "CHAT", systemPrompt,
                 scope.hasRole() ? scope.allowedMcpTools() : null,
-                catIds, strategy, useRag, streaming);
+                catIds, strategy, useRag, streaming, roleId, scope.name(), null, apiKey.getId());
     }
 
     /**

src/main/java/com/wok/supportbot/dao/LlmCallTraceMapper.java

@@ -0,0 +1,12 @@
+package com.wok.supportbot.dao;
+
+import com.baomidou.mybatisplus.core.mapper.BaseMapper;
+import com.wok.supportbot.entity.LlmCallTrace;
+import org.apache.ibatis.annotations.Mapper;
+
+/**
+ * LLM 调用追踪 Mapper
+ */
+@Mapper
+public interface LlmCallTraceMapper extends BaseMapper<LlmCallTrace> {
+}

src/main/java/com/wok/supportbot/entity/LlmCallTrace.java

@@ -0,0 +1,192 @@
+package com.wok.supportbot.entity;
+
+import com.baomidou.mybatisplus.annotation.FieldFill;
+import com.baomidou.mybatisplus.annotation.IdType;
+import com.baomidou.mybatisplus.annotation.TableField;
+import com.baomidou.mybatisplus.annotation.TableId;
+import com.baomidou.mybatisplus.annotation.TableName;
+import com.fasterxml.jackson.databind.annotation.JsonSerialize;
+import com.fasterxml.jackson.databind.ser.std.ToStringSerializer;
+import lombok.AllArgsConstructor;
+import lombok.Builder;
+import lombok.Data;
+import lombok.NoArgsConstructor;
+
+import java.io.Serial;
+import java.io.Serializable;
+import java.util.Date;
+
+/**
+ * LLM 调用追踪实体
+ * <p>
+ * 记录每次 LLM 对话的完整"案发现场"（最终 system prompt、分段来源、用户消息、AI 回复、
+ * 角色、模型参数、耗时、意图等），供管理员反推优化提示词。
+ * append-only，无逻辑删除。
+ */
+@Data
+@Builder
+@AllArgsConstructor
+@NoArgsConstructor
+@TableName("llm_call_trace")
+public class LlmCallTrace implements Serializable {
+
+    @Serial
+    @TableField(exist = false)
+    private static final long serialVersionUID = 1L;
+
+    /** 主键ID（雪花算法） */
+    @TableId(value = "id", type = IdType.ASSIGN_ID)
+    @JsonSerialize(using = ToStringSerializer.class)
+    private Long id;
+
+    /** 会话ID */
+    @TableField("conversation_id")
+    private String conversationId;
+
+    /** 客服角色ID（可空，无角色时为 null） */
+    @TableField("role_id")
+    @JsonSerialize(using = ToStringSerializer.class)
+    private Long roleId;
+
+    /** 角色名称快照（角色改名后仍可追溯） */
+    @TableField("role_name")
+    private String roleName;
+
+    /** 账户ID（可空，预留租户/个人信息删除权） */
+    @TableField("account_id")
+    private String accountId;
+
+    /** API Key ID（可空，预留租户隔离） */
+    @TableField("api_key_id")
+    @JsonSerialize(using = ToStringSerializer.class)
+    private Long apiKeyId;
+
+    /** 意图：CHAT / CHITCHAT / FAQ / RAG */
+    @TableField("intent")
+    private String intent;
+
+    /** 是否启用 RAG 增强 */
+    @TableField("enable_rag")
+    private Boolean enableRag;
+
+    /** 最终注入 LLM 的完整 system prompt */
+    @TableField("system_prompt")
+    private String systemPrompt;
+
+    /** 全局提示词快照（可空） */
+    @TableField("global_prompt")
+    private String globalPrompt;
+
+    /** 角色提示词快照（可空） */
+    @TableField("role_prompt")
+    private String rolePrompt;
+
+    /** 用户原始消息（脱敏后） */
+    @TableField("user_message")
+    private String userMessage;
+
+    /** AI 回复（截断，保留头部+尾部） */
+    @TableField("ai_response")
+    private String aiResponse;
+
+    /** AI 回复是否被截断 */
+    @TableField("ai_response_truncated")
+    private Boolean aiResponseTruncated;
+
+    /** MCP 工具调用事件 JSON 数组（可空） */
+    @TableField("tool_calls_json")
+    private String toolCallsJson;
+
+    /** 本次注入 LLM 的历史消息 JSON（可空） */
+    @TableField("history_messages_json")
+    private String historyMessagesJson;
+
+    /** 历史消息轮数（可空） */
+    @TableField("history_turns")
+    private Integer historyTurns;
+
+    /** RAG 资料块（可空） */
+    @TableField("rag_context")
+    private String ragContext;
+
+    /** 是否 FAQ 命中 */
+    @TableField("faq_hit")
+    private Boolean faqHit;
+
+    /** FAQ 命中 ID（可空，未命中或非 FAQ 意图为 null） */
+    @TableField("faq_id")
+    @JsonSerialize(using = ToStringSerializer.class)
+    private Long faqId;
+
+    /** FAQ 标准问题快照（可空） */
+    @TableField("faq_question")
+    private String faqQuestion;
+
+    /** FAQ 匹配类型：EXACT / KEYWORD / SEMANTIC（可空） */
+    @TableField("faq_match_type")
+    private String faqMatchType;
+
+    /** FAQ 匹配分数（0.0 ~ 1.0，可空） */
+    @TableField("faq_score")
+    private Double faqScore;
+
+    /** 检索模式：VECTOR / KEYWORD / HYBRID（可空） */
+    @TableField("search_mode")
+    private String searchMode;
+
+    /** 命中文档数（可空） */
+    @TableField("hit_count")
+    private Integer hitCount;
+
+    /** RAG 命中片段详情 JSON 数组（含 documentId/title/chunkIndex/score/searchMode，可空） */
+    @TableField("rag_hits_json")
+    private String ragHitsJson;
+
+    /** 模型名称 */
+    @TableField("model_name")
+    private String modelName;
+
+    /** 提供商 */
+    @TableField("provider")
+    private String provider;
+
+    /** 温度参数 */
+    @TableField("temperature")
+    private Double temperature;
+
+    /** 最大 Token */
+    @TableField("max_tokens")
+    private Integer maxTokens;
+
+    /** 提示词 token 数（可空，非 LLM 调用路径无此值） */
+    @TableField("prompt_tokens")
+    private Integer promptTokens;
+
+    /** 生成 token 数（可空） */
+    @TableField("completion_tokens")
+    private Integer completionTokens;
+
+    /** 总 token 数（可空） */
+    @TableField("total_tokens")
+    private Integer totalTokens;
+
+    /** 调用耗时（毫秒） */
+    @TableField("latency_ms")
+    private Integer latencyMs;
+
+    /** 错误类型：LLM_API / MCP / CIRCUIT_BREAK / VALIDATION / UNKNOWN（可空，非失败调用为 null） */
+    @TableField("error_type")
+    private String errorType;
+
+    /** 错误原始消息（已脱敏，可空） */
+    @TableField("error_message")
+    private String errorMessage;
+
+    /** 状态：COMPLETE / ERROR / CANCEL / FAQ / BYPASS */
+    @TableField("status")
+    private String status;
+
+    /** 创建时间 */
+    @TableField(value = "create_time", fill = FieldFill.INSERT)
+    private Date createTime;
+}

src/main/java/com/wok/supportbot/mcp/McpToolCallback.java

@@ -6,6 +6,7 @@ import com.wok.supportbot.config.McpClientManager;
 import io.modelcontextprotocol.client.McpSyncClient;
 import io.modelcontextprotocol.spec.McpSchema;
 import lombok.extern.slf4j.Slf4j;
+import org.springframework.ai.chat.model.ToolContext;
 import org.springframework.ai.model.ModelOptionsUtils;
 import org.springframework.ai.tool.ToolCallback;
 import org.springframework.ai.tool.definition.DefaultToolDefinition;
@@ -14,6 +15,7 @@ import org.springframework.ai.tool.definition.ToolDefinition;
 import java.util.ArrayList;
 import java.util.List;
 import java.util.Map;
+import java.util.concurrent.atomic.AtomicInteger;
 
 /**
  * MCP Tool -> Spring AI ToolCallback 适配器
@@ -30,6 +32,11 @@ public class McpToolCallback implements ToolCallback {
 
     private static final ObjectMapper OBJECT_MAPPER = new ObjectMapper();
 
+    /**
+     * 日志输出内容的最大长度，避免 MCP 工具返回数据过大导致日志刷屏。
+     */
+    private static final int MAX_LOG_LENGTH = 1000;
+
     // ==================== 工具调用事件收集 ====================
 
     /**
@@ -37,9 +44,20 @@ public class McpToolCallback implements ToolCallback {
      */
     public record ToolCallEvent(String tool, String input, String result, long latencyMs, boolean error) {}
 
+    /**
+     * ToolContext 中携带事件收集器的 key（AssistantApp 通过 ChatClientRequestSpec.toolContext 注入）。
+     * 用于解决 Reactor 流式场景下 ThreadLocal 跨线程丢失工具调用事件的问题。
+     */
+    public static final String MCP_EVENTS_KEY = "mcp_tool_events";
+
+    /**
+     * ToolContext 中携带调用轮次计数器的 key，避免跨线程时轮次限制失效。
+     */
+    public static final String MCP_ROUNDS_KEY = "mcp_tool_rounds";
+
     /**
      * 线程级事件收集器：在同一请求线程中收集所有工具调用事件
-     * SSE 流式输出完成后，从这里取出事件发送给前端
+     * 作为无 toolContext 场景（同步调用 / 兼容旧逻辑）的兜底。
      */
     private static final ThreadLocal<List<ToolCallEvent>> EVENTS = ThreadLocal.withInitial(ArrayList::new);
 
@@ -166,17 +184,57 @@ public class McpToolCallback implements ToolCallback {
      */
     @Override
     public String call(String toolInput) {
+        // 无 ToolContext 时回退到 ThreadLocal 收集器（同步调用 / 兼容旧逻辑）
+        return doCall(toolInput, EVENTS.get(), null);
+    }
+
+    /**
+     * 带 ToolContext 的执行入口。
+     * <p>
+     * 优先从 ToolContext 读取 AssistantApp 注入的事件收集器与轮次计数器，
+     * 解决 Reactor 流式场景下 ThreadLocal 跨线程丢失的问题；读取不到时回退 ThreadLocal。
+     */
+    @Override
+    @SuppressWarnings("unchecked")
+    public String call(String toolInput, ToolContext toolContext) {
+        List<ToolCallEvent> collector = EVENTS.get();
+        AtomicInteger rounds = null;
+        Map<String, Object> ctx = toolContext != null ? toolContext.getContext() : null;
+        if (ctx != null) {
+            Object ev = ctx.get(MCP_EVENTS_KEY);
+            if (ev instanceof List<?> list) {
+                collector = (List<ToolCallEvent>) list;
+            }
+            Object rd = ctx.get(MCP_ROUNDS_KEY);
+            if (rd instanceof AtomicInteger ai) {
+                rounds = ai;
+            }
+        }
+        return doCall(toolInput, collector, rounds);
+    }
+
+    /**
+     * 工具调用主逻辑。
+     *
+     * @param toolInput JSON 格式的工具输入参数
+     * @param collector 事件收集器（非 null，事件追加到此列表）
+     * @param rounds    调用轮次计数器（可 null，为 null 时回退 ThreadLocal 计数）
+     * @return 工具执行结果（JSON 字符串形式）
+     */
+    private String doCall(String toolInput, List<ToolCallEvent> collector, AtomicInteger rounds) {
         // 检查调用轮次是否超限
-        int currentRound = CALL_ROUNDS.get();
+        int currentRound = rounds != null ? rounds.getAndIncrement() : CALL_ROUNDS.get();
         if (currentRound >= maxCallRounds) {
             log.warn("MCP 工具调用轮次超限: tool={}, currentRound={}, maxRounds={}",
                     originalToolName, currentRound, maxCallRounds);
             return "{\"error\": \"工具调用轮次已达上限 (" + maxCallRounds + " 次)，已终止调用以防止无限循环。请优化提示词减少工具调用次数。\"}";
         }
-        CALL_ROUNDS.set(currentRound + 1);
+        if (rounds == null) {
+            CALL_ROUNDS.set(currentRound + 1);
+        }
 
         log.info("MCP 工具调用: serverId={}, tool={}, input={}, round={}/{}",
-                mcpServerConfigId, originalToolName, toolInput, currentRound + 1, maxCallRounds);
+                mcpServerConfigId, originalToolName, truncate(toolInput), currentRound + 1, maxCallRounds);
         long startTime = System.currentTimeMillis();
 
         try {
@@ -194,22 +252,22 @@ public class McpToolCallback implements ToolCallback {
             McpSchema.CallToolRequest request = new McpSchema.CallToolRequest(originalToolName, arguments);
             McpSchema.CallToolResult result = client.callTool(request);
 
-            // 调试日志：输出 MCP 调用原始返回结果
+            // 调试日志：输出 MCP 调用原始返回结果（截断，避免返回数据过大刷屏）
             log.info("🔧 [DEBUG] MCP callTool 原始结果 - serverId={}, tool={}, rawJson={}",
-                    mcpServerConfigId, originalToolName, ModelOptionsUtils.toJsonString(result));
+                    mcpServerConfigId, originalToolName, truncate(ModelOptionsUtils.toJsonString(result)));
 
             long latency = System.currentTimeMillis() - startTime;
             log.info("MCP 工具调用完成: tool={}, latency={}ms, isError={}",
                     originalToolName, latency, result.isError());
 
-            // 收集工具调用事件（供 SSE 流式输出使用）
+            // 收集工具调用事件（供提示词追踪 / SSE 流式输出使用）
             String resultStr = result.content() != null ? String.valueOf(result.content()) : "";
             boolean isError = result.isError() != null && result.isError();
-            EVENTS.get().add(new ToolCallEvent(originalToolName, toolInput, resultStr, latency, isError));
+            collector.add(new ToolCallEvent(originalToolName, toolInput, resultStr, latency, isError));
 
             // 检查是否为错误结果
             if (result.isError() != null && result.isError()) {
-                log.error("MCP 工具返回错误: tool={}, content={}", originalToolName, result.content());
+                log.error("MCP 工具返回错误: tool={}, content={}", originalToolName, truncate(String.valueOf(result.content())));
                 return "{\"error\": \"工具执行返回错误: " +
                         escapeJson(String.valueOf(result.content())) + "\"}";
             }
@@ -217,7 +275,7 @@ public class McpToolCallback implements ToolCallback {
             // 将 MCP Content 列表序列化为 JSON 字符串返回给 AI 模型
             // 与官方 SyncMcpToolCallback 保持一致，使用 ModelOptionsUtils 序列化
             String resultJson = ModelOptionsUtils.toJsonString(result.content());
-            log.debug("MCP 工具调用结果: tool={}, result={}", originalToolName, resultJson);
+            log.debug("MCP 工具调用结果: tool={}, result={}", originalToolName, truncate(resultJson));
             return resultJson;
 
         } catch (Exception e) {
@@ -235,16 +293,16 @@ public class McpToolCallback implements ToolCallback {
                         McpSchema.CallToolRequest retryRequest = new McpSchema.CallToolRequest(originalToolName, retryArgs);
                         McpSchema.CallToolResult retryResult = reconnected.callTool(retryRequest);
 
-                        // 调试日志：输出 MCP 重试调用原始返回结果
+                        // 调试日志：输出 MCP 重试调用原始返回结果（截断）
                         log.info("🔧 [DEBUG] MCP callTool 重试原始结果 - serverId={}, tool={}, rawJson={}",
-                                mcpServerConfigId, originalToolName, ModelOptionsUtils.toJsonString(retryResult));
+                                mcpServerConfigId, originalToolName, truncate(ModelOptionsUtils.toJsonString(retryResult)));
 
                         long retryLatency = System.currentTimeMillis() - startTime;
                         log.info("MCP 工具重试成功: tool={}, latency={}ms", originalToolName, retryLatency);
 
                         String retryResultStr = retryResult.content() != null ? String.valueOf(retryResult.content()) : "";
                         boolean retryIsError = retryResult.isError() != null && retryResult.isError();
-                        EVENTS.get().add(new ToolCallEvent(originalToolName, toolInput, retryResultStr, retryLatency, retryIsError));
+                        collector.add(new ToolCallEvent(originalToolName, toolInput, retryResultStr, retryLatency, retryIsError));
 
                         if (retryResult.isError() != null && retryResult.isError()) {
                             return "{\"error\": \"工具执行返回错误: " + escapeJson(String.valueOf(retryResult.content())) + "\"}";
@@ -257,11 +315,24 @@ public class McpToolCallback implements ToolCallback {
             }
 
             // 记录失败事件
-            EVENTS.get().add(new ToolCallEvent(originalToolName, toolInput, e.getMessage(), latency, true));
+            collector.add(new ToolCallEvent(originalToolName, toolInput, e.getMessage(), latency, true));
             return "{\"error\": \"" + escapeJson(e.getMessage()) + "\"}";
         }
     }
 
+    /**
+     * 截断日志内容，避免 MCP 工具返回数据过大导致日志刷屏。
+     *
+     * @param text 原始文本
+     * @return 未超限时返回原文本；超限时返回前 MAX_LOG_LENGTH 个字符 + 截断标记
+     */
+    private String truncate(String text) {
+        if (text == null || text.length() <= MAX_LOG_LENGTH) {
+            return text;
+        }
+        return text.substring(0, MAX_LOG_LENGTH) + "…（共 " + text.length() + " 字符，已截断）";
+    }
+
     /**
      * 转义字符串中的特殊字符，避免破坏 JSON 格式
      */

src/main/java/com/wok/supportbot/rag/RagContext.java

@@ -1,5 +1,6 @@
 package com.wok.supportbot.rag;
 
+import com.wok.supportbot.service.FaqMatchEngine.FaqMatchResult;
 import org.springframework.ai.document.Document;
 
 import java.util.List;
@@ -19,12 +20,16 @@ import java.util.Optional;
  * @param documents     检索命中的知识库片段（含 metadata），可为空
  * @param contextText   拼接后的资料文本（用 {@code "\n\n---\n\n"} 分隔）；无资料时为空串
  * @param rewrittenQuery 传给模型的用户消息：MULTI_QUERY 为原始 message，其余策略为重写后的查询
+ * @param searchMode    真实检索模式（VECTOR / KEYWORD / HYBRID），当前主管道固定为 VECTOR
+ * @param faqMatchResult FAQ 命中详情（含 matchType/score）；未命中为 null
  */
 public record RagContext(
         Optional<String> faqAnswer,
         List<Document> documents,
         String contextText,
-        String rewrittenQuery
+        String rewrittenQuery,
+        String searchMode,
+        FaqMatchResult faqMatchResult
 ) {
 
     /** FAQ 是否命中 */

src/main/java/com/wok/supportbot/rag/RagPipeline.java

@@ -8,6 +8,7 @@ import com.wok.supportbot.rag.preretrieval.MultiQueryExpanderRewriter;
 import com.wok.supportbot.rag.preretrieval.RewriteQueryRewriter;
 import com.wok.supportbot.rag.preretrieval.TranslationQueryRewriter;
 import com.wok.supportbot.service.FaqMatchEngine;
+import com.wok.supportbot.service.FaqMatchEngine.FaqMatchResult;
 import com.wok.supportbot.service.RagHitLogService;
 import jakarta.annotation.Resource;
 import lombok.extern.slf4j.Slf4j;
@@ -108,10 +109,12 @@ public class RagPipeline {
      */
     public RagContext retrieve(ChatContext ctx) {
         // 1. FAQ 优先匹配：命中则直接返回标准答案，跳过检索与生成
-        Optional<String> faqAnswer = tryFaqMatch(ctx.message());
-        if (faqAnswer.isPresent()) {
-            log.info("FAQ 命中，跳过知识库检索: chatId={}", ctx.chatId());
-            return new RagContext(faqAnswer, Collections.emptyList(), "", ctx.message());
+        Optional<FaqMatchResult> faqMatch = tryFaqMatchResult(ctx.message());
+        if (faqMatch.isPresent()) {
+            log.info("FAQ 命中，跳过知识库检索: chatId={}, matchType={}", ctx.chatId(), faqMatch.get().getMatchType());
+            String answer = faqMatch.get().getFaq().getAnswer();
+            return new RagContext(Optional.ofNullable(answer), Collections.emptyList(), "", ctx.message(),
+                    currentSearchMode(), faqMatch.get());
         }
 
         // 2. 统一检索 + 组装结果
@@ -150,7 +153,7 @@ public class RagPipeline {
         logRagHit(ctx.chatId(), ctx.message(), docs, rewrittenQuery);
 
         String contextText = joinContext(docs);
-        return new RagContext(Optional.empty(), docs, contextText, rewrittenQuery);
+        return new RagContext(Optional.empty(), docs, contextText, rewrittenQuery, currentSearchMode(), null);
     }
 
     /**
@@ -172,19 +175,36 @@ public class RagPipeline {
     // ==================== FAQ 匹配 ====================
 
     /**
-     * 尝试 FAQ 三级匹配（精确→关键词→语义），命中返回标准答案。
-     * 异常时降级为未命中，供 ChatPipeline 等调用方直接获取 FAQ 匹配结果。
+     * 尝试 FAQ 三级匹配（精确→关键词→语义），命中返回完整匹配结果（含 matchType/score）。
+     * 异常时降级为未命中。
      */
-    public Optional<String> tryFaqMatch(String message) {
+    public Optional<FaqMatchResult> tryFaqMatchResult(String message) {
         try {
-            return faqMatchEngine.match(message)
-                    .map(result -> result.getFaq().getAnswer());
+            return faqMatchEngine.match(message);
         } catch (Exception e) {
             log.warn("FAQ 匹配异常，降级到 RAG: {}", e.getMessage());
             return Optional.empty();
         }
     }
 
+    /**
+     * 尝试 FAQ 三级匹配，命中返回标准答案（仅答案文本）。
+     * 异常时降级为未命中，供仅需答案的调用方使用。
+     */
+    public Optional<String> tryFaqMatch(String message) {
+        return tryFaqMatchResult(message).map(result -> result.getFaq().getAnswer());
+    }
+
+    /**
+     * 当前主管道的真实检索模式。
+     * <p>
+     * 目前仅单路向量检索（VECTOR）；HybridSearchService 接入主管道后，
+     * 此方法应改为根据上下文透传 VECTOR / KEYWORD / HYBRID。
+     */
+    private String currentSearchMode() {
+        return "VECTOR";
+    }
+
     // ==================== 查询重写 ====================
 
     /**

src/main/java/com/wok/supportbot/service/LlmCallTraceService.java

@@ -0,0 +1,292 @@
+package com.wok.supportbot.service;
+
+import com.baomidou.mybatisplus.core.conditions.query.QueryWrapper;
+import com.wok.supportbot.dao.LlmCallTraceMapper;
+import com.wok.supportbot.entity.LlmCallTrace;
+import lombok.extern.slf4j.Slf4j;
+import org.springframework.beans.factory.annotation.Autowired;
+import org.springframework.jdbc.core.JdbcTemplate;
+import org.springframework.scheduling.annotation.Async;
+import org.springframework.scheduling.annotation.Scheduled;
+import org.springframework.stereotype.Service;
+
+import java.util.ArrayList;
+import java.util.LinkedHashMap;
+import java.util.List;
+import java.util.Map;
+import java.util.Set;
+
+/**
+ * LLM 调用追踪服务。
+ * <p>
+ * 异步写入每次 LLM 调用的完整现场，并提供分页查询、详情、聚合统计与自动清理。
+ * 列表查询不 select 大 TEXT 字段（system_prompt 等），避免分页 payload 膨胀。
+ */
+@Service
+@Slf4j
+public class LlmCallTraceService {
+
+    @Autowired
+    private LlmCallTraceMapper llmCallTraceMapper;
+
+    @Autowired
+    private JdbcTemplate jdbcTemplate;
+
+    @Autowired
+    private SystemConfigService systemConfigService;
+
+    /** 列表查询时 user_message 的摘要长度 */
+    private static final int SUMMARY_LENGTH = 120;
+
+    /** 列表查询需要排除的大 TEXT 字段 */
+    private static final Set<String> BIG_TEXT_FIELDS =
+            Set.of("system_prompt", "global_prompt", "role_prompt", "rag_context", "ai_response",
+                    "error_message", "tool_calls_json", "rag_hits_json", "history_messages_json");
+
+    /** 默认保留天数 */
+    private static final int DEFAULT_RETENTION_DAYS = 30;
+
+    /**
+     * 异步记录一次 LLM 调用追踪（不阻塞对话主流程）。
+     *
+     * @param trace 追踪实体
+     */
+    @Async("traceExecutor")
+    public void recordAsync(LlmCallTrace trace) {
+        try {
+            llmCallTraceMapper.insert(trace);
+            log.debug("记录 LLM 调用追踪: id={}, status={}, latencyMs={}", trace.getId(), trace.getStatus(), trace.getLatencyMs());
+        } catch (Exception e) {
+            log.error("记录 LLM 调用追踪失败: status={}, chatId={}, error={}", trace.getStatus(), trace.getConversationId(), e.getMessage());
+        }
+    }
+
+    /**
+     * 分页查询（列表不含大 TEXT 字段，user_message 截断为摘要）。
+     */
+    public Map<String, Object> pageQuery(int page, int size, Long roleId, String conversationId,
+                                         String intent, String startTime, String endTime, String keyword,
+                                         String errorType) {
+        if (page < 1) page = 1;
+        if (page > 10000) page = 10000;
+        if (size < 1 || size > 100) size = 20;
+
+        // 总数（只加 WHERE 条件）
+        Long total = llmCallTraceMapper.selectCount(
+                buildWhere(new QueryWrapper<>(), roleId, conversationId, intent, startTime, endTime, keyword, errorType));
+        if (total == null) total = 0L;
+
+        // 列表（排除大字段 + 排序 + 分页）
+        QueryWrapper<LlmCallTrace> listWrapper = buildWhere(
+                new QueryWrapper<>(), roleId, conversationId, intent, startTime, endTime, keyword, errorType);
+        listWrapper.select(LlmCallTrace.class, field -> !BIG_TEXT_FIELDS.contains(field.getColumn()));
+        listWrapper.orderByDesc("create_time").orderByDesc("id");
+        listWrapper.last("LIMIT " + size + " OFFSET " + ((page - 1L) * size));
+        List<LlmCallTrace> records = llmCallTraceMapper.selectList(listWrapper);
+
+        // user_message 截断为摘要，避免列表 payload 过大
+        for (LlmCallTrace r : records) {
+            if (r.getUserMessage() != null && r.getUserMessage().length() > SUMMARY_LENGTH) {
+                r.setUserMessage(r.getUserMessage().substring(0, SUMMARY_LENGTH) + "…");
+            }
+        }
+
+        Map<String, Object> result = new LinkedHashMap<>();
+        result.put("records", records);
+        result.put("total", total);
+        result.put("page", page);
+        result.put("size", size);
+        result.put("pages", (total + size - 1) / size);
+        return result;
+    }
+
+    /**
+     * 单条详情（含全部字段，system_prompt 全文）。
+     */
+    public LlmCallTrace getDetail(Long id) {
+        return llmCallTraceMapper.selectById(id);
+    }
+
+    /**
+     * 聚合统计：按角色（role）或按模型（model）分组。
+     */
+    public List<Map<String, Object>> stats(String groupBy) {
+        String sql;
+        if ("model".equalsIgnoreCase(groupBy)) {
+            sql = "SELECT model_name, provider, COUNT(*) AS count, "
+                    + "ROUND(AVG(latency_ms)) AS avg_latency_ms, "
+                    + "COUNT(*) FILTER (WHERE faq_hit = true) AS faq_hit_count "
+                    + "FROM llm_call_trace WHERE model_name IS NOT NULL "
+                    + "GROUP BY model_name, provider ORDER BY count DESC";
+        } else {
+            sql = "SELECT role_id::text AS role_id, role_name, COUNT(*) AS count, "
+                    + "ROUND(AVG(latency_ms)) AS avg_latency_ms, "
+                    + "COUNT(*) FILTER (WHERE faq_hit = true) AS faq_hit_count "
+                    + "FROM llm_call_trace GROUP BY role_id, role_name ORDER BY count DESC";
+        }
+        List<Map<String, Object>> rows = jdbcTemplate.queryForList(sql);
+        List<Map<String, Object>> result = new ArrayList<>(rows.size());
+        for (Map<String, Object> row : rows) {
+            result.add(snakeToCamel(row));
+        }
+        return result;
+    }
+
+    /**
+     * 时间维度趋势：按小时或天聚合调用量 / 平均耗时 / 错误率 / FAQ 命中率 / token 用量。
+     *
+     * @param groupBy   HOUR 或 DAY（其余值按 HOUR 处理）
+     * @param startTime 可选开始时间（含），格式 yyyy-MM-dd HH:mm:ss
+     * @param endTime   可选结束时间（含），格式 yyyy-MM-dd HH:mm:ss
+     * @return 按时间桶升序排列的统计列表
+     */
+    public List<Map<String, Object>> trend(String groupBy, String startTime, String endTime) {
+        boolean byDay = "DAY".equalsIgnoreCase(groupBy);
+        String bucket = byDay ? "day" : "hour";
+        String timeFormat = byDay ? "'YYYY-MM-DD'" : "'YYYY-MM-DD HH24'";
+
+        StringBuilder sql = new StringBuilder();
+        sql.append("SELECT to_char(date_trunc('").append(bucket).append("', create_time), ")
+                .append(timeFormat).append(") AS time_bucket, ")
+                .append("COUNT(*) AS call_count, ")
+                .append("ROUND(AVG(latency_ms)) AS avg_latency_ms, ")
+                .append("ROUND(100.0 * COUNT(*) FILTER (WHERE status = 'ERROR') / NULLIF(COUNT(*), 0), 2) AS error_rate, ")
+                .append("ROUND(100.0 * COUNT(*) FILTER (WHERE faq_hit = true) / NULLIF(COUNT(*), 0), 2) AS faq_hit_rate, ")
+                .append("COALESCE(SUM(total_tokens), 0) AS total_tokens ")
+                .append("FROM llm_call_trace WHERE 1=1 ");
+        List<Object> params = new ArrayList<>();
+        if (startTime != null && !startTime.isBlank()) {
+            sql.append("AND create_time >= ? ");
+            params.add(startTime.trim());
+        }
+        if (endTime != null && !endTime.isBlank()) {
+            sql.append("AND create_time <= ? ");
+            params.add(endTime.trim());
+        }
+        // PostgreSQL 不允许 GROUP BY 引用 SELECT 输出列别名，故此处重复 date_trunc 表达式；
+        // ORDER BY 则可引用别名 time_bucket
+        sql.append("GROUP BY date_trunc('").append(bucket).append("', create_time) ORDER BY time_bucket");
+
+        List<Map<String, Object>> rows = jdbcTemplate.queryForList(sql.toString(), params.toArray());
+        List<Map<String, Object>> result = new ArrayList<>(rows.size());
+        for (Map<String, Object> row : rows) {
+            result.add(snakeToCamel(row));
+        }
+        return result;
+    }
+
+    /**
+     * 分批删除 N 天前的记录，避免单次大表 DELETE 锁表。
+     *
+     * @param keepDays 保留天数
+     * @return 删除条数
+     */
+    public int cleanBefore(int keepDays) {
+        int batchSize = 5000;
+        int totalDeleted = 0;
+        while (true) {
+            int deleted = jdbcTemplate.update(
+                    "DELETE FROM llm_call_trace WHERE id IN ("
+                            + "SELECT id FROM llm_call_trace "
+                            + "WHERE create_time < NOW() - make_interval(days => ?) LIMIT ?)",
+                    keepDays, batchSize);
+            if (deleted <= 0) {
+                break;
+            }
+            totalDeleted += deleted;
+            if (deleted < batchSize) {
+                break;
+            }
+        }
+        return totalDeleted;
+    }
+
+    /**
+     * 每日凌晨 2 点自动清理过期追踪记录（保留天数从 system_config 读取）。
+     */
+    @Scheduled(cron = "0 0 2 * * ?")
+    public void scheduledClean() {
+        int keepDays = retentionDays();
+        try {
+            int deleted = cleanBefore(keepDays);
+            if (deleted > 0) {
+                log.info("LLM 调用追踪自动清理完成：删除 {} 条（保留 {} 天）", deleted, keepDays);
+            }
+        } catch (Exception e) {
+            log.error("LLM 调用追踪自动清理失败: {}", e.getMessage());
+        }
+    }
+
+    /**
+     * 读取保留天数配置，非法值回退默认 30 天。
+     */
+    private int retentionDays() {
+        String value = systemConfigService.getValueByKey("llm_trace_retention_days");
+        if (value != null && value.matches("\\d+")) {
+            int days = Integer.parseInt(value);
+            if (days >= 1) {
+                return days;
+            }
+        }
+        return DEFAULT_RETENTION_DAYS;
+    }
+
+    /**
+     * 组装 WHERE 条件（供 count 与 list 复用）。
+     */
+    private QueryWrapper<LlmCallTrace> buildWhere(QueryWrapper<LlmCallTrace> wrapper, Long roleId,
+            String conversationId, String intent, String startTime, String endTime, String keyword,
+            String errorType) {
+        if (roleId != null) {
+            wrapper.eq("role_id", roleId);
+        }
+        if (conversationId != null && !conversationId.isBlank()) {
+            wrapper.eq("conversation_id", conversationId.trim());
+        }
+        if (intent != null && !intent.isBlank()) {
+            wrapper.eq("intent", intent.trim().toUpperCase());
+        }
+        if (errorType != null && !errorType.isBlank()) {
+            wrapper.eq("error_type", errorType.trim().toUpperCase());
+        }
+        if (startTime != null && !startTime.isBlank()) {
+            wrapper.ge("create_time", startTime.trim());
+        }
+        if (endTime != null && !endTime.isBlank()) {
+            wrapper.le("create_time", endTime.trim());
+        }
+        if (keyword != null && !keyword.isBlank()) {
+            String raw = keyword.trim();
+            final String kw = raw.length() > 128 ? raw.substring(0, 128) : raw;
+            wrapper.and(w -> w.like("user_message", kw).or().like("ai_response", kw));
+        }
+        return wrapper;
+    }
+
+    /**
+     * snake_case → camelCase（供 JdbcTemplate 查询结果转驼峰）。
+     */
+    private Map<String, Object> snakeToCamel(Map<String, Object> row) {
+        Map<String, Object> out = new LinkedHashMap<>();
+        for (Map.Entry<String, Object> e : row.entrySet()) {
+            out.put(toCamel(e.getKey()), e.getValue());
+        }
+        return out;
+    }
+
+    private String toCamel(String snake) {
+        StringBuilder sb = new StringBuilder();
+        boolean upper = false;
+        for (char c : snake.toCharArray()) {
+            if (c == '_') {
+                upper = true;
+            } else if (upper) {
+                sb.append(Character.toUpperCase(c));
+                upper = false;
+            } else {
+                sb.append(c);
+            }
+        }
+        return sb.toString();
+    }
+}

src/main/java/com/wok/supportbot/service/RagHitLogService.java

@@ -28,7 +28,7 @@ public class RagHitLogService {
      * @param score          匹配得分
      * @param searchMode     检索模式
      */
-    @Async
+    @Async("traceExecutor")
     public void recordHit(String conversationId, String userQuery, Long documentId,
                           String documentTitle, String score, String searchMode) {
         try {
@@ -54,7 +54,7 @@ public class RagHitLogService {
      * @param userQuery      用户查询文本
      * @param searchMode     检索模式
      */
-    @Async
+    @Async("traceExecutor")
     public void recordMiss(String conversationId, String userQuery, String searchMode) {
         try {
             RagHitLog hitLog = RagHitLog.builder()

src/main/resources/init-database.sql

@@ -604,7 +604,161 @@ DROP INDEX IF EXISTS idx_faq_tenant_id;
 ALTER TABLE api_key ADD COLUMN IF NOT EXISTS role_ids JSONB DEFAULT '[]' NOT NULL;
 COMMENT ON COLUMN api_key.role_ids IS '绑定的客服角色 ID 列表（JSONB 数组，空数组=不限制，返回所有启用角色）';
 
+-- ============================================================
+-- 表 15: llm_call_trace — LLM 调用追踪表（提示词优化调试）
+-- ============================================================
+CREATE TABLE IF NOT EXISTS llm_call_trace (
+    id                    BIGINT       PRIMARY KEY,
+    conversation_id       VARCHAR(64),
+    role_id               BIGINT,
+    role_name             VARCHAR(100),
+    account_id            VARCHAR(64),
+    api_key_id            BIGINT,
+    intent                VARCHAR(16),
+    enable_rag            BOOLEAN,
+    system_prompt         TEXT,
+    global_prompt         TEXT,
+    role_prompt           TEXT,
+    user_message          TEXT,
+    ai_response           TEXT,
+    ai_response_truncated BOOLEAN,
+    tool_calls_json       TEXT,
+    history_messages_json TEXT,
+    history_turns         INTEGER,
+    rag_context           TEXT,
+    faq_hit               BOOLEAN,
+    faq_id                BIGINT,
+    faq_question          TEXT,
+    faq_match_type        VARCHAR(16),
+    faq_score             DOUBLE PRECISION,
+    search_mode           VARCHAR(20),
+    hit_count             INTEGER,
+    rag_hits_json         TEXT,
+    model_name            VARCHAR(128),
+    provider              VARCHAR(64),
+    temperature           DOUBLE PRECISION,
+    max_tokens            INTEGER,
+    prompt_tokens         INTEGER,
+    completion_tokens     INTEGER,
+    total_tokens          INTEGER,
+    error_type            VARCHAR(32),
+    error_message         TEXT,
+    latency_ms            INTEGER,
+    status                VARCHAR(16),
+    create_time           TIMESTAMP    NOT NULL DEFAULT CURRENT_TIMESTAMP
+);
+
+CREATE INDEX IF NOT EXISTS idx_llm_trace_created      ON llm_call_trace (create_time DESC, id DESC);
+CREATE INDEX IF NOT EXISTS idx_llm_trace_role_created ON llm_call_trace (role_id, create_time DESC);
+CREATE INDEX IF NOT EXISTS idx_llm_trace_conv_created ON llm_call_trace (conversation_id, create_time DESC);
+CREATE INDEX IF NOT EXISTS idx_llm_trace_error_type   ON llm_call_trace (error_type, create_time DESC);
+CREATE INDEX IF NOT EXISTS idx_llm_trace_api_key      ON llm_call_trace (api_key_id, create_time DESC);
+CREATE INDEX IF NOT EXISTS idx_llm_trace_search_mode  ON llm_call_trace (search_mode, create_time DESC);
+
+COMMENT ON TABLE  llm_call_trace                    IS 'LLM 调用追踪表（记录每次 LLM 调用的 system prompt/回复/模型参数/耗时，用于提示词优化调试，append-only）';
+COMMENT ON COLUMN llm_call_trace.id                 IS '主键（雪花算法生成）';
+COMMENT ON COLUMN llm_call_trace.conversation_id    IS '会话 ID';
+COMMENT ON COLUMN llm_call_trace.role_id            IS '客服角色 ID（可空）';
+COMMENT ON COLUMN llm_call_trace.role_name          IS '角色名称快照（角色改名后仍可追溯）';
+COMMENT ON COLUMN llm_call_trace.account_id         IS '账户 ID（可空，预留租户/个人信息删除权）';
+COMMENT ON COLUMN llm_call_trace.api_key_id         IS 'API Key ID（可空，预留租户隔离）';
+COMMENT ON COLUMN llm_call_trace.intent             IS '意图：CHAT / CHITCHAT / FAQ / RAG';
+COMMENT ON COLUMN llm_call_trace.enable_rag         IS '是否启用 RAG 增强';
+COMMENT ON COLUMN llm_call_trace.system_prompt      IS '最终注入 LLM 的完整 system prompt';
+COMMENT ON COLUMN llm_call_trace.global_prompt      IS '全局提示词快照（可空）';
+COMMENT ON COLUMN llm_call_trace.role_prompt        IS '角色提示词快照（可空）';
+COMMENT ON COLUMN llm_call_trace.user_message       IS '用户原始消息（脱敏后）';
+COMMENT ON COLUMN llm_call_trace.ai_response        IS 'AI 回复（截断，保留头部+尾部）';
+COMMENT ON COLUMN llm_call_trace.ai_response_truncated IS 'AI 回复是否被截断';
+COMMENT ON COLUMN llm_call_trace.tool_calls_json     IS 'MCP 工具调用事件 JSON 数组（可空）';
+COMMENT ON COLUMN llm_call_trace.history_messages_json IS '本次注入 LLM 的历史消息 JSON（可空）';
+COMMENT ON COLUMN llm_call_trace.history_turns       IS '历史消息条数（可空）';
+COMMENT ON COLUMN llm_call_trace.rag_context        IS 'RAG 资料块（可空）';
+COMMENT ON COLUMN llm_call_trace.faq_hit            IS '是否 FAQ 命中';
+COMMENT ON COLUMN llm_call_trace.faq_id             IS 'FAQ 命中 ID（可空）';
+COMMENT ON COLUMN llm_call_trace.faq_question       IS 'FAQ 标准问题快照（可空）';
+COMMENT ON COLUMN llm_call_trace.faq_match_type     IS 'FAQ 匹配类型：EXACT / KEYWORD / SEMANTIC（可空）';
+COMMENT ON COLUMN llm_call_trace.faq_score          IS 'FAQ 匹配分数（0.0~1.0，可空）';
+COMMENT ON COLUMN llm_call_trace.search_mode        IS '检索模式：VECTOR / KEYWORD / HYBRID（可空）';
+COMMENT ON COLUMN llm_call_trace.hit_count          IS '命中文档数（可空）';
+COMMENT ON COLUMN llm_call_trace.rag_hits_json      IS 'RAG 命中片段详情 JSON 数组（可空）';
+COMMENT ON COLUMN llm_call_trace.model_name         IS '模型名称';
+COMMENT ON COLUMN llm_call_trace.provider           IS '提供商';
+COMMENT ON COLUMN llm_call_trace.temperature        IS '温度参数';
+COMMENT ON COLUMN llm_call_trace.max_tokens         IS '最大 Token';
+COMMENT ON COLUMN llm_call_trace.prompt_tokens      IS '提示词 token 数（可空）';
+COMMENT ON COLUMN llm_call_trace.completion_tokens  IS '生成 token 数（可空）';
+COMMENT ON COLUMN llm_call_trace.total_tokens       IS '总 token 数（可空）';
+COMMENT ON COLUMN llm_call_trace.error_type         IS '错误类型：LLM_API / MCP / CIRCUIT_BREAK / VALIDATION / UNKNOWN（可空）';
+COMMENT ON COLUMN llm_call_trace.error_message      IS '错误原始消息（已脱敏，可空）';
+COMMENT ON COLUMN llm_call_trace.latency_ms         IS '调用耗时（毫秒）';
+COMMENT ON COLUMN llm_call_trace.status             IS '状态：COMPLETE / ERROR / CANCEL / FAQ / BYPASS';
+COMMENT ON COLUMN llm_call_trace.create_time        IS '创建时间';
+
+-- ============================================================
+-- 表 16: system_config — 系统配置表（key-value 模式）
+-- ============================================================
+CREATE TABLE IF NOT EXISTS system_config (
+    id              BIGINT          PRIMARY KEY,
+    config_key      VARCHAR(128)    NOT NULL UNIQUE,
+    config_value    TEXT            NOT NULL DEFAULT '',
+    description     VARCHAR(500),
+    create_time     TIMESTAMP       NOT NULL DEFAULT CURRENT_TIMESTAMP,
+    update_time     TIMESTAMP       NOT NULL DEFAULT CURRENT_TIMESTAMP,
+    is_delete       BOOLEAN         NOT NULL DEFAULT FALSE
+);
+
+-- 序列：供种子数据/手动 INSERT 生成 ID 兜底
+CREATE SEQUENCE IF NOT EXISTS system_config_id_seq START 1000 NO CYCLE;
+ALTER TABLE system_config ALTER COLUMN id SET DEFAULT nextval('system_config_id_seq');
+
+COMMENT ON TABLE  system_config              IS '系统配置表（key-value 模式，支持灵活扩展配置项）';
+COMMENT ON COLUMN system_config.id           IS '主键';
+COMMENT ON COLUMN system_config.config_key   IS '配置键（唯一）';
+COMMENT ON COLUMN system_config.config_value IS '配置值';
+COMMENT ON COLUMN system_config.description  IS '配置说明';
+COMMENT ON COLUMN system_config.create_time  IS '创建时间';
+COMMENT ON COLUMN system_config.update_time  IS '更新时间';
+COMMENT ON COLUMN system_config.is_delete    IS '逻辑删除';
+
+-- system_config 种子数据（与 DatabaseInitConfig.syncDefaultSystemConfigs 保持一致，幂等）
+INSERT INTO system_config (config_key, config_value, description)
+VALUES ('disclaimer',
+        '<p class="csk-disclaimer__title">特别声明 Claim of Confidential</p><p>✓ 本文件内容为 ④ 内部公开 请勿外传：禁止未授权的内部、第三方人员使用与访问</p><p>✓ 非专项必要的业务与项目负责人，收到此内容请立即删除</p><p>✓ 严禁在未经过主管领导审批，发送给无关业务团队与人员</p>',
+        'SDK 聊天窗口底部的保密声明')
+ON CONFLICT (config_key) DO NOTHING;
+
+INSERT INTO system_config (config_key, config_value, description)
+VALUES ('ai_system_prompt', '', 'AI 对话全局系统提示词（为空则不注入；修改后即时生效，无需重启）')
+ON CONFLICT (config_key) DO NOTHING;
+
+INSERT INTO system_config (config_key, config_value, description)
+VALUES ('suggestion_enabled', 'false', 'AI 推荐问题功能开关（true/false）')
+ON CONFLICT (config_key) DO NOTHING;
+
+INSERT INTO system_config (config_key, config_value, description)
+VALUES ('suggestion_prompt',
+        '【推荐问题生成规则】
+在回答正文结束后，请严格按以下格式生成 3 条用户可能继续追问的推荐问题：
+
+___SUGGESTIONS___
+["推荐问题1", "推荐问题2", "推荐问题3"]
+
+要求：
+1. 推荐问题需与当前回答内容相关，覆盖用户可能关心的不同方面
+2. 推荐问题是用户可直接点击发送的完整问题文本
+3. 必须是合法的 JSON 字符串数组
+4. ___SUGGESTIONS___ 是分隔标记，不要在回答正文中出现此标记',
+        'AI 推荐问题 Prompt 模板（仅在 suggestion_enabled=true 时生效）')
+ON CONFLICT (config_key) DO NOTHING;
+
+INSERT INTO system_config (config_key, config_value, description)
+VALUES ('llm_trace_retention_days', '30', 'LLM 调用追踪记录保留天数（自动清理，默认 30 天）')
+ON CONFLICT (config_key) DO NOTHING;
+
 -- ============================================================
 -- 完成！
--- 共创建 14 张表及对应索引。
+-- 注：本脚本与 DatabaseInitConfig 为双文件同步源；部分表（如 rag_hit_log、
+-- dashboard_snapshot、sys_user 等）历史上仅随 DatabaseInitConfig 演进，
+-- 属既有技术债，以 DatabaseInitConfig 为准。
 -- ============================================================

src/test/java/com/wok/supportbot/Phase1ComponentTests.java

@@ -1,123 +0,0 @@
-package com.wok.supportbot;
-
-import com.wok.supportbot.app.ChatContext;
-import com.wok.supportbot.rag.CategoryFilter;
-import com.wok.supportbot.rag.RagContext;
-import com.wok.supportbot.rag.RagPipeline;
-import jakarta.annotation.Resource;
-import org.junit.jupiter.api.Assertions;
-import org.junit.jupiter.api.DisplayName;
-import org.junit.jupiter.api.Test;
-import org.springframework.boot.test.context.SpringBootTest;
-
-import java.util.Collections;
-import java.util.List;
-
-/**
- * 阶段一公共组件验证测试。
- * 需要运行中的 PostgreSQL（RagPipeline 检索依赖向量库）。
- */
-@SpringBootTest
-class Phase1ComponentTests {
-
-    @Resource
-    private CategoryFilter categoryFilter;
-
-    @Resource
-    private RagPipeline ragPipeline;
-
-    // ==================== 1.1 CategoryFilter ====================
-
-    @Test @DisplayName("CF-01 parse(String) 正常逗号分隔")
-    void parseStringNormal() {
-        Assertions.assertEquals(List.of(1L, 2L, 3L), categoryFilter.parse("1,2,3"));
-    }
-
-    @Test @DisplayName("CF-02 parse(String) 含空格")
-    void parseStringWithSpaces() {
-        Assertions.assertEquals(List.of(1L, 2L, 3L), categoryFilter.parse(" 1 , 2 , 3 "));
-    }
-
-    @Test @DisplayName("CF-03 parse(String) null/空白")
-    void parseStringNullOrBlank() {
-        Assertions.assertEquals(Collections.emptyList(), categoryFilter.parse((String) null));
-        Assertions.assertEquals(Collections.emptyList(), categoryFilter.parse(""));
-        Assertions.assertEquals(Collections.emptyList(), categoryFilter.parse("   "));
-    }
-
-    @Test @DisplayName("CF-04/05 parse(Object) List")
-    void parseObjectList() {
-        // 数字元素
-        Assertions.assertEquals(List.of(1L, 2L, 3L), categoryFilter.parse((Object) List.of(1, 2, 3)));
-        // 字符串元素
-        Assertions.assertEquals(List.of(1L, 2L, 3L), categoryFilter.parse((Object) List.of("1", "2", "3")));
-    }
-
-    @Test @DisplayName("CF-06/07 parse(Object) 字符串 / null")
-    void parseObjectStringOrNull() {
-        Assertions.assertEquals(List.of(4L, 5L, 6L), categoryFilter.parse((Object) "4,5,6"));
-        Assertions.assertEquals(Collections.emptyList(), categoryFilter.parse((Object) null));
-    }
-
-    @Test @DisplayName("CF-08/09 normalize")
-    void normalize() {
-        Assertions.assertEquals(List.of("1", "2", "3"), categoryFilter.normalize(List.of(1L, 2L, 3L)));
-        Assertions.assertEquals(Collections.emptyList(), categoryFilter.normalize(null));
-        Assertions.assertEquals(Collections.emptyList(), categoryFilter.normalize(Collections.emptyList()));
-    }
-
-    @Test @DisplayName("CF-10/11 buildExpression")
-    void buildExpression() {
-        Assertions.assertNotNull(categoryFilter.buildExpression(List.of(1L, 2L)));
-        Assertions.assertNull(categoryFilter.buildExpression(Collections.emptyList()));
-    }
-
-    // ==================== 1.2 RagPipeline ====================
-
-    @Test @DisplayName("RP-08 未指定策略 → 原样查询")
-    void ragDefaultStrategy() {
-        ChatContext ctx = new ChatContext("退换货流程是什么", "test-rp-01", "CHAT",
-                null, null, null, null, true, false);
-        RagContext result = ragPipeline.retrieve(ctx);
-        Assertions.assertFalse(result.faqHit(), "不应触发 FAQ（除非该问题恰好命中 FAQ 库）");
-        Assertions.assertNotNull(result.documents(), "documents 不应为 null");
-        Assertions.assertNotNull(result.rewrittenQuery(), "rewrittenQuery 不应为 null");
-    }
-
-    @Test @DisplayName("RP-11/12 分类过滤")
-    void ragCategoryFilter() {
-        // 有分类过滤
-        ChatContext ctx1 = new ChatContext("平台使用说明", "test-rp-11", "CHAT",
-                null, null, List.of(1L), null, true, false);
-        RagContext r1 = ragPipeline.retrieve(ctx1);
-        Assertions.assertNotNull(r1.documents());
-        System.out.println(">>> 分类过滤(1) 命中数: " + r1.documents().size());
-
-        // 无分类过滤
-        ChatContext ctx2 = new ChatContext("平台使用说明", "test-rp-12", "CHAT",
-                null, null, null, null, true, false);
-        RagContext r2 = ragPipeline.retrieve(ctx2);
-        Assertions.assertNotNull(r2.documents());
-        System.out.println(">>> 无分类过滤 命中数: " + r2.documents().size());
-    }
-
-    @Test @DisplayName("RP-14 retrieveDocuments 跳过 FAQ")
-    void ragRetrieveDocuments() {
-        ChatContext ctx = new ChatContext("平台使用说明", "test-rp-14", "CHAT",
-                null, null, null, null, true, false);
-        var docs = ragPipeline.retrieveDocuments(ctx);
-        Assertions.assertNotNull(docs);
-        System.out.println(">>> retrieveDocuments 命中数: " + docs.size());
-    }
-
-    @Test @DisplayName("RP-04 REWRITE 策略")
-    void ragRewriteStrategy() {
-        ChatContext ctx = new ChatContext("订单退款流程", "test-rp-04", "CHAT",
-                null, null, null, "REWRITE", true, false);
-        RagContext result = ragPipeline.retrieve(ctx);
-        Assertions.assertNotNull(result.rewrittenQuery());
-        System.out.println(">>> REWRITE 重写前: 订单退款流程");
-        System.out.println(">>> REWRITE 重写后: " + result.rewrittenQuery());
-        Assertions.assertNotNull(result.documents());
-    }
-}

src/test/java/com/wok/supportbot/SupportBotApplicationTests.java

@@ -1,55 +0,0 @@
-package com.wok.supportbot;
-
-import com.wok.supportbot.app.AssistantApp;
-import com.wok.supportbot.app.ChatContext;
-import jakarta.annotation.Resource;
-import org.junit.jupiter.api.Assertions;
-import org.junit.jupiter.api.Test;
-import org.springframework.boot.test.context.SpringBootTest;
-
-import java.util.UUID;
-
-@SpringBootTest
-class SupportBotApplicationTests {
-    @Resource
-    private AssistantApp assistantApp;
-
-    @Test
-    void testChat() {
-        String chatId = UUID.randomUUID().toString();
-        // 第一轮：商品咨询
-        String message = "你好，我想买一台适合学生用的笔记本电脑，有推荐吗？";
-        String answer = assistantApp.chat(ChatContext.of(message, chatId));
-        Assertions.assertNotNull(answer);
-
-        // 第二轮：物流问题
-        message = "我上周买的那台电脑现在还没到，能查一下物流吗？";
-        answer = assistantApp.chat(ChatContext.of(message, chatId));
-        Assertions.assertNotNull(answer);
-
-        // 第三轮：售后问题
-        message = "电脑到了，但有点问题。你刚刚说的售后流程能再说一遍吗？";
-        answer = assistantApp.chat(ChatContext.of(message, chatId));
-        Assertions.assertNotNull(answer);
-    }
-
-    @Test
-    void doChatWithRag() {
-        String chatId = "1069b88d-eb85-47ac-bd2e-c393d118a5aa";
-        String message = "我之前询问了你什么问题？";
-        String answer = assistantApp.chat(new ChatContext(message, chatId, "CHAT", null, null,
-                null, null, true, false));
-        Assertions.assertNotNull(answer);
-    }
-
-    @Test
-    void doChatWithRagEnhance() {
-        String chatId = "1069b88d-eb85-47ac-bd2e-c393d118a5aa";
-        String message = "我之前询问了你什么问题？";
-        String answer = assistantApp.chat(new ChatContext(message, chatId, "CHAT", null, null,
-                null, null, true, false));
-        Assertions.assertNotNull(answer);
-    }
-
-
-}