fix: 修复重试计数、误报反馈、响应式阻塞等问题，版本号更新为v1.2.1

修复 AiReplyOrchestrator.retryOrFail 使用 .then() 导致 retryCount 失效；修复误报反馈端点 .subscribe() 过早触发覆盖 AI 回复；修复 PersonaResolver 在响应式上下文使用 .block() 阻塞线程；修复 penalize/approveOriginalComment 缺少乐观锁重试；修复误报反馈端点无法重试 FAIL 状态；修复 tag-NEUTRAL 缺少样式、handleTriggerAiReply 缺少加载保护；修复 filterKeyword 未做防抖、performCleanup 逻辑错误；优化 extractChoice 优先匹配违规类别避免误判
chore: 版本号更新为v1.2.0，更新日志同步
2026-07-01 10:52:28 +08:00 · 2026-06-25 08:48:38 +08:00 · 2026-06-25 08:33:33 +08:00 · 2026-06-25 07:56:16 +08:00 · 2026-06-24 16:34:02 +08:00 · 2026-06-23 19:59:25 +08:00
50 changed files with 3446 additions and 2921 deletions
@@ -6,7 +6,29 @@ on:
      - published

 jobs:
+  # Pre-cleanup: delete all existing assets from the release to avoid gh release upload failure
+  pre-release-cleanup:
+    runs-on: ubuntu-latest
+    steps:
+      - name: Delete Existing Release Assets
+        env:
+          GH_TOKEN: ${{ secrets.GITHUB_TOKEN }}
+        run: |
+          TAG_NAME="${{ github.event.release.tag_name }}"
+          # Capture asset list first to avoid pipefail issues
+          ASSETS=$(gh release view "$TAG_NAME" --json assets --jq '.assets[].name' 2>/dev/null || true)
+          if [ -n "$ASSETS" ]; then
+            echo "$ASSETS" | while read -r filename; do
+              echo "Deleting existing asset: $filename"
+              gh release delete-asset "$TAG_NAME" "$filename" --yes 2>/dev/null || true
+            done
+          else
+            echo "No existing assets to delete"
+          fi
+        shell: bash
+
  cd:
+    needs: pre-release-cleanup
    uses: halo-sigs/reusable-workflows/.github/workflows/plugin-cd.yaml@v4
    permissions:
      contents: write
@@ -63,6 +63,7 @@ lerna-debug.log*
 *.ctxt

 ### Package Files
+*.jar
 *.war
 *.nar
 *.ear
@@ -70,6 +71,12 @@ lerna-debug.log*
 *.tar.gz
 *.rar

+### UI build output
+ui/dist/
+ui/dist-ssr/
+ui/*.local
+ui/.eslintcache
+
 ### Local file
 application-local.yml
 application-local.yaml
@@ -1,13 +1,16 @@
 # AI回评 / Comment AI Autopilot

-基于 AI 的 Halo 博客评论自动回复插件，支持多 AI 角色、自审核、自动发布和对话式连续回复。
+基于 AI 的 Halo 博客评论自动回复插件，支持多 AI 角色、合规检测、自审核、自动发布和对话式连续回复。

 ## 功能特性

- **多 AI 角色** — 支持创建多个 AI 角色，每个角色有独立的昵称、人格提示词和 Gravatar 头像，可为不同文章指定不同角色
+- **多 AI 角色** — 支持创建多个 AI 角色，每个角色有独立的昵称、人格提示词、性别、语气风格和 Gravatar 头像，可为不同文章指定不同角色
+- **唤醒词** — 评论以唤醒词开头可唤醒指定角色回复，支持自定义唤醒词，可在未启用AI回评的页面使用唤醒词召唤AI
 - **自动回复** — 监听新评论，自动调用 AI 生成回复，支持多轮对话上下文
 - **多语言适配** — 根据评论语言自动用对应语言回复
- **情感分析** — 分析评论情感倾向（正面/中性/负面），根据情感调整回复语气
+- **情感分析** — 分析评论情感倾向（非常正面/正面/中性/负面/非常负面），根据情感调整回复语气
+- **前置过滤（合规检测）** — AI 回复前对评论进行合规性分类，自动拦截广告/辱骂攻击/敏感内容/乱码，违规评论停止生成 AI 回复以节省 Token，可选自动将违规评论设为待审核状态
+- **误报反馈** — 被误拦截的评论可进行误报反馈，支持"AI回复"和"仅通过"两种处理方式，"仅通过"后可随时补触发 AI 回复
 - **草稿模式** — AI 回复先存为草稿，管理员审核后再发布，支持批量操作
 - **失败重试** — AI 生成失败时自动重试，指数退避策略
 - **对话轮次限制** — 同一评论线程中限制 AI 最多回复轮次，防止无限对话
@@ -15,25 +18,30 @@
 - **文章/页面级开关** — 在文章编辑器中直接控制是否启用 AI 回复，文章默认开启，页面默认关闭
 - **评论者黑名单** — 支持按名称、邮箱和正则表达式屏蔽指定评论者，可从评论列表选择
 - **手动触发** — 在评论管理页面对历史评论手动触发 AI 回复
- **安全审核** — AI 生成的内容经过安全审核，不合规内容自动拒绝
- **Prompt 模板** — 支持自定义 Prompt 模板，提供多种模板变量（文章标题、发布日期、评论数等）
+- **两阶段安全审核** — 安全检查 + 质量评分（1-5 分映射到 0-100 分），不合规内容自动拒绝
+- **Prompt 模板** — 支持自定义 Prompt 模板，提供多种模板变量（文章标题、发布日期、评论数、对话历史等）
 - **Prompt 预设** — 内置友好型、专业型、幽默型、简洁型预设风格，可多选组合
- **仪表盘统计** — 显示回复数、情感分布、每日回复趋势等图表，支持时间范围切换
 - **插件健康检查** — 实时检测 AI Foundation 连接状态和模型可用性
- **日志筛选** — 按状态、情感筛选，关键词搜索
+- **日志筛选** — 按状态、情感筛选，关键词搜索，支持查看拦截原因和分类标签
 - **数据清理** — 自动清理超过指定天数的旧记录
- **AI Foundation 集成** — 必须安装 Halo AI Foundation 插件，使用其提供的 AI 模型能力
+- **AI Foundation 集成** — 通过 Halo 官方推荐的 `ExtensionGetter` 获取 AI 服务，需安装 AI Foundation 插件

 ## 前置要求

- Halo 2.23+
+- Halo 2.25+
 - AI Foundation 插件（必须）

 ## 安装

+### 应用商店安装
+
+进入 **插件** → **安装** → 应用市场搜索 **AI回评** → 安装，或前往 [Halo 应用商店](https://www.halo.run/store/apps/app-mo5tivjt) 一键安装。
+
+### 手动安装
+
 1. 前往 [Releases](https://github.com/sunny-335/plugin-comment-ai-autopilot/releases) 下载最新的 `.jar` 文件
 2. 登录 Halo 管理后台
-3. 进入 **插件** → **已安装** → 点击右上角 **安装** 按钮
+3. 进入 **插件** → **安装** → **本地上传**
 4. 选择下载的 `.jar` 文件上传
 5. 安装完成后启用插件

@@ -64,8 +72,8 @@ pnpm dev

 ## 文档

-完整文档请访问 [AI回评文档站](https://nxxy335.top/comment-ai-autopilot)
+完整文档及更新日志请访问 [AI回评文档站](https://nxxy335.top/comment-ai-autopilot)

 ## 许可证

-[GPL-3.0](./LICENSE) © 暖心向阳335
+[GPL-3.0](https://github.com/sunny-335/plugin-comment-ai-autopilot/blob/main/LICENSE) © 暖心向阳335
@@ -5,7 +5,7 @@ plugins {
 }

 group 'top.nxxy335.commentaiautopilot'
-version '1.0.0-beta.1'
+version project.property('version')

 repositories {
    mavenCentral()
@@ -41,13 +41,14 @@ export default defineConfig({
        text: "其他",
        items: [
          { text: "常见问题", link: "/guide/faq" },
+          { text: "更新日志", link: "/CHANGELOG" },
        ],
      },
    ],
    socialLinks: [
      {
        icon: "github",
-        link: "https://github.com/nxxy335/plugin-comment-ai-autopilot",
+        link: "https://github.com/sunny-335/plugin-comment-ai-autopilot",
      },
    ],
    search: {
@@ -0,0 +1,267 @@
+# 更新日志
+
+## v1.2.1
+
+> 2026-07-01
+
+### Bug 修复
+
+- **修复 `AiReplyOrchestrator.retryOrFail` 重试计数失效** — `.then()` 丢弃了更新后的记录导致 `retryCount` 始终为 0，AI 生成失败时陷入无限重试。改为 `.flatMap()` 传递更新后的记录
+- **修复误报反馈"AI 回复"被空字符串覆盖** — `.subscribe()` 在异步流程中过早触发，导致 AI 回复生成完成后被空字符串覆盖。改为在 `.doOnSuccess()` 中触发异步生成
+- **修复 `PersonaResolver` 在响应式上下文中使用 `.block()`** — 调用阻塞方法会阻塞 Reactor 线程。改为返回 `Mono<String>` 并使用 `Flux.concatMap().next()` 替代 for 循环
+- **修复 `penalizeComment`/`penalizeReply` 缺少乐观锁重试** — 并发更新 Comment/Reply 时可能静默失败。添加 `Retry.backoff(3, 100ms)` 重试
+- **修复 `approveOriginalComment` 缺少乐观锁重试** — 同上，添加 `Retry.backoff(3, 100ms)` 重试
+- **修复误报反馈端点无法重试 `FAIL` 状态记录** — 仅接受 `FILTERED` 和 `FALSE_POSITIVE` 状态，AI 生成失败的记录无法重试。现接受 `FAIL` 状态
+- **修复 `tag-NEUTRAL` 缺少 CSS 样式** — 中性情感标签无样式显示。补充样式定义
+- **修复 `handleTriggerAiReply` 缺少加载保护** — 触发 AI 回复按钮可被重复点击导致重复提交。添加 loading 状态
+- **修复 `filterKeyword` 输入未做防抖** — 每次按键都触发搜索，性能开销大。添加 300ms 防抖
+- **修复 `performCleanup` 逻辑错误** — 清理逻辑存在判断错误
+
+### 改进
+
+- **优化 `extractChoice` 分类匹配优先级** — 优先匹配违规类别（advertising/abuse/sensitive/meaningless），再匹配 `normal`，避免正常评论被误判为违规类别
+
+---
+
+## v1.2.0
+
+> 2026-06-25
+
+### 新增
+
+- **误报反馈功能** — 被拦截的评论可进行误报反馈，支持两种处理方式：
+  - **AI 回复**：标记为通过 + 触发 AI 生成回复
+  - **仅通过**：仅标记为通过，不生成回复
+- **误报通过状态** — 新增 `FALSE_POSITIVE` 状态，"仅通过"的记录显示为"误报通过"，不显示"通过/拒绝"按钮
+- **触发 AI 回复按钮** — "误报通过"状态的记录可随时点击"触发AI回复"按钮补生成 AI 回复
+- **上下文优先判断原则** — 前置过滤 AI 提示词重写，遵循五条核心原则：上下文优先、口语化宽容、恶意导向判定、宁放勿杀、闲聊不算无意义
+
+### Bug 修复
+
+- **修复误报反馈"AI 回复"被前置过滤再次拦截** — `processComment()` 始终调用 `preFilterService.check()`，用户已确认为误报的评论会被再次拦截。新增 `processFalsePositive()` 方法跳过前置过滤和去重检查
+- **修复误报反馈"AI 回复"被去重检查拦截** — `hasExistingReply()` 找到已有的 FILTERED→PENDING 记录导致 AI 回复无法生成。`processFalsePositive()` 复用已有记录，不经过去重检查
+- **修复误报反馈"AI 回复"导致全站崩溃** — `processComment()` 同步等待 AI 生成完成，HTTP 请求长时间不返回。改为 `.subscribe()` 异步执行，API 立即返回
+- **修复误报反馈"仅通过"后显示通过/拒绝按钮** — "仅通过"将记录设为 `status=PASS, published=false, reply=""`，导致显示"通过/拒绝"按钮且内容为空。改为 `status=FALSE_POSITIVE`
+- **修复 `extractChoice` 无匹配时返回原始文本** — AI 返回非预期文本时被误判为违规类别。改为返回空字符串触发安全拦截
+- **修复 `approveOriginalComment` 缺少乐观锁重试** — 并发更新 Comment/Reply 时可能静默失败。添加 `Retry.backoff(3, 100ms)` 重试
+
+### 改进
+
+- **消除 `checkBlockedCommenters` 重复代码** — `FilterService` 新增 `isCommenterBlocked(commentName)` 公共方法，`AiReplyOrchestrator` 改为调用它
+- **前端批量操作防重复提交** — 批量通过/拒绝/删除按钮添加 `batchLoading` 状态，操作期间禁用按钮
+
+---
+
+## v1.1.2
+
+> 2026-06-24
+
+### Bug 修复
+
+- **修复 AI 分类完全不可用** — `classifyWithChoice` 和 `classifyWithChat` 均使用了 `GenerateTextRequest.Builder.system()` 方法，而该方法在当前 AI Foundation 版本中不被支持或导致运行时错误，导致所有评论均被拦截并显示"AI分类服务不可用，安全拦截"。现改为将 system prompt 合并到 user prompt 中，与可用的 `chat()` 方法保持一致的调用方式
+- **修复 `classifyWithChoice` NPE** — `.map()` 返回 `null` 时触发 Reactor 内部 NullPointerException，改为 `.flatMap()` + `Mono.empty()` 正确触发 fallback
+
+### 改进
+
+- **分类调用诊断日志增强** — 在 `AiFoundationDelegate`、`AiFoundationClient`、`CommentPreFilterService` 中增加关键诊断日志（分类开始、fallback 触发、分类结果、异常详情），便于排查分类链路问题
+- **AI 分类空结果处理** — 当 AI 返回空字符串时单独拦截，区别于"服务不可用"场景
+
+---
+
+## v1.1.1
+
+> 2026-06-24
+
+### 改进
+
+- **"无意义"分类范围收窄** — 与文章主题无关的闲聊、灌水、打招呼不再被判为"无意义"，仅纯乱码和无意义字符堆砌（如随机符号、键盘乱敲）才归类为"无意义"
+- **AI 分类降级方案** — 当 `OutputSpec.choice` 结构化输出不被模型支持时，自动退回到普通 chat 调用并从响应文本中提取分类值（`classifyWithChat` fallback）
+
+---
+
+## v1.1.0
+
+> 2026-06-23
+
+### 新增
+
+- **评论前置过滤（合规检测）** — AI 回复前对评论进行合规性分类，识别广告/辱骂攻击/敏感内容/无意义内容，违规评论停止生成 AI 回复，节省 Token
+- **违规评论自动设为待审核** — 检测到违规评论时自动将原评论 `approved` 置为 `false`，进入待审核队列，前端不再展示该评论
+- **FILTERED 日志状态** — 被拦截的评论生成"已拦截"状态记录，日志页支持按"已拦截"状态筛选
+- **拦截原因分类标签** — 日志页显示拦截分类标签（广告/辱骂攻击/敏感内容/无意义）和详细拦截原因（含评论内容摘要）
+- **安全优先策略** — AI 分类服务不可用或异常时，默认拦截评论而非放行，防止违规内容漏网
+
+### 改进
+
+- **AI Foundation 隔离加载** — 将 AI Foundation API 引用隔离到 `AiFoundationDelegate` 类，`AiFoundationClient` 不再直接引用 AI Foundation 类，修复未安装 AI Foundation 时插件无法启动的问题（`NoClassDefFoundError`）
+- **评论内容 HTML 剥离** — 前置过滤检测前自动剥离评论 HTML 标签，提升 AI 分类准确性
+- **对话场景精准处罚** — AI 对话场景下违规内容来自 Reply 时，仅取消通过该 Reply 而非父级 Comment，避免误伤
+- **升级配置自动迁移** — 从 v1.0.x 升级时自动将 `preFilterEnabled` 从 `false` 迁移为 `true`（新默认值）
+
+### Bug 修复
+
+- **修复未安装 AI Foundation 时插件无法启动** — `BeanDefinitionStoreException: Failed to parse AiFoundationClient`，将 AI Foundation API 引用隔离到委托类
+- **修复前置过滤默认关闭** — `preFilterEnabled` 默认值从 `false` 改为 `true`，新安装和升级用户均默认启用
+- **修复 `penalize()` 遗漏 `approved=null`** — Halo 评论创建时 `approved` 可能为 `null`，原代码仅处理 `approved=true` 的情况
+- **修复 `classify()` 失败时放行违规评论** — `defaultIfEmpty` 和 `onErrorResume` 改为拦截而非放行
+- **修复 Windows 构建失败** — Gradle Worker Daemon 执行 pnpm 退出码 268435659，改用系统 pnpm Exec 任务并禁用 Daemon
+
+---
+
+## v1.0.4
+
+> 2026-06-19
+
+### 改进
+
+- **对话弹窗头像显示** — 对话弹窗中每条消息显示 Gravatar 头像，基于评论者或 AI 角色的邮箱自动匹配
+- **对话引用摘要** — 对话弹窗中回复消息显示引用摘要框，标明引用了谁的什么内容，支持截断显示
+- **UI 全面重构** — LogsView 和 SettingsView 改用纯 Scoped CSS，移除所有 Tailwind 类和自定义 CSS 依赖，避免 Halo 主题冲突
+- **标签去 Emoji 化** — 状态、情感标签改用纯色背景标签，去除所有 Emoji
+- **移动端适配优化** — 全面优化移动端响应式布局，解决排版错位问题
+- **AI角色设置完善** — 支持 CRUD、Gravatar 头像预览、性别/唤醒词/默认角色配置
+- **配置导入导出** — 支持将插件配置（ConfigMap + AI角色）导出为 JSON 文件，方便备份和迁移
+- **评论者黑名单弹窗选择** — 设置页面可从已有评论列表中选择评论者添加到黑名单
+
+### Bug 修复
+
+- **修复对话弹窗引用溯源** — 后端 `getConversation` 重写，构建 Reply 映射字典正确溯源引用关系
+- **修复 ConversationMessage 数据结构** — 新增 `quoteOwner`/`quoteContent` 字段支持引用摘要展示
+- **修复 AI 角色邮箱提取** — 后端新增 `extractOwnerEmail` 方法，正确从 CommentOwner 提取邮箱用于头像生成
+
+---
+
+## v1.0.3
+
+### 改进
+
+- **SettingsView 完整功能版** — 5个设置面板（基本设置、AI角色、模型设置、Prompt、数据清理）全部实现
+- **AI角色管理** — 支持 CRUD、Gravatar 头像、性别/唤醒词/默认角色配置
+- **数据清理** — 自动清理开关、保留天数滑块、手动清理
+- **导入导出** — JSON 配置导入导出
+- **评论者黑名单弹窗选择** — 从已有评论列表中选择评论者
+
+---
+
+## v1.0.2
+
+### 改进
+
+- **LogsView & SettingsView 样式重构** — 移除所有 Tailwind 类，改用 `<style scoped>` 原生 CSS
+- **标签配色、气泡样式、引用框** — 全部使用纯 CSS 实现，避免 Halo 主题冲突
+
+---
+
+## v1.0.1
+
+### 改进
+
+- **版本号升级** — 强制刷新 Halo 前端缓存
+- **历史数据兼容** — LogsView 增加历史 Markdown 引用文本清理正则，防止旧版测试数据套娃显示
+
+---
+
+## v1.0.0
+
+> 2026-06-18
+
+### 新功能
+
+- **唤醒词**：评论以唤醒词开头可唤醒指定角色回复，支持自定义唤醒词，可在未启用AI回评的页面使用唤醒词召唤AI，二级评论同样支持
+- **性别配置**：AI角色支持性别设置（男/女），AI回复时会保持对应性别身份
+- **语气风格**：支持中性语气复选框，勾选后使用中性语气，取消勾选则跟随性别语气（女性温柔细腻/男性沉稳理性）
+- **身份提示词强化**：角色身份信息前置到Prompt最开头（【核心身份】），安全规范中增加身份约束，确保AI始终保持角色身份
+
+### 改进
+
+- **优化情感分析系统**：从 3 级分类（正面/中性/负面）升级为 5 级分类（非常正面/正面/中性/负面/非常负面），情感判断更精细
+- **优化日志页面 UI**：批量操作按钮重写样式，确保底色和白色文字清晰可见；搜索框添加搜索图标；重置按钮添加图标和底色
+- **优化评分显示**：评分数字与等级标签之间添加间距，等级标签增加底色背景（优秀/良好/一般/较差）
+- **优化状态标签**：通过状态、发布状态、情感标签统一使用带底色的标签样式
+- **支持页面链接显示**：日志中新增独立页面（SinglePage）链接显示，之前仅支持文章链接
+- **移动端适配**：仪表盘、配置、日志页面全面适配移动端
+- **ObjectMapper 统一注入**：FilterService 和 PromptBuilder 中的 `new ObjectMapper()` 改为 Spring 构造函数注入
+- **服务端过滤优化**：日志列表查询改用 `Queries.equal()` 服务端过滤 status/sentiment，减少内存过滤开销
+- **新增索引**：为 AiCommentReply 扩展添加 `spec.sentiment`、`spec.published`、`spec.postKind` 索引
+- **新增 postKind 字段**：区分关联内容类型（Post/SinglePage），支持页面评论的链接生成
+- **PromptBuilder 情感提示**：适配 5 级情感分类，新增 VERY_POSITIVE 和 VERY_NEGATIVE 的语气提示
+
+### Bug 修复
+
+- **修复 ObjectMapper Bean 不存在**：Halo 插件上下文中没有自动注册 ObjectMapper Bean，创建 ObjectMapperConfiguration 手动注册
+- **修复 AI 回复仍说没有性别**：将身份信息前置到 Prompt 最开头，安全规范中删除"作为AI助手"措辞，新增身份约束
+- **修复唤醒词无法唤醒**：评论内容提取时对 raw 也做 HTML strip（Jsoup.clean），所有内容做 trim()，wakeWord 也做 trim()
+- **修复二级评论唤醒词检查位置错误**：唤醒词检查提前到 isReplyToAi 判断之前
+- **修复 SinglePage 内容获取 404**：PostContentService 不能用于 SinglePage，改用 SinglePage.getStatus().getExcerpt()
+- **修复 Post/SinglePage 404 容错**：fetch 添加 onErrorResume 降级为空上下文继续处理
+- **修复 Sort 参数 null 警告**：listAll 调用改为 Sort.unsorted()
+
+---
+
+## v1.0.0-beta.2
+
+> 2026-06-17
+
+### 改进
+
+- **改用 ExtensionGetter 集成 AI Foundation**：通过 Halo 官方推荐的 `ExtensionGetter.getEnabledExtension(AiModelService.class)` 获取 AI 服务，替换原先的跨 ClassLoader 反射调用方式（[Issue #1](https://github.com/sunny-335/plugin-comment-ai-autopilot/issues/1)）
+- **声明插件依赖**：在 `plugin.yaml` 中声明可选插件依赖 `ai-foundation?: "*"`，建立正确的插件依赖关系，插件在未安装 AI Foundation 时仍可正常加载
+- **应用市场推荐**：新增 `store.halo.run/recommended-apps` 注解，安装本插件后可在应用市场推荐安装 AI Foundation 插件
+- **使用结构化输出**：情感分析和内容审核改用 AI Foundation 的 `OutputSpec.choice` 结构化输出，替换原先的字符串匹配解析，分类更可靠
+- **使用 GenerateTextRequest**：AI 调用改用 `GenerateTextRequest` 并设置 `maxRetries=2`，由 SDK 自动重试瞬时错误
+- **多轮对话上下文**：AI 对话续接时自动获取之前的回复历史并注入到 Prompt 中，AI 能更好地理解对话上下文
+- **优化 AI 自审核评分机制**：审核改为两阶段评估（安全检查 + 质量评分 1-5 分），评分映射到 0-100 分（0/30/50/70/85/100），替代原先的二值评分（0/100），评分更有区分度
+- **精简仪表盘**：移除情感分布、近7日回复趋势、平均审核评分三个卡片，快捷操作精简为回复日志、插件设置、刷新数据
+- **重做设置页面**：基本设置、AI角色设置、模型设置、Prompt设置、数据清理各为独立页面，通过标签栏切换
+- **优化设置页面布局**：按钮统一排版并添加图标，侧边栏保存卡片高亮显示，新增"未保存"状态指示器
+- **优化日志页面**：评分增加等级标签（优秀/良好/一般/较差），筛选下拉框修复文本与箭头重叠
+- **优化 AI Foundation 状态提示宽度**：状态提示与内容区宽度一致
+- **优化插件文档**：修复版本要求（2.23→2.25）、变量名（`{{conversation}}`→`{{conversation_history}}`）、GitHub 链接（`nxxy335`→`sunny-335`）、Cravatar 链接（`cravatar.cn`→`cn.cravatar.com`）等错误，补充缺失的配置项文档
+
+### Bug 修复
+
+- **修复自动发布后日志显示未发布**：`generateAndPublish` 中 `Mono<Void>` 的 empty 信号触发 `switchIfEmpty` 导致 `publishReply` 被调用两次，第二次覆盖 `published=false`
+- **修复 RateLimitService 线程泄漏**：清理线程未在插件停止时关闭，实现 `DisposableBean` 正确释放资源
+- **修复 ReviewService 提示词不匹配**：审核提示词要求"重新生成"但代码未使用重新生成的内容，移除误导性指令
+- **修复设置页面按钮图标文字对齐**：通过 `:deep(.btn-content)` 设置 inline-flex 布局，图标和文字并排显示
+- **修复设置页面标签栏无法点击切换**：替换不工作的 VTabbar 组件为自定义按钮实现
+- **修复设置页面配置区域布局错误**：将标签栏移出 grid 容器，避免挤占配置区域宽度
+- **修复日志页面筛选下拉框文本与箭头重叠**：将 `px-3` 改为 `pl-3 pr-8` 为下拉箭头预留空间
+
+---
+
+## v1.0.0-beta.1
+
+> 2026-06-05
+
+### 新功能
+
+- **自动回复**：监听新评论，自动调用 AI 生成回复
+- **多语言适配**：根据评论语言自动用对应语言回复
+- **情感分析**：分析评论情感倾向，根据情感调整回复语气
+- **草稿模式**：关闭"自动发布"后，AI 回复将保存为草稿，需站长审核后才发布
+- **多 AI 角色支持**：支持配置多个 AI 虚拟角色，每个角色有独立的提示词、头像和模型
+- **提示词预设**：内置多种回复风格预设（专业型、幽默型、简洁型等），可自由组合
+- **文章/页面级开关**：在文章编辑器中直接控制是否启用 AI 回复
+- **评论者黑名单**：屏蔽指定评论者，不触发 AI 回复
+- **安全审核**：AI 生成的内容经过安全审核，不合规内容自动拒绝
+- **Prompt 模板**：支持自定义 Prompt 模板，提供多种模板变量
+- **仪表盘统计**：显示回复数等统计信息
+- **日志筛选搜索**：按状态、情感筛选，关键词搜索
+- **数据清理**：自动清理超过指定天数的旧记录
+- **AI Foundation 集成**：使用 AI Foundation 插件提供的 AI 模型能力
+- **对话历史查看**：支持查看 AI 回复的完整对话上下文
+
+### Bug 修复
+
+- 修复草稿模式下审批失败（"AI回复已存在，无法重复发布"）的问题
+- 修复批量审批时同样的去重检查冲突问题
+- 修复 AI Foundation 不可用的问题（`PluginManager` 无法通过 Spring 依赖注入获取）
+- 修复 `DefaultSpringPlugin` 包级私有类反射访问权限问题
+- 修复 CI 构建失败（`gradlew` 缺少执行权限）
+
+### 改进
+
+- 审批逻辑优化：先查找已有 Reply 扩展再决定创建或更新
+- 移除不必要的 `AiFoundationConfiguration` 配置类
+- 前端 UI 优化：移除编辑功能、简化角色排序逻辑、清理无用代码
@@ -21,6 +21,32 @@

 插件启动时间之前的评论不会触发自动回复，避免安装插件后对大量历史评论批量回复。

+## 唤醒词机制
+
+唤醒词允许用户在评论中通过特定词语唤醒指定AI角色回复，即使该页面未启用AI回评。
+
+### 工作方式
+
+1. 用户发表以唤醒词开头的评论（如"小回小回你好"）
+2. 插件检测到唤醒词匹配，自动唤醒对应角色
+3. 唤醒词后的内容作为实际评论内容传递给AI
+4. AI生成回复时自动获取上下文（文章内容、对话历史等）
+
+### 唤醒词特点
+
+- **跨页面唤醒**：在未启用AI回评的页面也能使用唤醒词召唤AI
+- **二级评论支持**：回复中同样可以使用唤醒词
+- **独立唤醒**：每个角色有独立的唤醒词，可以唤醒不同角色
+- **绕过限制**：唤醒词触发时绕过页面级启用检查和"必须是回复给AI"的检查，但仍检查黑名单
+
+### 配置唤醒词
+
+在 **AI回评** → **插件设置** → **AI角色设置** 中，为每个角色配置唤醒词。唤醒词留空则不启用该角色的唤醒功能。
+
+::: tip
+唤醒词建议设置为容易记忆且不易与正常评论混淆的词语。
+:::
+
 ## 对话式回复

 当评论者回复AI的评论时，插件会自动提取对话上下文（最近5条回复），让AI的回复更连贯自然。
@@ -31,6 +57,8 @@
 |--------|------|--------|
 | 自动回复 | 是否启用自动回复功能 | 开启 |
 | 自动发布 | AI回复是否自动发布，关闭则存为草稿 | 开启 |
+| 最大对话轮次 | 同一评论线程中AI最多自动回复的轮次 | 8 |
+| 速率限制 | 每分钟最大AI回复数量 | 10 |
 | 最大重试次数 | AI生成失败时的最大重试次数 | 3 |

 ## 重试机制
@@ -22,6 +22,18 @@
 - 草稿记录显示 **审核通过** 和 **拒绝** 按钮
 - 已发布的记录显示正常状态
 - 被拒绝的记录显示 REJECTED 标签
+- 失败的记录显示 FAIL 标签，并显示重试次数
+- 每条记录可点击 **查看对话** 查看完整对话上下文
+
+## 对话上下文查看
+
+点击日志记录的 **查看对话** 按钮，弹出对话上下文窗口：
+
+- 以气泡形式展示完整对话（评论 + 所有回复）
+- AI 回复和用户回复以不同颜色气泡区分
+- 每条消息显示发送者头像（通过 Gravatar 服务生成）
+- 回复消息显示引用摘要框，标明该回复引用了哪条消息
+- 支持移动端响应式布局

 ## 批量操作

@@ -12,7 +12,7 @@

 1. 确认在插件设置中填写了AI角色邮箱
 2. 邮箱需要在 [Gravatar](https://gravatar.com) 上注册并设置头像
-3. 插件使用 [Cravatar](https://cravatar.cn) 作为Gravatar镜像服务
+3. 插件使用 [Cravatar](https://cn.cravatar.com) 作为Gravatar镜像服务

 ## 评论没有触发AI回复？

@@ -23,6 +23,7 @@
 3. **已有AI回复记录** — 同一评论不会重复触发
 4. **历史评论** — 插件启动前的评论不会自动触发，可使用手动触发
 5. **AI生成失败** — 检查AI模型配置和日志
+6. **被前置过滤拦截** — 若启用"前置过滤"，违规评论会被拦截，可在日志页通过"已拦截"状态筛选查看

 ## 如何对历史评论触发AI回复？

@@ -66,3 +67,57 @@
 ## 黑名单支持邮箱吗？

 支持。黑名单同时匹配评论者的显示名称和邮箱地址，不区分大小写。你也可以在设置页面点击"添加评论者"按钮从评论列表中选择。
+
+## 对话窗口中的头像是怎么来的？
+
+对话窗口中每条消息的头像通过 [Gravatar](https://gravatar.com) 服务生成（使用 [Cravatar](https://cn.cravatar.com) 镜像）。头像基于评论者或 AI 角色的邮箱自动匹配。如果未设置邮箱，则显示默认图标。
+
+## 对话窗口中的引用框是什么？
+
+当一条回复是针对另一条回复的（即层级回复），对话窗口会在该消息气泡内显示一个引用摘要框，标明该回复引用了谁的什么内容。引用内容会截断显示（最多35个字符），方便快速了解对话脉络。
+
+## 如何备份和迁移插件配置？
+
+在插件设置页面顶部点击 **导出** 按钮，将当前配置导出为 JSON 文件。在目标实例中点击 **导入** 按钮选择该文件即可恢复配置。导入会覆盖当前配置，请谨慎操作。
+
+## AI Foundation 显示"部分功能不可用"怎么办？
+
+这通常表示 AI Foundation 插件未正确配置模型。请检查：
+
+1. AI Foundation 插件已安装并启用
+2. 在 AI Foundation 中配置了至少一个 AI 模型
+3. 如果回评插件未指定模型名称，将使用 AI Foundation 的默认模型
+
+## 前置过滤会误伤正常评论吗？
+
+前置过滤默认启用。AI 会对评论进行分类判断，若 AI 服务不可用或分类失败，为安全起见会拦截评论而非放行。如果你发现正常评论被误拦截，可以在设置中关闭"启用前置过滤"开关。被拦截的评论会在日志页生成一条"已拦截"状态的记录，可查看具体分类标签和拦截原因。
+
+## 所有评论都显示"AI分类服务不可用，安全拦截"怎么办？
+
+这表示 AI 分类调用链路存在问题，可能的原因：
+
+1. **AI Foundation 插件未安装或未启用** — 请确保 AI Foundation 插件已正确安装并启用
+2. **AI Foundation 中未配置模型** — 请在 AI Foundation 中配置至少一个 AI 模型
+3. **模型名称配置错误** — 检查插件设置中的模型名称是否与 AI Foundation 中的 AiModel 资源名称一致，留空则使用默认模型
+4. **AI Foundation 版本过旧** — 请确保使用最新版本的 AI Foundation 插件
+
+::: tip 排查步骤
+1. 检查插件设置页面顶部的 AI Foundation 连接状态
+2. 查看插件日志中 `[Delegate]` 和 `[PreFilter]` 前缀的诊断信息
+3. 确认 AI 回复功能（非前置过滤）是否正常工作 — 如果 AI 回复也无法生成，说明是 AI Foundation 连接问题
+:::
+
+## 被前置过滤拦截的评论会怎样？
+
+1. **停止生成 AI 回复** — 不会消耗后续 Token
+2. **创建拦截记录** — 在日志页显示为"已拦截"状态，标注分类标签（如"辱骂攻击"）和详细原因（含评论内容摘要）
+3. **自动设为待审核** — 原评论的 `approved` 会被置为 `false`，前端不再展示该评论，需人工判断后审核通过
+
+## 被误拦截的评论怎么处理？
+
+在日志页的"已拦截"记录右侧，点击 **误报反馈** 按钮，可选择：
+
+- **AI 回复** — 标记为误报 + 自动通过评论 + 触发 AI 生成回复
+- **仅通过** — 仅标记为误报 + 自动通过评论，不生成 AI 回复
+
+选择"仅通过"后，记录状态变为"误报通过"，可随时点击 **触发AI回复** 按钮补生成 AI 回复。
@@ -32,9 +32,17 @@

 1. 进入插件设置页面
 2. 在 **基本设置** 中找到 **评论者黑名单**
-3. 输入评论者的显示名称或邮箱，多个用逗号分隔
+3. 输入评论者的显示名称、邮箱或正则表达式，多个用逗号分隔
 4. 保存设置

+### 支持的格式
+
+| 格式 | 示例 | 说明 |
+|------|------|------|
+| 名称 | `张三` | 匹配评论者的显示名称 |
+| 邮箱 | `spam@example.com` | 匹配评论者邮箱（不区分大小写） |
+| 正则表达式 | `regex:^spam.*` | 以 `regex:` 开头，按正则匹配 |
+
 ### 从评论列表选择

 1. 在黑名单输入框旁点击 **添加评论者** 按钮
@@ -45,7 +53,7 @@
 ### 示例

 ```
-张三,spam@example.com,李四
+张三, spam@example.com, 李四, regex:^spam.*
 ```

-黑名单中的评论者发布评论时，插件会同时匹配显示名称和邮箱地址（不区分大小写），匹配成功则跳过AI回复。
+黑名单中的评论者发布评论时，插件会同时匹配显示名称和邮箱地址（不区分大小写），正则表达式则按模式匹配，匹配成功则跳过AI回复。
@@ -2,24 +2,28 @@

 ## 前置要求

- Halo 2.23+
- AI Foundation 插件（必须） — 本插件依赖 AI Foundation 提供的AI模型能力，请先安装并配置 AI Foundation
+- Halo 2.25+
+- AI Foundation 插件（必须） — 本插件通过 `ExtensionGetter` 调用 AI Foundation 提供的 `AiModelService` 扩展点，请先安装并配置 AI Foundation

 ## 安装

-### 方式一：从 Release 下载
+### 方式一：应用商店安装

-1. 前往 [GitHub Releases](https://github.com/nxxy335/plugin-comment-ai-autopilot/releases) 下载最新的 `.jar` 文件
+进入 **插件** → **安装** → 应用市场搜索 **AI回评** → 安装，或前往 [Halo 应用商店](https://www.halo.run/store/apps/app-mo5tivjt) 一键安装。
+
+### 方式二：从 Release 下载
+
+1. 前往 [GitHub Releases](https://github.com/sunny-335/plugin-comment-ai-autopilot/releases) 下载最新的 `.jar` 文件
 2. 登录 Halo 管理后台
-3. 进入 **插件** → **已安装** → 点击右上角 **安装** 按钮
+3. 进入 **插件** → **安装** → **本地上传**
 4. 选择下载的 `.jar` 文件上传
 5. 安装完成后启用插件

-### 方式二：从源码构建
+### 方式三：从源码构建

 ```bash
 # 克隆仓库
-git clone https://github.com/nxxy335/plugin-comment-ai-autopilot.git
+git clone https://github.com/sunny-335/plugin-comment-ai-autopilot.git
 cd plugin-comment-ai-autopilot

 # 构建
@@ -5,45 +5,55 @@ AI回评（Comment AI Autopilot）是一个 Halo 博客系统的插件，能够
 ## 核心功能

 - **自动回复** — 监听新评论，自动调用AI生成回复，支持多轮对话上下文
+- **多 AI 角色** — 支持创建多个 AI 角色，每个角色有独立的昵称、人格提示词、性别、语气风格和 Gravatar 头像，可为不同文章指定不同角色
+- **唤醒词** — 评论以唤醒词开头可唤醒指定角色回复，支持自定义唤醒词，可在未启用AI回评的页面使用唤醒词召唤AI
+- **性别与语气** — AI角色支持性别配置（男/女）和中性语气风格，AI回复时保持对应性别身份
 - **多语言适配** — 根据评论语言自动用对应语言回复
- **情感分析** — 分析评论情感倾向（正面/中性/负面），根据情感调整回复语气
+- **情感分析** — 分析评论情感倾向（非常正面/正面/中性/负面/非常负面），根据情感调整回复语气
 - **草稿模式** — AI回复先存为草稿，管理员审核后再发布
 - **失败重试** — AI生成失败时自动重试，指数退避策略
 - **批量操作** — 草稿模式下支持批量通过/拒绝/删除
 - **文章/页面级开关** — 在文章编辑器中直接控制是否启用AI回复，文章默认开启，页面默认关闭
- **评论者黑名单** — 屏蔽指定评论者，不触发AI回复
+- **评论者黑名单** — 屏蔽指定评论者，不触发AI回复，支持名称、邮箱和正则表达式
+- **前置过滤（合规检测）** — AI回复前对评论进行合规性分类，自动拦截广告/辱骂/敏感/乱码内容，节省Token；可选将违规评论设为待审核状态
+- **误报反馈** — 被误拦截的评论可进行误报反馈，支持"AI回复"和"仅通过"两种处理方式，"仅通过"后可随时补触发 AI 回复
 - **手动触发** — 在评论管理页面对历史评论手动触发AI回复
- **AI角色** — 自定义AI回复者的昵称、人格提示词和Gravatar头像
- **安全审核** — AI生成的内容经过安全审核，不合规内容自动拒绝
- **仪表盘统计** — 显示回复数、情感分布、每日回复趋势等图表
+- **安全审核** — AI生成的内容经过两阶段安全审核（安全检查 + 质量评分），不合规内容自动拒绝
+- **Prompt 预设** — 内置友好型、专业型、幽默型、简洁型预设风格，可多选组合
+- **对话轮次限制** — 同一评论线程中限制 AI 最多回复轮次，防止无限对话
+- **速率限制** — 每分钟最大 AI 回复数量，防止批量评论消耗过多额度
 - **日志筛选搜索** — 按状态、情感筛选，关键词搜索
+- **对话上下文查看** — 在日志页面查看完整对话上下文，支持引用摘要展示和 Gravatar 头像显示
 - **数据清理** — 自动清理超过指定天数的旧记录
- **AI Foundation 集成** — 必须安装 Halo AI Foundation 插件，使用其提供的AI模型能力
+- **配置导入导出** — 支持将插件配置导出为 JSON 文件，方便备份和迁移
+- **AI Foundation 集成** — 通过 Halo 官方推荐的 `ExtensionGetter` 获取 AI 服务，需安装 AI Foundation 插件

 ## 工作流程

 ```
-新评论 → 过滤检查 → 情感分析 → 构建Prompt → AI生成 → 安全审核 → 发布/草稿
+新评论 → 唤醒词检查 → 过滤检查 → 前置过滤（合规检测） → 情感分析 → 构建Prompt → AI生成 → 安全审核 → 发布/草稿
                                                          ↓ (失败)
                                                        重试 → ... → 最终失败
 ```

 1. **新评论到达** — Reconciler 监听到新评论创建事件
-2. **过滤检查** — 检查文章/页面是否启用AI回复、评论者是否在黑名单中
-3. **情感分析** — 调用AI分析评论情感倾向
-4. **构建Prompt** — 结合AI角色人格、情感提示、文章内容、评论上下文构建Prompt
-5. **AI生成** — 调用AI模型生成回复内容
-6. **安全审核** — 对生成内容进行安全审核
-7. **发布/草稿** — 根据设置自动发布或存为草稿等待审核
-8. **重试** — 如果AI生成失败，系统会自动重试（最多 maxRetryCount 次），每次重试间隔递增
+2. **唤醒词检查** — 检查评论是否以某个角色的唤醒词开头，匹配则唤醒对应角色
+3. **过滤检查** — 检查文章/页面是否启用AI回复、评论者是否在黑名单中（唤醒词触发时绕过页面级启用检查）
+4. **前置过滤（合规检测）** — 若启用，AI 对评论内容进行合规性分类（正常/广告/辱骂攻击/敏感内容/无意义）。违规评论将停止后续流程，可选自动设为待审核状态
+5. **情感分析** — 调用AI分析评论情感倾向
+6. **构建Prompt** — 结合AI角色人格、情感提示、文章内容、评论上下文构建Prompt
+7. **AI生成** — 调用AI模型生成回复内容
+8. **安全审核** — 对生成内容进行两阶段审核（安全检查 + 质量评分 1-5 分映射到 0-100）
+9. **发布/草稿** — 根据设置自动发布或存为草稿等待审核
+10. **重试** — 如果AI生成失败，系统会自动重试（最多 maxRetryCount 次），每次重试间隔递增

 ## 前置要求

- Halo 2.23+
- AI Foundation 插件（必须） — 本插件依赖 AI Foundation 提供的AI模型能力
+- Halo 2.25+
+- AI Foundation 插件（必须） — 本插件通过 `ExtensionGetter` 调用 AI Foundation 提供的 `AiModelService` 扩展点

 ## 技术栈

 - **后端**：Java + Spring WebFlux + Reactive
 - **前端**：Vue 3 + @halo-dev/components
- **AI**：支持 AI Foundation 插件集成
+- **AI**：通过 AI Foundation 插件集成
@@ -37,3 +37,15 @@ POST /apis/console.api.comment-ai-autopilot.nxxy335.top/v1alpha1/replies/{replyN
 ```

 对指定回复触发对话式AI回复。
+
+### 更新草稿回复内容
+
+```
+PUT /apis/console.api.comment-ai-autopilot.nxxy335.top/v1alpha1/replies/{name}/content
+```
+
+更新草稿状态的AI回复内容。请求体为 JSON 格式：`{"reply": "新的回复内容"}`。仅未发布的草稿回复可编辑。
+
+::: warning
+已发布的回复不可编辑。
+:::
@@ -1,6 +1,6 @@
 # AI角色

-AI角色定义了回复评论的虚拟身份，包括昵称、人格和头像。
+AI角色定义了回复评论的虚拟身份，包括昵称、人格、性别、语气风格和头像。

 ## 角色配置

@@ -8,6 +8,28 @@ AI角色定义了回复评论的虚拟身份，包括昵称、人格和头像。

 AI回复者的显示名称，默认为「小回」。修改后新回复将使用新昵称，已有回复不受影响。

+### 性别与语气
+
+每个角色可以设置性别（男/女），AI回复时会保持对应性别身份。语气风格通过"中性语气"复选框控制：
+
+- **勾选中性语气**：AI使用中性语气回复
+- **取消勾选**：AI根据性别使用对应语气风格（女性→温柔细腻，男性→沉稳理性）
+
+默认角色「小回」的性别为女，勾选中性语气。
+
+### 唤醒词
+
+评论以唤醒词开头可唤醒该角色回复。唤醒词功能的特点：
+
+- **跨页面唤醒**：在未启用AI回评的页面，使用唤醒词也能召唤AI回复
+- **二级评论支持**：回复中同样可以使用唤醒词唤醒指定角色
+- **独立唤醒**：每个角色有独立的唤醒词，可以唤醒不同角色
+- **留空不启用**：唤醒词留空则不启用该角色的唤醒功能
+
+::: tip
+唤醒词匹配时，插件会自动去除评论中的HTML标签并去除首尾空格，确保匹配准确。
+:::
+
 ### 人格提示词

 人格提示词定义了AI角色的性格和回复风格，是影响回复质量的关键配置。
@@ -31,7 +53,7 @@ AI回复者的显示名称，默认为「小回」。修改后新回复将使用
 填写邮箱后，AI回复者的头像将通过 Gravatar 服务自动生成：

 1. 插件根据邮箱生成 SHA-256 哈希
-2. 构造 Gravatar URL：`https://cn.cravatar.com/avatar/{hash}`
+2. 构造 Gravatar URL：`https://cn.cravatar.com/avatar/{hash}`（使用 [Cravatar](https://cn.cravatar.com) 镜像服务）
 3. 头像URL存储在评论的 `owner.annotations["avatar"]` 中

 ::: warning
@@ -9,6 +9,8 @@ Prompt模板控制AI生成回复时的完整提示词结构。

 {{safety_prompt}}

+【语言要求】请用评论所使用的语言回复。如果评论是英文，请用英文回复；如果是中文，请用中文回复；如果是日文，请用日文回复；以此类推。
+
 请回复以下评论。注意：
 - 回复长度应与评论长度匹配，简短问候简短回复
 - 不要复述或总结文章内容
@@ -21,6 +23,7 @@ Prompt模板控制AI生成回复时的完整提示词结构。
 文章（仅供理解上下文，不要复述）：
 {{article}}

+{{conversation_history}}
 评论：
 {{comment}}
 ```
@@ -29,23 +32,28 @@ Prompt模板控制AI生成回复时的完整提示词结构。

 | 变量 | 说明 | 注入时机 |
 |------|------|---------|
-| `{{persona_prompt}}` | AI角色人格提示词 | 始终注入 |
+| `{{persona_prompt}}` | AI角色人格提示词（含已启用的预设） | 始终注入 |
 | `{{safety_prompt}}` | 安全规范提示词 | 始终注入 |
-| `{{sentiment_prompt}}` | 情感语气提示词 | 情感分析后自动注入，不在模板中显式使用 |
 | `{{post_title}}` | 文章标题 | 始终注入 |
 | `{{post_date}}` | 文章发布日期（如 2024-01-15） | 始终注入 |
 | `{{comment_count}}` | 该文章的评论数 | 始终注入 |
 | `{{article}}` | 文章/页面内容（含标题） | 始终注入 |
+| `{{conversation_history}}` | 对话历史上下文 | 多轮对话时注入 |
 | `{{comment}}` | 评论内容（含评论者名称） | 始终注入 |
-| `{{conversation}}` | 对话上下文 | 多轮对话时注入 |
+
+::: warning 变量名注意
+对话上下文变量是 `{{conversation_history}}`（不是 `{{conversation}}`）。如果模板中使用了错误的变量名，该变量不会被替换。
+:::

 ## 情感提示

-情感提示由插件根据情感分析结果自动注入到Prompt中，不需要在模板中手动添加：
+情感提示由插件根据情感分析结果自动追加到 Prompt 末尾，不需要在模板中手动添加：

- **正面** → "评论者情绪积极友好，请用热情友好的语气回复，表达感谢和共鸣。"
- **负面** → "评论者情绪偏消极或不满，请用理性温和的语气回复，避免激化矛盾，适当表示理解。"
- **中性** → 不注入额外提示
+- **非常正面** → 追加"评论者情绪非常正面积极，请用热情洋溢的语气回复，表达真诚的感谢和共鸣。"
+- **正面** → 追加"评论者情绪正面积极，请用热情友好的语气回复，可以表达感谢和共鸣。"
+- **负面** → 追加"评论者情绪偏负面，请用理性温和的语气回复，避免激化矛盾，展现理解和包容。"
+- **非常负面** → 追加"评论者情绪非常负面，请用非常温和、理性的语气回复，避免任何可能激化矛盾的表达，展现充分的理解和耐心。"
+- **中性** → 不追加额外提示

 ## 安全提示

@@ -56,6 +64,17 @@ Prompt模板控制AI生成回复时的完整提示词结构。
 - 不生成虚假信息
 - 回复内容与评论相关

+## 预设风格
+
+在 Prompt 设置页面可以多选启用预设风格，启用后预设提示词会自动合并到 `{{persona_prompt}}` 之后：
+
+| 预设 | 说明 |
+|------|------|
+| 友好型 | 热情友好，多用感叹号和表情符号，口语化表达 |
+| 专业型 | 专业严谨，正式语言风格，有逻辑性 |
+| 幽默型 | 适当加入幽默元素，轻松诙谐但不过度 |
+| 简洁型 | 非常简洁，一两句话即可，不展开讨论 |
+
 ## 自定义建议

 自定义Prompt模板时，建议：
@@ -63,20 +82,7 @@ Prompt模板控制AI生成回复时的完整提示词结构。
 1. 保留 `{{persona_prompt}}` 和 `{{safety_prompt}}` 变量
 2. 保留 `{{article}}` 和 `{{comment}}` 变量
 3. 利用 `{{post_title}}`、`{{post_date}}`、`{{comment_count}}` 提供更丰富的上下文
-4. 在变量之间添加清晰的分隔和指令
-5. 避免让AI复述文章内容
-6. 控制回复长度和风格
-
-## 变量使用示例
-
-### 根据评论数调整回复风格
-
-```
-{{comment_count}}条评论说明这篇文章{{#if comment_count > 10}}很受欢迎{{/if}}。
-```
-
-### 利用发布日期
-
-```
-这篇文章发布于{{post_date}}，回复时请考虑时效性。
-```
+4. 保留 `{{conversation_history}}` 以支持多轮对话上下文
+5. 在变量之间添加清晰的分隔和指令
+6. 避免让AI复述文章内容
+7. 控制回复长度和风格
@@ -6,25 +6,29 @@

 | 分类 | 说明 | AI回复语气 |
 |------|------|-----------|
-| 正面 | 评论情绪积极、友好、感谢 | 热情友好，表达感谢和共鸣 |
-| 中性 | 评论情绪平淡、普通提问 | 正常语气回复，不加额外提示 |
-| 负面 | 评论情绪偏消极、不满、批评 | 理性温和，避免激化矛盾 |
+| 非常正面 | 强烈的感谢、赞美、认同（如"太棒了"、"非常感谢"） | 热情洋溢，表达真诚的感谢和共鸣 |
+| 正面 | 友好、肯定、支持（如"不错"、"学习了"） | 热情友好，表达感谢和共鸣 |
+| 中性 | 提问、讨论、陈述事实（如"请问..."、"这个怎么用"） | 正常语气回复，不加额外提示 |
+| 负面 | 不满、质疑、批评（如"不好用"、"有问题"） | 理性温和，避免激化矛盾 |
+| 非常负面 | 攻击、辱骂、极端情绪（如"垃圾"、"骗子"） | 非常温和理性，避免激化矛盾，展现理解和耐心 |

 ## 工作原理

 1. 评论通过过滤检查后，调用AI对评论内容进行情感分析
-2. AI返回情感分类结果（POSITIVE / NEUTRAL / NEGATIVE）
+2. AI 使用结构化输出（`OutputSpec.choice`）返回情感分类结果（VERY_POSITIVE / POSITIVE / NEUTRAL / NEGATIVE / VERY_NEGATIVE）
 3. 如果情感分析失败（如AI不可用），默认降级为 NEUTRAL
-4. 情感结果传入 PromptBuilder，在生成Prompt时注入对应的语气提示
+4. 情感结果传入 PromptBuilder，在生成Prompt时追加对应的语气提示
 5. 情感结果同时记录在 `AiCommentReply` 的 `sentiment` 字段中

 ## 日志展示

-在AI回复日志页面，每条记录会显示情感标签：
+在AI回复日志页面，每条记录会显示情感标签（纯色背景标签）：

- 🟢 **正面** — 绿色标签
- ⚪ **中性** — 灰色标签
- 🔴 **负面** — 红色标签
+- **非常正面** — 深绿色标签
+- **正面** — 浅绿色标签
+- **中性** — 灰色标签
+- **负面** — 浅红色标签
+- **非常负面** — 深红色标签

 ## 性能影响

@@ -1,30 +1,78 @@
 # 插件设置

-插件设置页面位于 **AI回评** → **插件设置**，包含以下配置组：
+插件设置页面位于 **AI回评** → **插件设置**，通过标签栏切换以下五个配置页面：
+
+- 基本设置
+- AI角色设置
+- 模型设置
+- Prompt设置
+- 数据清理
+
+页面右侧为操作控制侧边栏，显示保存按钮和未保存状态指示器。在 Prompt 设置页面，侧边栏还会显示可用模板变量列表。

 ## 基本设置

 | 配置项 | 说明 | 默认值 |
 |--------|------|--------|
 | 自动回复 | 是否启用自动回复功能 | 开启 |
-| 自动发布 | AI回复是否自动发布 | 开启 |
+| 自动发布 | AI回复是否自动发布，关闭则存为草稿 | 开启 |
+| 最大对话轮次 | 同一评论线程中AI最多自动回复的轮次 | 8 |
+| 速率限制 | 每分钟最大AI回复数量，防止批量评论消耗过多额度 | 10 |
 | 最大重试次数 | AI生成失败时的最大重试次数 | 3 |
-| 评论者黑名单 | 不触发AI回复的评论者显示名称或邮箱，逗号分隔 | 空 |
+| 评论者黑名单 | 不触发AI回复的评论者，支持名称、邮箱和正则表达式（`regex:` 开头），逗号分隔 | 空 |
+| 启用前置过滤 | AI回复前检测评论合规性，拦截广告/辱骂/敏感内容，节省Token | 开启 |
+| 违规评论设为待审核 | 检测到违规评论时自动取消通过，需人工审核 | 开启 |
+
+::: tip 评论者黑名单
+黑名单支持三种格式：
+- **名称**：如 `张三`
+- **邮箱**：如 `spam@example.com`（不区分大小写）
+- **正则表达式**：以 `regex:` 开头，如 `regex:^spam.*`
+
+点击"添加评论者"按钮可从已有评论列表中选择评论者自动添加到黑名单。
+:::
+
+::: tip 前置过滤（合规检测）
+启用前置过滤后，AI 在生成回复前会先对评论内容进行合规性分类，识别以下类别：
+
+- **正常**：放行，继续走 AI 回复流程
+- **广告**：包含推广链接、产品推销、引流信息等
+- **辱骂攻击**：包含辱骂、人身攻击、恶意挑衅、歧视性言论等
+- **敏感内容**：涉及政治敏感、违法违规、色情暴力等
+- **无意义**：纯乱码、无意义字符堆砌（如随机符号、键盘乱敲）
+
+对于非"正常"类别的评论，插件会：
+
+1. **停止生成 AI 回复**，节省 Token 与 API 调用
+2. 创建一条 `FILTERED` 状态的日志记录（可在日志页通过"已拦截"状态筛选查看）
+3. 若启用"违规评论设为待审核"，会自动将原评论的 `approved` 置为 `false`，使其进入待审核队列，需人工判断后审核通过
+
+被误拦截的评论可在日志页点击 **误报反馈** 按钮处理，支持"AI 回复"和"仅通过"两种方式。选择"仅通过"后记录变为"误报通过"状态，可随时点击"触发AI回复"按钮补生成回复。
+
+::: warning
+前置过滤依赖 AI Foundation 插件进行分类判断，会额外消耗少量 Token。若 AI 服务不可用或分类失败，为安全起见将拦截评论而非放行，防止违规内容漏网。
+:::
+:::

 ## AI角色设置

+AI角色定义了回复评论的虚拟身份。支持创建多个角色，每个角色有独立的昵称、人格提示词、性别、语气风格和 Gravatar 头像，可指定一个为默认角色。
+
 | 配置项 | 说明 | 默认值 |
 |--------|------|--------|
-| AI角色昵称 | AI回复者的显示名称 | 小回 |
-| AI角色人格提示词 | 定义AI角色的人格和回复风格 | 见下方 |
-| AI角色邮箱 | 用于Gravatar头像服务展示头像 | 空 |
+| 角色昵称 | AI回复者的显示名称 | 小回 |
+| 性别与语气 | 角色性别（男/女）+ 中性语气复选框（勾选=中性语气，取消勾选=跟随性别语气） | 女 + 中性语气 |
+| 唤醒词 | 评论以此词开头则唤醒该角色回复，留空不启用 | 空 |
+| 人格提示词 | 定义AI角色的人格和回复风格 | 见下方 |
+| 邮箱 | 用于 Gravatar 头像服务展示头像 | 空 |
+| 设为默认 | 将该角色设为默认角色 | 第一个角色默认 |

 默认人格提示词：

 > 你是「小回」，一个友善的评论者。你的回复简洁自然，像朋友聊天一样。简短的评论就简短回复，有深度的讨论才展开回应。不要长篇大论，不要复述文章内容。

 ::: tip Gravatar头像
-填写邮箱后，AI回复者的头像将通过 [Gravatar](https://gravatar.com) 服务自动生成。如果不填写邮箱，将使用默认头像。
+填写邮箱后，AI回复者的头像将通过 [Gravatar](https://gravatar.com) 服务自动生成，使用 [Cravatar](https://cn.cravatar.com) 镜像。如果不填写邮箱，将使用默认头像。
 :::

 ## 模型设置
@@ -42,17 +90,40 @@
 | 配置项 | 说明 | 默认值 |
 |--------|------|--------|
 | 自定义Prompt模板 | AI生成回复时使用的Prompt模板 | 见下方 |
+| 启用预设 | 选择要启用的Prompt预设风格（可多选） | 空 |
+
+### 预设风格
+
+| 预设 | 说明 |
+|------|------|
+| 友好型 | 热情友好，多用感叹号和表情符号，口语化表达 |
+| 专业型 | 专业严谨，正式语言风格，有逻辑性 |
+| 幽默型 | 适当加入幽默元素，轻松诙谐但不过度 |
+| 简洁型 | 非常简洁，一两句话即可，不展开讨论 |
+
+预设提示词会自动合并到角色人格提示词之后。

 ### 模板变量

-| 变量 | 说明 |
-|------|------|
-| `{{persona_prompt}}` | AI角色人格提示词 |
-| `{{safety_prompt}}` | 安全规范提示词 |
-| `{{sentiment_prompt}}` | 情感语气提示词（自动注入） |
-| `{{article}}` | 文章内容 |
-| `{{comment}}` | 评论内容 |
-| `{{conversation}}` | 对话上下文（多轮对话时） |
+| 变量 | 说明 | 注入时机 |
+|------|------|---------|
+| `{{persona_prompt}}` | AI角色人格提示词（含已启用的预设） | 始终注入 |
+| `{{safety_prompt}}` | 安全规范提示词 | 始终注入 |
+| `{{post_title}}` | 文章标题 | 始终注入 |
+| `{{post_date}}` | 文章发布日期（如 2024-01-15） | 始终注入 |
+| `{{comment_count}}` | 该文章的评论数 | 始终注入 |
+| `{{article}}` | 文章/页面内容（含标题） | 始终注入 |
+| `{{conversation_history}}` | 对话历史上下文 | 多轮对话时注入 |
+| `{{comment}}` | 评论内容（含评论者名称） | 始终注入 |
+
+::: tip 情感提示
+情感提示由插件根据情感分析结果自动追加到 Prompt 末尾，不需要在模板中手动添加：
+- **非常正面** → 追加热情洋溢的语气提示
+- **正面** → 追加热情友好的语气提示
+- **负面** → 追加理性温和的语气提示
+- **非常负面** → 追加冷静关怀的语气提示
+- **中性** → 不追加额外提示
+:::

 ## 数据清理

@@ -64,3 +135,25 @@
 ::: tip
 你也可以在数据清理页面点击"立即清理"按钮手动触发清理操作。
 :::
+
+::: warning
+清理操作仅删除 `AiCommentReply` 记录（插件内部的日志记录），不会删除已发布的 Halo Reply 评论。
+:::
+
+## 配置导入导出
+
+插件设置页面顶部提供导入导出按钮，方便备份和迁移配置。
+
+### 导出配置
+
+点击 **导出** 按钮，将当前配置（包括 ConfigMap 数据和所有 AI 角色）导出为 JSON 文件。
+
+### 导入配置
+
+1. 点击 **导入** 按钮，选择 JSON 配置文件
+2. 确认导入操作（导入会覆盖当前配置，不可撤销）
+3. 导入完成后自动刷新设置和角色列表
+
+::: warning
+导入操作会覆盖当前配置，请谨慎操作。建议在导入前先导出当前配置作为备份。
+:::
@@ -16,14 +16,16 @@ hero:
 features:
  - title: 自动回复
    details: 监听新评论，自动调用AI生成回复，支持对话式上下文和失败重试
-  - title: 多语言适配
-    details: 根据评论语言自动用对应语言回复，中文评论中文回复，英文评论英文回复
+  - title: 多 AI 角色
+    details: 创建多个虚拟角色，独立昵称、人格、性别、语气和 Gravatar 头像
  - title: 情感分析
    details: 分析评论情感倾向，根据正面/中性/负面调整回复语气
+  - title: 前置过滤
+    details: AI回复前检测评论合规性，拦截广告/辱骂/敏感内容，节省Token
  - title: 草稿模式
    details: AI回复先存为草稿，管理员审核后再发布，支持批量操作
-  - title: 灵活过滤
-    details: 文章/页面级开关控制，评论者黑名单支持名称和邮箱匹配
+  - title: 对话上下文
+    details: 查看完整对话上下文，支持引用摘要展示和头像显示
  - title: 数据管理
-    details: 仪表盘统计、日志筛选搜索、自动清理旧记录
+    details: 仪表盘统计、日志筛选搜索、自动清理旧记录、配置导入导出
 ---
@@ -1 +1,4 @@
-version=1.0.0-SNAPSHOT
+version=1.2.1
+
+# Fix Windows Gradle Worker Daemon exit code 268435659 when running pnpm via Exec tasks
+org.gradle.daemon=false
@@ -1,6 +1,10 @@
 package top.nxxy335.commentaiautopilot;

+import com.fasterxml.jackson.databind.JsonNode;
+import com.fasterxml.jackson.databind.ObjectMapper;
+import com.fasterxml.jackson.databind.node.ObjectNode;
 import org.springframework.stereotype.Component;
+import run.halo.app.extension.ConfigMap;
 import run.halo.app.extension.ReactiveExtensionClient;
 import run.halo.app.extension.index.IndexSpecs;
 import run.halo.app.extension.Scheme;
@@ -25,13 +29,18 @@ import reactor.core.publisher.Mono;
@Component
 public class CommentAiAutopilotPlugin extends BasePlugin {

+    private static final String CONFIG_MAP_NAME = "comment-ai-autopilot-configmap";
+
    private final SchemeManager schemeManager;
    private final ReactiveExtensionClient client;
+    private final ObjectMapper objectMapper;

-    public CommentAiAutopilotPlugin(PluginContext pluginContext, SchemeManager schemeManager, ReactiveExtensionClient client) {
+    public CommentAiAutopilotPlugin(PluginContext pluginContext, SchemeManager schemeManager,
+                                     ReactiveExtensionClient client, ObjectMapper objectMapper) {
        super(pluginContext);
        this.schemeManager = schemeManager;
        this.client = client;
+        this.objectMapper = objectMapper;
    }

    @Override
@@ -43,11 +52,60 @@ public class CommentAiAutopilotPlugin extends BasePlugin {
                .indexFunc(ext -> ext.getSpec().getPostId()));
            indexSpecs.add(IndexSpecs.<AiCommentReply, String>single("spec.status", String.class)
                .indexFunc(ext -> ext.getSpec().getStatus()));
+            indexSpecs.add(IndexSpecs.<AiCommentReply, String>single("spec.sentiment", String.class)
+                .indexFunc(ext -> ext.getSpec().getSentiment()));
+            indexSpecs.add(IndexSpecs.<AiCommentReply, String>single("spec.published", String.class)
+                .indexFunc(ext -> String.valueOf(ext.getSpec().getPublished())));
+            indexSpecs.add(IndexSpecs.<AiCommentReply, String>single("spec.postKind", String.class)
+                .indexFunc(ext -> ext.getSpec().getPostKind()));
        });
        schemeManager.register(AiPersona.class);

        // 初始化默认AI角色"小回"
        initDefaultPersona();
+
+        // 迁移：确保升级用户的前置过滤配置正确
+        migratePreFilterConfig();
+    }
+
+    /**
+     * 迁移前置过滤配置：从 v1.0.x 升级到 v1.1.0 时，
+     * ConfigMap 中可能保存了旧默认值 preFilterEnabled=false，
+     * 需要将其更新为 true（新默认值）。
+     */
+    private void migratePreFilterConfig() {
+        client.fetch(ConfigMap.class, CONFIG_MAP_NAME)
+            .flatMap(cm -> {
+                var data = cm.getData();
+                if (data == null) return Mono.empty();
+                String basicJson = data.get("basic");
+                if (basicJson == null || basicJson.isBlank()) return Mono.empty();
+                try {
+                    JsonNode node = objectMapper.readTree(basicJson);
+                    if (!node.has("preFilterEnabled")) {
+                        // 字段不存在，添加并设为 true
+                        ((ObjectNode) node).put("preFilterEnabled", true);
+                        data.put("basic", objectMapper.writeValueAsString(node));
+                        return client.update(cm)
+                            .doOnSuccess(c -> log.info("[Migration] Added preFilterEnabled=true to ConfigMap"));
+                    }
+                    if (node.has("preFilterEnabled") && !node.get("preFilterEnabled").asBoolean(true)) {
+                        // 字段存在但为 false（旧默认值），迁移为 true
+                        ((ObjectNode) node).put("preFilterEnabled", true);
+                        data.put("basic", objectMapper.writeValueAsString(node));
+                        return client.update(cm)
+                            .doOnSuccess(c -> log.info("[Migration] Migrated preFilterEnabled from false to true"));
+                    }
+                } catch (Exception e) {
+                    log.warn("[Migration] Failed to migrate preFilter config: {}", e.getMessage());
+                }
+                return Mono.empty();
+            })
+            .subscribe(
+                null,
+                err -> log.debug("[Migration] PreFilter config migration skipped: {}", err.getMessage()),
+                () -> log.debug("[Migration] PreFilter config migration check completed")
+            );
    }

    private void initDefaultPersona() {
@@ -61,6 +119,9 @@ public class CommentAiAutopilotPlugin extends BasePlugin {
                spec.setDisplayName("小回");
                spec.setPrompt("你是一个友善的评论者，回复简洁自然，像朋友聊天一样。");
                spec.setEmail("");
+                spec.setGender("female");
+                spec.setNeutralVoice(true);
+                spec.setWakeWord("小回小回");
                spec.setIsDefault(true);
                persona.setSpec(spec);
                return client.create(persona);
@@ -0,0 +1,14 @@
+package top.nxxy335.commentaiautopilot;
+
+import com.fasterxml.jackson.databind.ObjectMapper;
+import org.springframework.context.annotation.Bean;
+import org.springframework.context.annotation.Configuration;
+
+@Configuration
+public class ObjectMapperConfiguration {
+
+    @Bean
+    ObjectMapper objectMapper() {
+        return new ObjectMapper();
+    }
+}
@@ -8,7 +8,6 @@ import org.springframework.web.reactive.function.server.ServerResponse;
 import reactor.core.publisher.Flux;
 import reactor.core.publisher.Mono;
 import run.halo.app.core.extension.content.Comment;
-import run.halo.app.core.extension.content.Post;
 import run.halo.app.core.extension.content.Reply;
 import run.halo.app.core.extension.endpoint.CustomEndpoint;
 import run.halo.app.extension.ConfigMap;
@@ -16,6 +15,7 @@ import run.halo.app.extension.Metadata;
 import run.halo.app.extension.GroupVersion;
 import run.halo.app.extension.ListOptions;
 import run.halo.app.extension.ListResult;
+import run.halo.app.extension.index.query.Queries;
 import run.halo.app.extension.ReactiveExtensionClient;
 import run.halo.app.extension.PageRequestImpl;
 import top.nxxy335.commentaiautopilot.extension.AiCommentReply;
@@ -24,6 +24,8 @@ import top.nxxy335.commentaiautopilot.service.AiFoundationClient;
 import top.nxxy335.commentaiautopilot.service.AiReplyCleanupService;
 import top.nxxy335.commentaiautopilot.service.AiReplyOrchestrator;
 import top.nxxy335.commentaiautopilot.service.CommentReplyPublisher;
+import top.nxxy335.commentaiautopilot.service.PersonaResolver;
+import top.nxxy335.commentaiautopilot.util.GravatarUtil;

 import com.fasterxml.jackson.databind.JsonNode;
 import com.fasterxml.jackson.databind.ObjectMapper;
@@ -56,16 +58,18 @@ public class CommentAiAutopilotEndpoint implements CustomEndpoint {
    private final AiFoundationClient aiFoundationClient;
    private final CommentReplyPublisher commentReplyPublisher;
    private final ObjectMapper objectMapper;
+    private final PersonaResolver personaResolver;

    private static final String CONFIG_MAP_NAME = "comment-ai-autopilot-configmap";

-    public CommentAiAutopilotEndpoint(ReactiveExtensionClient client, AiReplyOrchestrator orchestrator, AiReplyCleanupService cleanupService, AiFoundationClient aiFoundationClient, CommentReplyPublisher commentReplyPublisher) {
+    public CommentAiAutopilotEndpoint(ReactiveExtensionClient client, AiReplyOrchestrator orchestrator, AiReplyCleanupService cleanupService, AiFoundationClient aiFoundationClient, CommentReplyPublisher commentReplyPublisher, ObjectMapper objectMapper, PersonaResolver personaResolver) {
        this.client = client;
        this.orchestrator = orchestrator;
        this.cleanupService = cleanupService;
        this.aiFoundationClient = aiFoundationClient;
        this.commentReplyPublisher = commentReplyPublisher;
-        this.objectMapper = new ObjectMapper();
+        this.objectMapper = objectMapper;
+        this.personaResolver = personaResolver;
    }

    @Override
@@ -97,6 +101,8 @@ public class CommentAiAutopilotEndpoint implements CustomEndpoint {
            .POST("/import", this::importConfig)
            // 更新草稿回复内容（同时更新 AiCommentReply 和 Reply 扩展）
            .PUT("/replies/{name}/content", this::updateReplyContent)
+            // 误报反馈：将拦截记录标记为误报，可选触发AI回复
+            .POST("/replies/{name}/false-positive", this::falsePositive)
            .build();
    }

@@ -133,19 +139,26 @@ public class CommentAiAutopilotEndpoint implements CustomEndpoint {
        final Instant finalStartInstant = startInstant;
        final Instant finalEndInstant = endInstant;

-        // Check if we need in-memory filtering (keyword, date range, status, or sentiment)
-        boolean needsMemoryFilter = !keywordFilter.isBlank() || finalStartInstant != null || finalEndInstant != null
-            || !statusFilter.isBlank() || !sentimentFilter.isBlank();
+        // Check if we need in-memory filtering (keyword or date range)
+        boolean needsMemoryFilter = !keywordFilter.isBlank() || finalStartInstant != null || finalEndInstant != null;
+
+        // Build server-side query for status and sentiment (indexed fields)
+        var listOptionsBuilder = ListOptions.builder();
+        if (!statusFilter.isBlank()) {
+            listOptionsBuilder.andQuery(Queries.equal("spec.status", statusFilter));
+        }
+        if (!sentimentFilter.isBlank()) {
+            listOptionsBuilder.andQuery(Queries.equal("spec.sentiment", sentimentFilter));
+        }
+        var listOptions = listOptionsBuilder.build();

        if (needsMemoryFilter) {
-            // Fall back to listAll + in-memory filter for complex queries
-            return client.listAll(AiCommentReply.class, ListOptions.builder().build(), Sort.unsorted())
+            // Fall back to listAll + in-memory filter for keyword/date queries
+            return client.listAll(AiCommentReply.class, listOptions, Sort.unsorted())
                .collectList()
                .map(replies -> {
                    var filtered = replies.stream()
                        .filter(r -> {
-                            if (!statusFilter.isBlank() && !statusFilter.equals(r.getSpec().getStatus())) return false;
-                            if (!sentimentFilter.isBlank() && !sentimentFilter.equals(r.getSpec().getSentiment())) return false;
                            if (!keywordFilter.isBlank()) {
                                String reply = r.getSpec().getReply();
                                if (reply == null || !reply.contains(keywordFilter)) return false;
@@ -184,13 +197,11 @@ public class CommentAiAutopilotEndpoint implements CustomEndpoint {
                .flatMap(result -> ServerResponse.ok().bodyValue(result));
        }

-        // No filters - use server-side pagination directly
+        // No memory filters needed - use server-side pagination directly
        Sort sort = "asc".equalsIgnoreCase(sortOrder)
            ? Sort.by(Sort.Order.asc("metadata.creationTimestamp"))
            : Sort.by(Sort.Order.desc("metadata.creationTimestamp"));

-        var listOptions = ListOptions.builder().build();
-
        return client.listBy(AiCommentReply.class, listOptions,
                PageRequestImpl.of(page - 1, size, sort))
            .map(listResult -> {
@@ -216,100 +227,25 @@ public class CommentAiAutopilotEndpoint implements CustomEndpoint {
    }

    private Mono<ServerResponse> getStats(ServerRequest request) {
-        String range = request.queryParam("range").orElse("7");
-
        return client.listAll(AiCommentReply.class, ListOptions.builder().build(), Sort.unsorted())
            .collectList()
-            .map(allReplies -> {
-                // 根据 range 计算截止时间
-                ZoneId zoneId = ZoneId.systemDefault();
-                LocalDate today = LocalDate.now(zoneId);
-                Instant cutoffInstant;
-                int trendDays;
-
-                if ("all".equals(range)) {
-                    cutoffInstant = null; // 不做时间过滤
-                    trendDays = 30; // "all" 时趋势也展示最近30天
-                } else {
-                    int days = Integer.parseInt(range);
-                    cutoffInstant = today.minusDays(days).atStartOfDay(zoneId).toInstant();
-                    trendDays = days;
-                }
-
-                // 根据 range 过滤记录
-                List<AiCommentReply> replies;
-                if (cutoffInstant != null) {
-                    replies = allReplies.stream()
-                        .filter(r -> {
-                            Instant ts = r.getMetadata().getCreationTimestamp();
-                            return ts != null && !ts.isBefore(cutoffInstant);
-                        })
-                        .toList();
-                } else {
-                    replies = allReplies;
-                }
-
+            .map(replies -> {
                long total = replies.size();
                long passCount = replies.stream()
                    .filter(r -> "PASS".equals(r.getSpec().getStatus())).count();
                long failCount = replies.stream()
                    .filter(r -> "FAIL".equals(r.getSpec().getStatus())).count();
-                double avgScore = replies.stream()
-                    .filter(r -> r.getSpec().getScore() != null && r.getSpec().getScore() > 0)
-                    .mapToInt(r -> r.getSpec().getScore())
-                    .average().orElse(0.0);

                long reviewingCount = replies.stream()
                    .filter(r -> "PASS".equals(r.getSpec().getStatus())
                        && !Boolean.TRUE.equals(r.getSpec().getPublished()))
                    .count();

-                Map<String, Long> sentimentDistribution = new HashMap<>();
-                sentimentDistribution.put("POSITIVE", 0L);
-                sentimentDistribution.put("NEUTRAL", 0L);
-                sentimentDistribution.put("NEGATIVE", 0L);
-                sentimentDistribution.put("UNKNOWN", 0L);
-                for (var r : replies) {
-                    String sentiment = r.getSpec().getSentiment();
-                    if (sentiment == null || sentiment.isBlank()) {
-                        sentimentDistribution.merge("UNKNOWN", 1L, Long::sum);
-                    } else {
-                        sentimentDistribution.merge(sentiment, 1L, Long::sum);
-                    }
-                }
-
-                // 计算 dailyTrend
-                DateTimeFormatter formatter = DateTimeFormatter.ofPattern("yyyy-MM-dd");
-                Map<LocalDate, Long> dailyMap = new HashMap<>();
-                for (int i = 0; i < trendDays; i++) {
-                    dailyMap.put(today.minusDays(i), 0L);
-                }
-                for (var r : replies) {
-                    Instant timestamp = r.getMetadata().getCreationTimestamp();
-                    if (timestamp != null) {
-                        try {
-                            LocalDate date = timestamp.atZone(zoneId).toLocalDate();
-                            if (dailyMap.containsKey(date)) {
-                                dailyMap.merge(date, 1L, Long::sum);
-                            }
-                        } catch (Exception ignored) {
-                        }
-                    }
-                }
-                List<DailyCount> dailyTrend = new ArrayList<>();
-                for (int i = 0; i < trendDays; i++) {
-                    LocalDate date = today.minusDays(i);
-                    dailyTrend.add(new DailyCount(date.format(formatter), dailyMap.get(date)));
-                }
-
-                return new StatsResponse(total, passCount, failCount, avgScore,
-                    reviewingCount, sentimentDistribution, dailyTrend);
+                return new StatsResponse(total, passCount, failCount, reviewingCount);
            })
            .onErrorResume(e -> {
                log.warn("Failed to fetch stats: {}", e.getMessage());
-                return Mono.just(new StatsResponse(0, 0, 0, 0.0, 0L,
-                    Map.of("POSITIVE", 0L, "NEUTRAL", 0L, "NEGATIVE", 0L, "UNKNOWN", 0L),
-                    List.of()));
+                return Mono.just(new StatsResponse(0, 0, 0, 0));
            })
            .flatMap(stats -> ServerResponse.ok().bodyValue(stats));
    }
@@ -321,18 +257,7 @@ public class CommentAiAutopilotEndpoint implements CustomEndpoint {
            .next()
            .flatMap(persona -> {
                String email = persona.getSpec().getEmail();
-                String avatarUrl = "";
-                if (email != null && !email.isBlank()) {
-                    try {
-                        var digest = java.security.MessageDigest.getInstance("SHA-256");
-                        var hashBytes = digest.digest(email.trim().toLowerCase().getBytes(java.nio.charset.StandardCharsets.UTF_8));
-                        var hexString = new StringBuilder();
-                        for (byte b : hashBytes) {
-                            hexString.append(String.format("%02x", b));
-                        }
-                        avatarUrl = "https://cn.cravatar.com/avatar/" + hexString;
-                    } catch (Exception ignored) {}
-                }
+                String avatarUrl = GravatarUtil.generateUrl(email);
                return ServerResponse.ok().bodyValue(Map.of(
                    "name", persona.getSpec().getDisplayName(),
                    "prompt", persona.getSpec().getPrompt() != null ? persona.getSpec().getPrompt() : "",
@@ -346,16 +271,11 @@ public class CommentAiAutopilotEndpoint implements CustomEndpoint {
            )));
    }

-    public record DailyCount(String date, long count) {}
-
    public record StatsResponse(
        long total,
        long passCount,
        long failCount,
-        double avgScore,
-        long reviewingCount,
-        Map<String, Long> sentimentDistribution,
-        List<DailyCount> dailyTrend
+        long reviewingCount
    ) {}

    public record PersonaResponse(
@@ -373,25 +293,51 @@ public class CommentAiAutopilotEndpoint implements CustomEndpoint {
                var commentTime = String.valueOf(comment.getMetadata().getCreationTimestamp());
                var isCommentAi = isAiOwner(comment.getSpec().getOwner());

+                // 首条评论没有引用对象
                var commentMsg = new ConversationMessage(
-                    "comment", commentOwner, commentContent, commentTime, isCommentAi
+                    "comment", commentOwner, commentContent, commentTime, isCommentAi, null, null
                );

-                return client.listAll(Reply.class, ListOptions.builder().build(), Sort.unsorted())
-                    .filter(reply -> commentName.equals(reply.getSpec().getCommentName()))
+                return client.list(Reply.class,
+                        reply -> commentName.equals(reply.getSpec().getCommentName()),
+                        null)
                    .sort(Comparator.comparing(r -> r.getMetadata().getCreationTimestamp()))
-                    .map(reply -> {
+                    .collectList() // 收集为List以便统一处理引用映射
+                    .map(replyList -> {
+                        List<ConversationMessage> messages = new ArrayList<>();
+                        messages.add(commentMsg);
+
+                        // 构建 Reply 的映射字典，方便查找引用关系
+                        Map<String, Reply> replyMap = new HashMap<>();
+                        for (Reply r : replyList) {
+                            replyMap.put(r.getMetadata().getName(), r);
+                        }
+
+                        for (Reply reply : replyList) {
                            var replyOwner = extractOwnerName(reply.getSpec().getOwner());
                            var replyContent = extractContent(reply.getSpec().getRaw(), reply.getSpec().getContent());
                            var replyTime = String.valueOf(reply.getMetadata().getCreationTimestamp());
                            var isAi = isAiOwner(reply.getSpec().getOwner());
-                        return new ConversationMessage("reply", replyOwner, replyContent, replyTime, isAi);
-                    })
-                    .collectList()
-                    .map(replyList -> {
-                        List<ConversationMessage> messages = new ArrayList<>();
-                        messages.add(commentMsg);
-                        messages.addAll(replyList);
+
+                            String quoteOwner = null;
+                            String quoteContent = null;
+
+                            // 获取引用的 Reply 名称 (Halo中如果为空，代表直接回复顶级 Comment)
+                            String quoteReplyName = reply.getSpec().getQuoteReply();
+                            if (quoteReplyName != null && !quoteReplyName.isBlank()) {
+                                Reply quotedReply = replyMap.get(quoteReplyName);
+                                if (quotedReply != null) {
+                                    quoteOwner = extractOwnerName(quotedReply.getSpec().getOwner());
+                                    quoteContent = extractContent(quotedReply.getSpec().getRaw(), quotedReply.getSpec().getContent());
+                                }
+                            } else {
+                                // 没有 quoteReply 表示直接回复首条评论
+                                quoteOwner = commentOwner;
+                                quoteContent = commentContent;
+                            }
+
+                            messages.add(new ConversationMessage("reply", replyOwner, replyContent, replyTime, isAi, quoteOwner, quoteContent));
+                        }
                        return messages;
                    });
            })
@@ -722,9 +668,9 @@ public class CommentAiAutopilotEndpoint implements CustomEndpoint {
                        .bodyValue(Map.of("message", "该评论已有AI回复记录"));
                }
                // Read persona name from post annotations
-                return getPersonaNameFromComment(commentName)
+                return personaResolver.getPersonaNameFromComment(commentName)
                    .flatMap(personaName ->
-                        orchestrator.processComment(commentName, null, false, personaName)
+                        orchestrator.processComment(commentName, null, false, personaName, false)
                            .then(ServerResponse.ok().bodyValue(Map.of("message", "已触发AI回复")))
                    );
            });
@@ -749,9 +695,9 @@ public class CommentAiAutopilotEndpoint implements CustomEndpoint {
                            return ServerResponse.badRequest()
                                .bodyValue(Map.of("message", "该回复已有AI对话记录"));
                        }
-                        return getPersonaNameFromComment(commentName)
+                        return personaResolver.getPersonaNameFromComment(commentName)
                            .flatMap(personaName ->
-                                orchestrator.processComment(commentName, replyName, true, personaName)
+                                orchestrator.processComment(commentName, replyName, true, personaName, false)
                                    .then(ServerResponse.ok().bodyValue(Map.of("message", "已触发AI对话回复")))
                            );
                    });
@@ -759,31 +705,6 @@ public class CommentAiAutopilotEndpoint implements CustomEndpoint {
            .switchIfEmpty(ServerResponse.notFound().build());
    }

-    private static final String AI_PERSONA_ANNOTATION = "comment-ai-autopilot.nxxy335.top/ai-persona";
-
-    private Mono<String> getPersonaNameFromComment(String commentName) {
-        return client.fetch(Comment.class, commentName)
-            .flatMap(comment -> {
-                var subjectRef = comment.getSpec().getSubjectRef();
-                if (subjectRef == null || !"Post".equals(subjectRef.getKind())) {
-                    return Mono.justOrEmpty(null);
-                }
-                String postName = subjectRef.getName();
-                return client.fetch(Post.class, postName)
-                    .mapNotNull(post -> {
-                        var annotations = post.getMetadata().getAnnotations();
-                        if (annotations != null) {
-                            String persona = annotations.get(AI_PERSONA_ANNOTATION);
-                            if (persona != null && !persona.isBlank()) {
-                                return persona;
-                            }
-                        }
-                        return null;
-                    });
-            })
-            .defaultIfEmpty("");
-    }
-
    private Mono<Reply> findReplyForRecord(AiCommentReply record) {
        // First try using replyName if available
        String replyName = record.getSpec().getReplyName();
@@ -818,7 +739,9 @@ public class CommentAiAutopilotEndpoint implements CustomEndpoint {
        String owner,
        String content,
        String time,
-        boolean isAi
+        boolean isAi,
+        String quoteOwner,
+        String quoteContent
    ) {}

    public record CommenterInfo(
@@ -828,7 +751,8 @@ public class CommentAiAutopilotEndpoint implements CustomEndpoint {
    ) {}

    private Mono<ServerResponse> listCommenters(ServerRequest request) {
-        return client.listAll(Comment.class, ListOptions.builder().build(), Sort.unsorted())
+        return client.list(Comment.class, null, null)
+            .take(1000)
            .collectList()
            .map(comments -> {
                Set<String> seen = new HashSet<>();
@@ -846,7 +770,7 @@ public class CommentAiAutopilotEndpoint implements CustomEndpoint {
                        if (owner.getAnnotations() != null && owner.getAnnotations().get(Comment.CommentOwner.AVATAR_ANNO) != null) {
                            avatarUrl = owner.getAnnotations().get(Comment.CommentOwner.AVATAR_ANNO);
                        } else if (!email.isBlank()) {
-                            avatarUrl = generateGravatarUrl(email);
+                            avatarUrl = GravatarUtil.generateUrl(email);
                        }
                        result.add(new CommenterInfo(displayName, email, avatarUrl));
                    }
@@ -856,26 +780,11 @@ public class CommentAiAutopilotEndpoint implements CustomEndpoint {
            .flatMap(commenters -> ServerResponse.ok().bodyValue(commenters));
    }

-    private String generateGravatarUrl(String email) {
-        try {
-            var digest = java.security.MessageDigest.getInstance("SHA-256");
-            var hashBytes = digest.digest(email.trim().toLowerCase().getBytes(java.nio.charset.StandardCharsets.UTF_8));
-            var hexString = new StringBuilder();
-            for (byte b : hashBytes) {
-                hexString.append(String.format("%02x", b));
-            }
-            return "https://cn.cravatar.com/avatar/" + hexString;
-        } catch (Exception e) {
-            return "";
-        }
-    }
-
    private Mono<ServerResponse> triggerCleanup(ServerRequest request) {
-        return Mono.fromCallable(() -> {
-                int retentionDays = cleanupService.getRetentionDays();
-                long deleted = cleanupService.executeCleanup(retentionDays);
-                return Map.of("deletedCount", deleted, "retentionDays", retentionDays);
-            })
+        return cleanupService.getRetentionDays()
+            .flatMap(retentionDays -> cleanupService.executeCleanup(retentionDays)
+                .map(deleted -> Map.of("deletedCount", deleted, "retentionDays", retentionDays))
+            )
            .flatMap(result -> ServerResponse.ok().bodyValue(result))
            .onErrorResume(e -> {
                log.warn("Failed to trigger cleanup: {}", e.getMessage());
@@ -1041,7 +950,6 @@ public class CommentAiAutopilotEndpoint implements CustomEndpoint {
                    for (var personaData : personasList) {
                        importMono = importMono.then(Mono.defer(() -> {
                            try {
-                                var objectMapper = new com.fasterxml.jackson.databind.ObjectMapper();
                                var personaJson = objectMapper.writeValueAsString(personaData);
                                var persona = objectMapper.readValue(personaJson, AiPersona.class);
                                var personaName = persona.getMetadata().getName();
@@ -1129,4 +1037,131 @@ public class CommentAiAutopilotEndpoint implements CustomEndpoint {
                    .switchIfEmpty(ServerResponse.notFound().build());
            });
    }
+
+    /**
+     * 误报反馈：将被拦截的评论标记为误报（正常），并可选触发 AI 回复。
+     *
+     * 请求体：{ "action": "aiReply" | "approveOnly" }
+     * - aiReply: 将评论审核状态设为已通过 + 触发 AI 生成回复
+     * - approveOnly: 仅将评论审核状态设为已通过，不触发 AI 回复
+     */
+    private Mono<ServerResponse> falsePositive(ServerRequest request) {
+        var name = request.pathVariable("name");
+        return request.bodyToMono(String.class)
+            .flatMap(body -> {
+                String actionStr;
+                try {
+                    JsonNode node = objectMapper.readTree(body);
+                    actionStr = node.has("action") ? node.get("action").asText("approveOnly") : "approveOnly";
+                } catch (Exception e) {
+                    actionStr = "approveOnly";
+                }
+                final String action = actionStr;
+
+                return client.fetch(AiCommentReply.class, name)
+                    .flatMap(record -> {
+                        String currentStatus = record.getSpec().getStatus();
+                        // 允许：FILTERED（拦截误报）、FALSE_POSITIVE（已通过但可触发AI）、FAIL（AI生成失败可重试）
+                        if (!"FILTERED".equals(currentStatus)
+                            && !"FALSE_POSITIVE".equals(currentStatus)
+                            && !"FAIL".equals(currentStatus)) {
+                            return ServerResponse.badRequest()
+                                .bodyValue(Map.of("message", "仅已拦截、误报通过或AI生成失败的记录可进行此操作"));
+                        }
+
+                        String commentName = record.getSpec().getCommentId();
+                        String replyName = record.getSpec().getReplyTo();
+
+                        // 1. 将原评论/回复的审核状态设为已通过
+                        Mono<Void> approveMono = approveOriginalComment(commentName, replyName);
+
+                        // 2. 更新 AiCommentReply 记录状态
+                        Mono<Void> updateRecordMono = Mono.defer(() -> client.fetch(AiCommentReply.class, name)
+                            .flatMap(latest -> {
+                                latest.getSpec().setFilterCategory("误报");
+                                latest.getSpec().setFilterReason("用户确认为误报，已通过");
+                                if ("aiReply".equals(action)) {
+                                    latest.getSpec().setStatus("PENDING");
+                                    latest.getSpec().setReply("");
+                                } else {
+                                    // 仅通过：使用 FALSE_POSITIVE 状态，区别于 PASS
+                                    // 避免前端显示"通过/拒绝"按钮和"未发布"标签
+                                    latest.getSpec().setStatus("FALSE_POSITIVE");
+                                    latest.getSpec().setPublished(false);
+                                }
+                                return client.update(latest);
+                            })
+                            .retryWhen(Retry.backoff(3, Duration.ofMillis(100))
+                                .filter(e -> e instanceof OptimisticLockingFailureException))
+                            .then());
+
+                        // 3. 异步触发 AI 回复（在记录更新完成后，不阻塞 HTTP 响应）
+                        // 使用 processFalsePositive 跳过前置过滤和去重检查
+                        final boolean isConversation = Boolean.TRUE.equals(record.getSpec().getIsAiConversation());
+                        final String recordName = record.getMetadata().getName();
+
+                        return approveMono
+                            .then(updateRecordMono)
+                            .doOnSuccess(v -> {
+                                if ("aiReply".equals(action)) {
+                                    personaResolver.getPersonaNameFromComment(commentName)
+                                        .flatMap(personaName ->
+                                            orchestrator.processFalsePositive(commentName, replyName, isConversation, personaName, recordName)
+                                        )
+                                        .subscribe(
+                                            null,
+                                            err -> log.warn("[FalsePositive] AI reply trigger failed for {}: {}", commentName, err.getMessage()),
+                                            () -> log.info("[FalsePositive] AI reply trigger completed for {}", commentName)
+                                        );
+                                }
+                            })
+                            .then(ServerResponse.ok().bodyValue(Map.of(
+                                "message", "aiReply".equals(action) ? "已标记为误报，AI回复正在后台生成" : "已标记为误报并通过"
+                            )));
+                    })
+                    .switchIfEmpty(ServerResponse.notFound().build());
+            });
+    }
+
+    /**
+     * 将被拦截评论的原 Comment 或 Reply 审核状态设为已通过。
+     */
+    private Mono<Void> approveOriginalComment(String commentName, String replyName) {
+        // 优先处理 Reply（AI 对话场景下违规内容来自 Reply）
+        if (replyName != null && !replyName.isBlank()) {
+            return client.fetch(Reply.class, replyName)
+                .flatMap(reply -> {
+                    var spec = reply.getSpec();
+                    if (spec != null && !Boolean.TRUE.equals(spec.getApproved())) {
+                        spec.setApproved(true);
+                        spec.setApprovedTime(Instant.now());
+                        return client.update(reply)
+                            .retryWhen(Retry.backoff(3, Duration.ofMillis(100))
+                                .filter(e -> e instanceof OptimisticLockingFailureException))
+                            .doOnSuccess(r -> log.info("[FalsePositive] Reply {} approved", replyName))
+                            .then();
+                    }
+                    return Mono.empty();
+                })
+                .switchIfEmpty(Mono.defer(() -> approveComment(commentName)));
+        }
+        return approveComment(commentName);
+    }
+
+    private Mono<Void> approveComment(String commentName) {
+        return client.fetch(Comment.class, commentName)
+            .flatMap(comment -> {
+                var spec = comment.getSpec();
+                if (spec != null && !Boolean.TRUE.equals(spec.getApproved())) {
+                    spec.setApproved(true);
+                    spec.setApprovedTime(Instant.now());
+                    return client.update(comment)
+                        .retryWhen(Retry.backoff(3, Duration.ofMillis(100))
+                            .filter(e -> e instanceof OptimisticLockingFailureException))
+                        .doOnSuccess(c -> log.info("[FalsePositive] Comment {} approved", commentName))
+                        .then();
+                }
+                return Mono.empty();
+            });
+    }
 }
@@ -33,6 +33,9 @@ public class AiCommentReply extends AbstractExtension {
        @Schema(description = "关联文章Slug，用于生成文章链接")
        private String postSlug;

+        @Schema(description = "关联内容类型: Post/SinglePage")
+        private String postKind;
+
        @Schema(description = "AI回复内容")
        private String reply;

@@ -62,5 +65,11 @@ public class AiCommentReply extends AbstractExtension {

        @Schema(description = "已发布的回复名称")
        private String replyName;
+
+        @Schema(description = "前置过滤拦截分类（广告/辱骂攻击/敏感内容/无意义，为空表示未被拦截）")
+        private String filterCategory;
+
+        @Schema(description = "前置过滤拦截原因详情（为空表示未被拦截）")
+        private String filterReason;
    }
 }
@@ -34,6 +34,16 @@ public class AiPersona extends AbstractExtension {
        @Schema(description = "邮箱（用于Gravatar头像）")
        private String email;

+        @Schema(description = "角色性别（male/female）")
+        private String gender;
+
+        @Schema(description = "是否使用中性语气，默认false即跟随性别语气")
+        @JsonProperty("neutralVoice")
+        private Boolean neutralVoice;
+
+        @Schema(description = "唤醒词，评论以此开头则唤醒该角色回复，留空则不启用唤醒")
+        private String wakeWord;
+
        @Schema(description = "是否为默认角色")
        @JsonProperty("isDefault")
        private Boolean isDefault;
@@ -1,31 +0,0 @@
-package top.nxxy335.commentaiautopilot.listener;
-
-import lombok.RequiredArgsConstructor;
-import lombok.extern.slf4j.Slf4j;
-import org.springframework.stereotype.Component;
-import run.halo.app.extension.ReactiveExtensionClient;
-import run.halo.app.extension.controller.Controller;
-import run.halo.app.extension.controller.ControllerBuilder;
-import run.halo.app.extension.controller.Reconciler;
-import top.nxxy335.commentaiautopilot.extension.AiPersona;
-
-@Component
-@Slf4j
-@RequiredArgsConstructor
-public class AiPersonaReconciler implements Reconciler<Reconciler.Request> {
-
-    private final ReactiveExtensionClient client;
-
-    @Override
-    public Result reconcile(Request request) {
-        return new Result(false, null);
-    }
-
-    @Override
-    public Controller setupWith(ControllerBuilder builder) {
-        return builder
-            .extension(new AiPersona())
-            .syncAllOnStart(false)
-            .build();
-    }
-}
@@ -4,22 +4,19 @@ import lombok.RequiredArgsConstructor;
 import lombok.extern.slf4j.Slf4j;
 import org.springframework.stereotype.Component;
 import reactor.core.scheduler.Schedulers;
-import run.halo.app.core.extension.content.Category;
 import run.halo.app.core.extension.content.Comment;
-import run.halo.app.core.extension.content.Post;
-import run.halo.app.core.extension.content.Tag;
 import run.halo.app.extension.ExtensionClient;
 import run.halo.app.extension.controller.Controller;
 import run.halo.app.extension.controller.ControllerBuilder;
 import run.halo.app.extension.controller.Reconciler;
 import top.nxxy335.commentaiautopilot.extension.AiCommentReply;
 import top.nxxy335.commentaiautopilot.service.AiReplyOrchestrator;
+import top.nxxy335.commentaiautopilot.service.PersonaResolver;
+import top.nxxy335.commentaiautopilot.service.WakeWordService;

 import java.time.Instant;
 import java.util.HashMap;
 import java.util.Map;
-import java.util.concurrent.ConcurrentHashMap;
-import java.util.concurrent.atomic.AtomicBoolean;

@Component
@Slf4j
@@ -28,31 +25,20 @@ public class CommentReconciler implements Reconciler<Reconciler.Request> {

    private final ExtensionClient client;
    private final AiReplyOrchestrator orchestrator;
+    private final PersonaResolver personaResolver;
+    private final WakeWordService wakeWordService;

    private static final String PROCESSED_ANNOTATION = "comment-ai-autopilot.nxxy335.top/processed";
-    private static final String AI_MARKER_ANNOTATION = "comment-ai-autopilot.nxxy335.top/is-ai";
    private static final String AI_PERSONA_OWNER_PREFIX = "ai-persona-";
-    private static final String AI_PERSONA_ANNOTATION = "comment-ai-autopilot.nxxy335.top/ai-persona";
+    private static final String AI_MARKER_ANNOTATION = "comment-ai-autopilot.nxxy335.top/is-ai";

    // Record the time when this bean was created (plugin startup time)
    private final Instant pluginStartTime = Instant.now();

-    // In-memory dedup lock: prevents the same comment from being processed multiple times
-    // even if reconcile is triggered concurrently
-    private final ConcurrentHashMap<String, Boolean> processingLocks = new ConcurrentHashMap<>();
-
    @Override
    public Result reconcile(Request request) {
        var name = request.name();

-        // Acquire lock at the very beginning to prevent any concurrent processing
-        if (processingLocks.putIfAbsent(name, Boolean.TRUE) != null) {
-            log.debug("[CommentReconciler] Already processing comment: {}, skipping", name);
-            return Result.doNotRetry();
-        }
-
-        AtomicBoolean asyncStarted = new AtomicBoolean(false);
-        try {
        client.fetch(Comment.class, name).ifPresent(comment -> {
            if (isProcessed(comment.getMetadata().getAnnotations())) {
                return;
@@ -92,33 +78,39 @@ public class CommentReconciler implements Reconciler<Reconciler.Request> {
            markProcessed(comment);
            client.update(comment);

-                // Read persona name from the post's annotations
-                String personaName = getPersonaNameFromComment(comment);
+            // Check for wake word in comment content
+            String commentContent = getCommentContent(comment);
+            log.info("[CommentReconciler] Wake word check for comment {}: content='{}'",
+                name, commentContent.length() > 80 ? commentContent.substring(0, 80) + "..." : commentContent);
+            var wakeMatch = wakeWordService.checkWakeWordBlocking(client, commentContent);
+
+            if (wakeMatch != null) {
+                // Wake word matched: trigger AI reply with the matched persona,
+                // bypassing normal page-level enable check
+                log.info("[CommentReconciler] Wake word '{}' matched for persona '{}', triggering reply for: {}",
+                    wakeMatch.wakeWord(), wakeMatch.personaName(), name);
+                orchestrator.processComment(name, null, false, wakeMatch.personaName(), true)
+                    .subscribeOn(Schedulers.boundedElastic())
+                    .subscribe(
+                        null,
+                        e -> log.error("[CommentReconciler] Error processing wake word comment {}: {}", name, e.getMessage(), e),
+                        () -> log.info("[CommentReconciler] Wake word processing completed for comment: {}", name)
+                    );
+            } else {
+                // Normal flow: read persona name from the post's annotations
+                String personaName = personaResolver.getPersonaNameFromCommentBlocking(client, comment);

                // Top-level comment → always trigger AI reply
                log.info("[CommentReconciler] New top-level comment detected: {}, personaName: {}", name, personaName);
-                asyncStarted.set(true);
-                orchestrator.processComment(name, null, false, personaName)
+                orchestrator.processComment(name, null, false, personaName, false)
                    .subscribeOn(Schedulers.boundedElastic())
-                    .doFinally(signal -> {
-                        processingLocks.remove(name);
-                        log.debug("[CommentReconciler] Released processing lock for: {}", name);
-                    })
                    .subscribe(
                        null,
                        e -> log.error("[CommentReconciler] Error processing comment {}: {}", name, e.getMessage(), e),
                        () -> log.info("[CommentReconciler] Processing completed for comment: {}", name)
                    );
+            }
        });
-        } catch (Exception e) {
-            log.error("[CommentReconciler] Error in reconcile for {}: {}", name, e.getMessage(), e);
-        } finally {
-            // Only release lock here if async processing was NOT started
-            // (async path releases lock in doFinally)
-            if (!asyncStarted.get()) {
-                processingLocks.remove(name);
-            }
-        }

        return Result.doNotRetry();
    }
@@ -139,61 +131,6 @@ public class CommentReconciler implements Reconciler<Reconciler.Request> {
        return false;
    }

-    /**
-     * Read persona name from the post's annotations associated with this comment.
-     */
-    private String getPersonaNameFromComment(Comment comment) {
-        var subjectRef = comment.getSpec().getSubjectRef();
-        if (subjectRef == null || !"Post".equals(subjectRef.getKind())) {
-            return null;
-        }
-        String postName = subjectRef.getName();
-        return client.fetch(Post.class, postName)
-            .map(post -> {
-                // 1. 文章注解优先
-                var annotations = post.getMetadata().getAnnotations();
-                if (annotations != null) {
-                    String persona = annotations.get(AI_PERSONA_ANNOTATION);
-                    if (persona != null && !persona.isBlank()) {
-                        return persona;
-                    }
-                }
-                // 2. 分类注解
-                var spec = post.getSpec();
-                if (spec != null && spec.getCategories() != null) {
-                    for (String categoryName : spec.getCategories()) {
-                        var cat = client.fetch(Category.class, categoryName).orElse(null);
-                        if (cat != null) {
-                            var catAnnotations = cat.getMetadata().getAnnotations();
-                            if (catAnnotations != null) {
-                                String catPersona = catAnnotations.get(AI_PERSONA_ANNOTATION);
-                                if (catPersona != null && !catPersona.isBlank()) {
-                                    return catPersona;
-                                }
-                            }
-                        }
-                    }
-                }
-                // 3. 标签注解
-                if (spec != null && spec.getTags() != null) {
-                    for (String tagName : spec.getTags()) {
-                        var tag = client.fetch(Tag.class, tagName).orElse(null);
-                        if (tag != null) {
-                            var tagAnnotations = tag.getMetadata().getAnnotations();
-                            if (tagAnnotations != null) {
-                                String tagPersona = tagAnnotations.get(AI_PERSONA_ANNOTATION);
-                                if (tagPersona != null && !tagPersona.isBlank()) {
-                                    return tagPersona;
-                                }
-                            }
-                        }
-                    }
-                }
-                return null;
-            })
-            .orElse(null);
-    }
-
    private boolean isProcessed(Map<String, String> annotations) {
        return annotations != null && "true".equals(annotations.get(PROCESSED_ANNOTATION));
    }
@@ -207,6 +144,22 @@ public class CommentReconciler implements Reconciler<Reconciler.Request> {
        annotations.put(PROCESSED_ANNOTATION, "true");
    }

+    private String getCommentContent(Comment comment) {
+        if (comment.getSpec() == null) return "";
+        // Always strip HTML to get plain text for wake word matching
+        String raw = comment.getSpec().getRaw();
+        if (raw != null && !raw.isBlank()) {
+            // raw might still contain HTML in some cases, always strip
+            String plain = org.jsoup.Jsoup.clean(raw, org.jsoup.safety.Safelist.none()).trim();
+            if (!plain.isBlank()) return plain;
+        }
+        String content = comment.getSpec().getContent();
+        if (content != null && !content.isBlank()) {
+            return org.jsoup.Jsoup.clean(content, org.jsoup.safety.Safelist.none()).trim();
+        }
+        return "";
+    }
+
    @Override
    public Controller setupWith(ControllerBuilder builder) {
        return builder
@@ -4,17 +4,16 @@ import lombok.RequiredArgsConstructor;
 import lombok.extern.slf4j.Slf4j;
 import org.springframework.stereotype.Component;
 import reactor.core.scheduler.Schedulers;
-import run.halo.app.core.extension.content.Category;
 import run.halo.app.core.extension.content.Comment;
-import run.halo.app.core.extension.content.Post;
 import run.halo.app.core.extension.content.Reply;
-import run.halo.app.core.extension.content.Tag;
 import run.halo.app.extension.ExtensionClient;
 import run.halo.app.extension.controller.Controller;
 import run.halo.app.extension.controller.ControllerBuilder;
 import run.halo.app.extension.controller.Reconciler;
 import top.nxxy335.commentaiautopilot.extension.AiCommentReply;
 import top.nxxy335.commentaiautopilot.service.AiReplyOrchestrator;
+import top.nxxy335.commentaiautopilot.service.PersonaResolver;
+import top.nxxy335.commentaiautopilot.service.WakeWordService;

 import java.time.Instant;
 import java.util.HashMap;
@@ -27,11 +26,12 @@ public class ReplyReconciler implements Reconciler<Reconciler.Request> {

    private final ExtensionClient client;
    private final AiReplyOrchestrator orchestrator;
+    private final PersonaResolver personaResolver;
+    private final WakeWordService wakeWordService;

    private static final String PROCESSED_ANNOTATION = "comment-ai-autopilot.nxxy335.top/processed";
    private static final String AI_PERSONA_OWNER_PREFIX = "ai-persona-";
    private static final String AI_MARKER_ANNOTATION = "comment-ai-autopilot.nxxy335.top/is-ai";
-    private static final String AI_PERSONA_ANNOTATION = "comment-ai-autopilot.nxxy335.top/ai-persona";

    // Record the time when this bean was created (plugin startup time)
    private final Instant pluginStartTime = Instant.now();
@@ -77,25 +77,21 @@ public class ReplyReconciler implements Reconciler<Reconciler.Request> {
                return;
            }

-            // Check if this reply is specifically replying to an AI reply
-            // by checking the quoteReply field
-            String quoteReply = reply.getSpec().getQuoteReply();
-
-            if (quoteReply == null || quoteReply.isBlank()) {
-                // No quoteReply - this is a direct reply to the top-level comment,
-                // NOT a reply to AI. Skip it (CommentReconciler handles top-level comments).
-                log.debug("[ReplyReconciler] Reply {} has no quoteReply, skipping (not a reply to AI)", name);
-                markProcessed(reply);
-                client.update(reply);
-                return;
-            }
+            // Check for wake word FIRST - wake word can bypass "must be reply to AI" check
+            String replyContent = getReplyContent(reply);
+            log.info("[ReplyReconciler] Wake word check for reply {}: content='{}'",
+                name, replyContent.length() > 80 ? replyContent.substring(0, 80) + "..." : replyContent);
+            var wakeMatch = wakeWordService.checkWakeWordBlocking(client, replyContent);

            // This reply quotes another reply - check if the quoted reply is from AI
-            boolean isReplyToAi = isAiReply(quoteReply);
+            String quoteReply = reply.getSpec().getQuoteReply();
+            boolean isReplyToAi = quoteReply != null && !quoteReply.isBlank() && isAiReply(quoteReply);
            log.debug("[ReplyReconciler] Reply {} quotes {}, isAiReply={}", name, quoteReply, isReplyToAi);

-            if (!isReplyToAi) {
-                log.debug("[ReplyReconciler] Not a reply to AI, skipping: {}", name);
+            // Skip if not a reply to AI AND no wake word matched
+            if (!isReplyToAi && wakeMatch == null) {
+                // No quoteReply or not replying to AI, and no wake word - skip
+                log.debug("[ReplyReconciler] Not a reply to AI and no wake word, skipping: {}", name);
                markProcessed(reply);
                client.update(reply);
                return;
@@ -119,16 +115,32 @@ public class ReplyReconciler implements Reconciler<Reconciler.Request> {
            markProcessed(reply);
            client.update(reply);

-            // Reply to AI → trigger AI reply (conversation continuation)
-            String personaName = getPersonaNameFromComment(parentCommentName);
+            if (wakeMatch != null) {
+                // Wake word matched: trigger AI reply with the matched persona,
+                // bypassing the "must be reply to AI" check and page-level enable check
+                log.info("[ReplyReconciler] Wake word '{}' matched for persona '{}', triggering reply for: {}",
+                    wakeMatch.wakeWord(), wakeMatch.personaName(), name);
+                orchestrator.processComment(parentCommentName, name, true, wakeMatch.personaName(), true)
+                    .subscribeOn(Schedulers.boundedElastic())
+                    .subscribe(
+                        null,
+                        e -> log.error("[ReplyReconciler] Error processing wake word reply {}: {}", name, e.getMessage(), e),
+                        () -> log.info("[ReplyReconciler] Wake word processing completed for reply: {}", name)
+                    );
+            } else if (isReplyToAi) {
+                // Normal flow: reply to AI → trigger AI reply (conversation continuation)
+                String personaName = client.fetch(Comment.class, parentCommentName)
+                    .map(comment -> personaResolver.getPersonaNameFromCommentBlocking(client, comment))
+                    .orElse(null);
                log.info("[ReplyReconciler] Reply to AI detected: {}, triggering conversation, personaName: {}", name, personaName);
-            orchestrator.processComment(parentCommentName, name, true, personaName)
+                orchestrator.processComment(parentCommentName, name, true, personaName, false)
                    .subscribeOn(Schedulers.boundedElastic())
                    .subscribe(
                        null,
                        e -> log.error("[ReplyReconciler] Error processing reply {}: {}", name, e.getMessage(), e),
                        () -> log.info("[ReplyReconciler] Processing completed for reply: {}", name)
                    );
+            }
        });

        return Result.doNotRetry();
@@ -154,65 +166,6 @@ public class ReplyReconciler implements Reconciler<Reconciler.Request> {
            .orElse(false);
    }

-    /**
-     * Read persona name from the post's annotations associated with the parent comment.
-     */
-    private String getPersonaNameFromComment(String commentName) {
-        return client.fetch(Comment.class, commentName)
-            .map(comment -> {
-                var subjectRef = comment.getSpec().getSubjectRef();
-                if (subjectRef == null || !"Post".equals(subjectRef.getKind())) {
-                    return null;
-                }
-                String postName = subjectRef.getName();
-                return client.fetch(Post.class, postName)
-                    .map(post -> {
-                        // 1. 文章注解优先
-                        var annotations = post.getMetadata().getAnnotations();
-                        if (annotations != null) {
-                            String persona = annotations.get(AI_PERSONA_ANNOTATION);
-                            if (persona != null && !persona.isBlank()) {
-                                return persona;
-                            }
-                        }
-                        // 2. 分类注解
-                        var spec = post.getSpec();
-                        if (spec != null && spec.getCategories() != null) {
-                            for (String categoryName : spec.getCategories()) {
-                                var cat = client.fetch(Category.class, categoryName).orElse(null);
-                                if (cat != null) {
-                                    var catAnnotations = cat.getMetadata().getAnnotations();
-                                    if (catAnnotations != null) {
-                                        String catPersona = catAnnotations.get(AI_PERSONA_ANNOTATION);
-                                        if (catPersona != null && !catPersona.isBlank()) {
-                                            return catPersona;
-                                        }
-                                    }
-                                }
-                            }
-                        }
-                        // 3. 标签注解
-                        if (spec != null && spec.getTags() != null) {
-                            for (String tagName : spec.getTags()) {
-                                var tag = client.fetch(Tag.class, tagName).orElse(null);
-                                if (tag != null) {
-                                    var tagAnnotations = tag.getMetadata().getAnnotations();
-                                    if (tagAnnotations != null) {
-                                        String tagPersona = tagAnnotations.get(AI_PERSONA_ANNOTATION);
-                                        if (tagPersona != null && !tagPersona.isBlank()) {
-                                            return tagPersona;
-                                        }
-                                    }
-                                }
-                            }
-                        }
-                        return null;
-                    })
-                    .orElse(null);
-            })
-            .orElse(null);
-    }
-
    private boolean isProcessed(Map<String, String> annotations) {
        return annotations != null && "true".equals(annotations.get(PROCESSED_ANNOTATION));
    }
@@ -226,6 +179,22 @@ public class ReplyReconciler implements Reconciler<Reconciler.Request> {
        annotations.put(PROCESSED_ANNOTATION, "true");
    }

+    private String getReplyContent(Reply reply) {
+        if (reply.getSpec() == null) return "";
+        // Always strip HTML to get plain text for wake word matching
+        String raw = reply.getSpec().getRaw();
+        if (raw != null && !raw.isBlank()) {
+            // raw might still contain HTML in some cases, always strip
+            String plain = org.jsoup.Jsoup.clean(raw, org.jsoup.safety.Safelist.none()).trim();
+            if (!plain.isBlank()) return plain;
+        }
+        String content = reply.getSpec().getContent();
+        if (content != null && !content.isBlank()) {
+            return org.jsoup.Jsoup.clean(content, org.jsoup.safety.Safelist.none()).trim();
+        }
+        return "";
+    }
+
    @Override
    public Controller setupWith(ControllerBuilder builder) {
        return builder
@@ -1,219 +1,101 @@
 package top.nxxy335.commentaiautopilot.service;

 import lombok.extern.slf4j.Slf4j;
-import org.springframework.context.ApplicationContext;
 import org.springframework.stereotype.Component;
 import reactor.core.publisher.Mono;
-import run.halo.app.core.extension.Plugin;
-import run.halo.app.extension.ReactiveExtensionClient;
+import run.halo.app.plugin.extensionpoint.ExtensionGetter;

-import java.lang.reflect.Method;
-import java.util.Map;
+import java.util.List;

 /**
- * AI Foundation client that uses runtime class loading and reflection
- * to call the AI Foundation plugin's AiModelService.
- * <p>
- * This approach avoids classloader identity issues by loading AiModelService
- * from ai-foundation's own classloader, so that Spring's getBeansOfType()
- * can correctly match the implementation bean.
- * <p>
- * No @ConditionalOnClass or pluginDependencies needed.
- * Always registered as a bean; availability is checked at runtime.
+ * AI Foundation 客户端，通过 Halo 的 {@link ExtensionGetter} 获取 AI 服务。
+ *
+ * <p>此类不直接引用任何 AI Foundation API 类（AiModelService、GenerateTextRequest 等），
+ * 所有 AI Foundation 交互委托给 {@link AiFoundationDelegate}。
+ * 当 AI Foundation 插件未安装时，{@link AiFoundationDelegate} 的类加载会触发
+ * {@link NoClassDefFoundError}，在 {@code Mono.defer()} 中被捕获，
+ * 保证插件在无 AI Foundation 环境下仍可正常启动。
+ *
+ * <p>需要在 plugin.yaml 中声明可选依赖：
+ * <pre>
+ * spec:
+ *   pluginDependencies:
+ *     ai-foundation?: "*"
+ * </pre>
 */
@Slf4j
@Component
 public class AiFoundationClient {

-    private static final String AI_FOUNDATION_PLUGIN_NAME = "ai-foundation";
-    private static final String AI_MODEL_SERVICE_CLASS = "run.halo.aifoundation.AiModelService";
+    private final ExtensionGetter extensionGetter;

-    private final ReactiveExtensionClient client;
-    private final ApplicationContext applicationContext;
-
-    public AiFoundationClient(ReactiveExtensionClient client, ApplicationContext applicationContext) {
-        this.client = client;
-        this.applicationContext = applicationContext;
+    public AiFoundationClient(ExtensionGetter extensionGetter) {
+        this.extensionGetter = extensionGetter;
    }

    /**
-     * Call AI Foundation to generate a chat response using the specified model.
+     * 调用 AI Foundation 生成聊天回复。
     *
-     * @param prompt    the prompt text
-     * @param modelName the AiModel metadata.name, null or blank to use default model
-     * @return the generated text, or empty if AI Foundation is unavailable
+     * @param prompt    提示词文本
+     * @param modelName AiModel metadata.name，null 或空则使用默认模型
+     * @return 生成的文本，AI Foundation 不可用时返回 empty
     */
    public Mono<String> chat(String prompt, String modelName) {
-        return isAiFoundationEnabled()
-            .flatMap(enabled -> {
-                if (!enabled) {
-                    log.warn("AI Foundation plugin is not installed or not enabled, skipping AI reply");
+        return Mono.defer(() -> {
+            try {
+                return AiFoundationDelegate.chat(extensionGetter, prompt, modelName);
+            } catch (NoClassDefFoundError e) {
+                log.debug("AI Foundation API not on classpath: {}", e.getMessage());
                return Mono.empty();
            }
-                return doChat(prompt, modelName);
-            });
-    }
-
-    /**
-     * Check if AI Foundation is available: plugin installed, enabled, and AiModelService bean found.
-     */
-    public Mono<Boolean> isAvailable() {
-        return isAiFoundationEnabled()
-            .flatMap(enabled -> {
-                if (!enabled) return Mono.just(false);
-                return findAiModelService().hasElement();
-            });
-    }
-
-    private Mono<Boolean> isAiFoundationEnabled() {
-        return client.fetch(Plugin.class, AI_FOUNDATION_PLUGIN_NAME)
-            .map(plugin -> plugin.getSpec().getEnabled())
-            .defaultIfEmpty(false)
-            .onErrorResume(e -> {
-                log.debug("Failed to check AI Foundation plugin status: {}", e.getMessage());
-                return Mono.just(false);
-            });
-    }
-
-    private Mono<String> doChat(String prompt, String modelName) {
-        return findAiModelService()
-            .flatMap(service -> invokeLanguageModel(service, modelName)
-                .flatMap(model -> invokeGenerateText(model, prompt))
-            )
-            .doOnError(e -> log.error("AI Foundation call failed: {}", e.getMessage()))
-            .onErrorResume(e -> {
+        })
+        .onErrorResume(NoClassDefFoundError.class, e -> {
            log.warn("AI Foundation not available: {}", e.getMessage());
            return Mono.empty();
        });
    }

    /**
-     * Get PluginManager via the pluginWrapper bean registered in our plugin context.
-     * Halo's DefaultPluginApplicationContextFactory registers pluginWrapper as a singleton:
-     *   beanFactory.registerSingleton("pluginWrapper", pluginWrapper);
-     * Then PluginWrapper.getPluginManager() gives us the PluginManager instance.
+     * 调用 AI Foundation 进行文本分类，使用结构化输出（OutputSpec.choice）。
+     *
+     * @param systemPrompt 系统提示词
+     * @param userPrompt   待分类的用户输入
+     * @param choices      允许的分类值列表
+     * @param modelName    AiModel metadata.name，null 或空则使用默认模型
+     * @return 选中的分类字符串，AI Foundation 不可用时返回 empty
     */
-    private Object findPluginManager() {
+    public Mono<String> classify(String systemPrompt, String userPrompt,
+                                  List<String> choices, String modelName) {
+        return Mono.defer(() -> {
            try {
-            Object pluginWrapper = applicationContext.getBean("pluginWrapper");
-            Method getPluginManagerMethod = pluginWrapper.getClass().getMethod("getPluginManager");
-            getPluginManagerMethod.setAccessible(true);
-            Object pm = getPluginManagerMethod.invoke(pluginWrapper);
-            if (pm != null) {
-                log.info("Found PluginManager via pluginWrapper bean: {}", pm.getClass().getName());
+                return AiFoundationDelegate.classify(extensionGetter, systemPrompt, userPrompt, choices, modelName);
+            } catch (NoClassDefFoundError e) {
+                log.warn("[Client] AI Foundation API not on classpath (classify): {}", e.getMessage());
+                return Mono.empty();
            }
-            return pm;
-        } catch (NoSuchMethodException e) {
-            log.warn("pluginWrapper does not have getPluginManager() method: {}", e.getMessage());
-        } catch (Exception e) {
-            log.warn("Failed to get PluginManager via pluginWrapper: {}", e.getMessage());
-        }
-        log.warn("PluginManager not found");
-        return null;
-    }
-
-    /**
-     * Find the AiModelService bean from ai-foundation's PluginApplicationContext.
-     * Uses PluginManager.getPlugin() to get the plugin wrapper, then reflection
-     * to get the plugin's ApplicationContext.
-     */
-    private Mono<Object> findAiModelService() {
-        return Mono.fromCallable(() -> {
-            Object pm = findPluginManager();
-            if (pm == null) return null;
-
-            // Call pm.getPlugin("ai-foundation") via reflection
-            Method getPluginMethod = pm.getClass().getMethod("getPlugin", String.class);
-            getPluginMethod.setAccessible(true);
-            Object pluginWrapper = getPluginMethod.invoke(pm, AI_FOUNDATION_PLUGIN_NAME);
-            if (pluginWrapper == null) {
-                log.debug("ai-foundation plugin not found in PluginManager");
-                return null;
-            }
-
-            // Call pluginWrapper.getPlugin() to get the plugin instance
-            Method getPluginInstanceMethod = pluginWrapper.getClass().getMethod("getPlugin");
-            getPluginInstanceMethod.setAccessible(true);
-            Object pluginInstance = getPluginInstanceMethod.invoke(pluginWrapper);
-            if (pluginInstance == null) {
-                log.debug("ai-foundation plugin instance is null");
-                return null;
-            }
-
-            // Get the plugin's ApplicationContext via reflection on SpringPlugin
-            // DefaultSpringPlugin is package-private, so we need setAccessible
-            Method getCtxMethod = pluginInstance.getClass().getMethod("getApplicationContext");
-            getCtxMethod.setAccessible(true);
-            ApplicationContext pluginAppContext = (ApplicationContext) getCtxMethod.invoke(pluginInstance);
-
-            // Get the plugin classloader
-            Method getClassLoaderMethod = pluginWrapper.getClass().getMethod("getPluginClassLoader");
-            getClassLoaderMethod.setAccessible(true);
-            ClassLoader pluginClassLoader = (ClassLoader) getClassLoaderMethod.invoke(pluginWrapper);
-
-            // Load AiModelService from ai-foundation's classloader
-            Class<?> aiModelServiceClass = pluginClassLoader.loadClass(AI_MODEL_SERVICE_CLASS);
-
-            // Find the AiModelService bean in ai-foundation's ApplicationContext
-            Map<String, ?> beans = pluginAppContext.getBeansOfType(aiModelServiceClass);
-            if (beans.isEmpty()) {
-                log.debug("AiModelService bean not found in ai-foundation's ApplicationContext");
-                return null;
-            }
-
-            log.info("Found AiModelService bean in ai-foundation's ApplicationContext");
-            Object result = beans.values().iterator().next();
-            return (Object) result;
-        }).doOnError(e -> log.error("Failed to find AiModelService: {}", e.getMessage()));
-    }
-
-    /**
-     * Call service.languageModel(modelName) or service.languageModel() via reflection.
-     * Returns Mono&lt;LanguageModel&gt; from ai-foundation's classloader.
-     */
-    private Mono<Object> invokeLanguageModel(Object service, String modelName) {
-        return Mono.fromCallable(() -> {
-            Method method;
-            if (modelName != null && !modelName.isBlank()) {
-                method = service.getClass().getMethod("languageModel", String.class);
-                method.setAccessible(true);
-                return method.invoke(service, modelName);
-            } else {
-                method = service.getClass().getMethod("languageModel");
-                method.setAccessible(true);
-                return method.invoke(service);
-            }
-        }).flatMap(result -> {
-            if (result instanceof Mono<?> mono) return mono;
-            return Mono.justOrEmpty(result);
+        })
+        .onErrorResume(NoClassDefFoundError.class, e -> {
+            log.warn("[Client] AI Foundation NoClassDefFoundError during classify: {}", e.getMessage());
+            return Mono.empty();
        });
    }

    /**
-     * Call model.generateText(prompt) via reflection, then extract text from result.
-     * Returns the generated text string.
+     * 检查 AI Foundation 是否可用（插件已安装且 AiModelService 扩展已启用）。
     */
-    private Mono<String> invokeGenerateText(Object model, String prompt) {
-        return Mono.fromCallable(() -> {
-            Method method = model.getClass().getMethod("generateText", String.class);
-            method.setAccessible(true);
-            return method.invoke(model, prompt);
-        }).flatMap(result -> {
-            if (result instanceof Mono<?> mono) {
-                return mono.map(this::extractText);
-            }
-            return Mono.justOrEmpty(extractText(result));
-        });
-    }
-
-    private String extractText(Object result) {
-        if (result == null) return null;
+    public Mono<Boolean> isAvailable() {
+        return Mono.defer(() -> {
            try {
-            Method getText = result.getClass().getMethod("getText");
-            getText.setAccessible(true);
-            return (String) getText.invoke(result);
-        } catch (Exception e) {
-            throw new RuntimeException("Failed to call getText() on GenerateTextResult: " + e.getMessage(), e);
+                return AiFoundationDelegate.isAvailable(extensionGetter);
+            } catch (NoClassDefFoundError e) {
+                log.debug("AI Foundation API not on classpath: {}", e.getMessage());
+                return Mono.just(false);
            }
+        })
+        .onErrorResume(NoClassDefFoundError.class, e -> Mono.just(false))
+        .onErrorResume(e -> {
+            log.debug("AI Foundation not available: {}", e.getMessage());
+            return Mono.just(false);
+        });
    }
 }
@@ -0,0 +1,164 @@
+package top.nxxy335.commentaiautopilot.service;
+
+import lombok.extern.slf4j.Slf4j;
+import reactor.core.publisher.Mono;
+import run.halo.aifoundation.AiModelService;
+import run.halo.aifoundation.chat.GenerateTextRequest;
+import run.halo.aifoundation.chat.GenerateTextResult;
+import run.halo.aifoundation.schema.OutputSpec;
+import run.halo.app.plugin.extensionpoint.ExtensionGetter;
+
+import java.util.List;
+
+/**
+ * AI Foundation API 隔离层。
+ *
+ * <p>此类集中了所有对 AI Foundation 插件 API 的直接引用（AiModelService、
+ * GenerateTextRequest、GenerateTextResult、OutputSpec）。
+ *
+ * <p>关键设计：此类不是 Spring 组件，由 {@link AiFoundationClient} 通过
+ * {@code Mono.defer()} 懒加载调用。当 AI Foundation 插件未安装时，
+ * JVM 加载此类会触发 NoClassDefFoundError，该错误在
+ * {@code AiFoundationClient} 的 defer + try-catch 中被捕获，
+ * 从而保证插件在无 AI Foundation 的环境下仍可正常启动。
+ */
+@Slf4j
+class AiFoundationDelegate {
+
+    private AiFoundationDelegate() {}
+
+    static Mono<String> chat(ExtensionGetter extensionGetter, String prompt, String modelName) {
+        return extensionGetter.getEnabledExtension(AiModelService.class)
+            .flatMap(service -> service.languageModel(modelName != null ? modelName : "")
+                .flatMap(model -> model.generateText(
+                    GenerateTextRequest.builder().prompt(prompt).maxRetries(2).build()))
+                .map(GenerateTextResult::getText))
+            .doOnError(e -> log.error("[Delegate] chat call failed: {}", e.getMessage()))
+            .onErrorResume(e -> {
+                log.warn("[Delegate] chat not available: {}", e.getMessage());
+                return Mono.empty();
+            });
+    }
+
+    /**
+     * 使用 AI 进行文本分类。
+     * 优先使用 OutputSpec.choice 结构化输出，失败时退回到普通 chat 并解析响应。
+     */
+    static Mono<String> classify(ExtensionGetter extensionGetter, String systemPrompt,
+                                  String userPrompt, List<String> choices, String modelName) {
+        log.info("[Delegate] Starting classification, modelName='{}'", modelName);
+        return classifyWithChoice(extensionGetter, systemPrompt, userPrompt, choices, modelName)
+            .switchIfEmpty(
+                Mono.defer(() -> {
+                    log.info("[Delegate] classifyWithChoice returned empty, falling back to classifyWithChat");
+                    return classifyWithChat(extensionGetter, systemPrompt, userPrompt, choices, modelName);
+                })
+            )
+            .doOnNext(result -> log.info("[Delegate] Classification succeeded: '{}'", result))
+            .doOnSuccess(result -> {
+                if (result == null) {
+                    log.warn("[Delegate] Classification completed with no result (both methods returned empty)");
+                }
+            });
+    }
+
+    /**
+     * 使用 OutputSpec.choice 结构化输出分类（部分模型不支持）。
+     * 注意：不使用 system() 方法，因为部分 AI Foundation 版本可能不支持，
+     * 将 system prompt 合并到 user prompt 中。
+     */
+    private static Mono<String> classifyWithChoice(ExtensionGetter extensionGetter, String systemPrompt,
+                                                    String userPrompt, List<String> choices, String modelName) {
+        // 合并 system prompt 和 user prompt，避免使用 system() 方法
+        String combinedPrompt = systemPrompt + "\n\n" + userPrompt;
+        return extensionGetter.getEnabledExtension(AiModelService.class)
+            .flatMap(service -> service.languageModel(modelName != null ? modelName : "")
+                .flatMap(model -> model.generateText(
+                    GenerateTextRequest.builder()
+                        .prompt(combinedPrompt)
+                        .output(OutputSpec.choice(choices))
+                        .maxRetries(2)
+                        .build()))
+                .flatMap(result -> {
+                    Object output = result.getOutput();
+                    if (output != null) {
+                        String value = String.valueOf(output).trim();
+                        if (!value.isEmpty()) {
+                            log.debug("[Delegate] classifyWithChoice got output: '{}'", value);
+                            return Mono.just(value);
+                        }
+                    }
+                    // output 为空可能是模型不支持结构化输出，返回 empty 触发 fallback
+                    log.info("[Delegate] classifyWithChoice: output is null/empty, triggering fallback");
+                    return Mono.empty();
+                }))
+            .onErrorResume(e -> {
+                log.warn("[Delegate] classifyWithChoice failed, will fallback to chat: {}", e.getMessage());
+                return Mono.empty();
+            });
+    }
+
+    /**
+     * 使用普通 chat 调用进行分类，从响应文本中提取匹配的分类值。
+     * 作为 OutputSpec.choice 不可用时的降级方案。
+     * 不使用 system() 方法，将 system prompt 合并到 user prompt 中，
+     * 与可用的 chat() 方法保持一致的调用方式。
+     */
+    private static Mono<String> classifyWithChat(ExtensionGetter extensionGetter, String systemPrompt,
+                                                  String userPrompt, List<String> choices, String modelName) {
+        // 合并 system prompt 和 user prompt，与 chat() 方法保持一致的调用方式
+        String combinedPrompt = systemPrompt + "\n\n" + userPrompt;
+        return extensionGetter.getEnabledExtension(AiModelService.class)
+            .flatMap(service -> service.languageModel(modelName != null ? modelName : "")
+                .flatMap(model -> model.generateText(
+                    GenerateTextRequest.builder()
+                        .prompt(combinedPrompt)
+                        .maxRetries(2)
+                        .build()))
+                .map(GenerateTextResult::getText)
+                .map(text -> extractChoice(text, choices)))
+            .doOnError(e -> log.error("[Delegate] classifyWithChat failed: {}", e.getMessage()))
+            .onErrorResume(e -> {
+                log.warn("[Delegate] classifyWithChat error: {}", e.getMessage());
+                return Mono.empty();
+            });
+    }
+
+    /**
+     * 从 chat 响应文本中提取匹配的分类值。
+     * 优先精确匹配，其次包含匹配。
+     * 包含匹配时优先匹配违规类别（广告/辱骂/敏感/无意义），最后才匹配"正常"，
+     * 避免 AI 解释性文本中同时出现"正常"和违规词时误判为"正常"。
+     * 无匹配时返回空字符串（触发 defaultIfEmpty 安全拦截），避免原始文本被误判为违规类别。
+     */
+    static String extractChoice(String text, List<String> choices) {
+        if (text == null || text.isBlank()) return "";
+        String trimmed = text.trim();
+        // 精确匹配
+        for (String choice : choices) {
+            if (trimmed.equals(choice)) return choice;
+        }
+        // 包含匹配：先匹配违规类别，最后匹配"正常"
+        // 避免"该评论属于广告，不是正常评论"被误匹配为"正常"
+        for (String choice : choices) {
+            if ("正常".equals(choice)) continue;
+            if (trimmed.contains(choice)) return choice;
+        }
+        // 最后检查"正常"
+        for (String choice : choices) {
+            if ("正常".equals(choice) && trimmed.contains(choice)) return choice;
+        }
+        // 无匹配，返回空字符串触发安全拦截
+        log.warn("[Delegate] No matching choice found in response: '{}', returning empty for safety", trimmed);
+        return "";
+    }
+
+    static Mono<Boolean> isAvailable(ExtensionGetter extensionGetter) {
+        return extensionGetter.getEnabledExtension(AiModelService.class)
+            .hasElement()
+            .onErrorResume(e -> {
+                log.debug("[Delegate] AI Foundation not available: {}", e.getMessage());
+                return Mono.just(false);
+            });
+    }
+}
@@ -17,6 +17,9 @@ import java.util.concurrent.Executors;
 import java.util.concurrent.ScheduledExecutorService;
 import java.util.concurrent.TimeUnit;

+import reactor.core.publisher.Mono;
+import reactor.core.publisher.Flux;
+
@Component
@Slf4j
 public class AiReplyCleanupService implements DisposableBean {
@@ -27,9 +30,9 @@ public class AiReplyCleanupService implements DisposableBean {

    private static final String CONFIG_MAP_NAME = "comment-ai-autopilot-configmap";

-    public AiReplyCleanupService(ReactiveExtensionClient client) {
+    public AiReplyCleanupService(ReactiveExtensionClient client, ObjectMapper objectMapper) {
        this.client = client;
-        this.objectMapper = new ObjectMapper();
+        this.objectMapper = objectMapper;
        this.scheduler = Executors.newSingleThreadScheduledExecutor(r -> {
            Thread t = new Thread(r, "ai-reply-cleanup");
            t.setDaemon(true);
@@ -40,8 +43,25 @@ public class AiReplyCleanupService implements DisposableBean {
    }

    public void dailyCleanup() {
-        try {
-            Boolean enabled = client.fetch(ConfigMap.class, CONFIG_MAP_NAME)
+        isCleanupEnabled()
+            .flatMap(enabled -> {
+                if (!Boolean.TRUE.equals(enabled)) {
+                    log.debug("[Cleanup] Auto cleanup is disabled, skipping");
+                    return Mono.empty();
+                }
+                return getRetentionDays()
+                    .flatMap(retentionDays -> executeCleanup(retentionDays)
+                        .doOnNext(deleted -> log.info("[Cleanup] Auto cleanup completed, deleted {} records older than {} days", deleted, retentionDays))
+                    );
+            })
+            .subscribe(
+                null,
+                e -> log.error("[Cleanup] Error during daily cleanup: {}", e.getMessage(), e)
+            );
+    }
+
+    private Mono<Boolean> isCleanupEnabled() {
+        return client.fetch(ConfigMap.class, CONFIG_MAP_NAME)
            .mapNotNull(cm -> {
                var data = cm.getData();
                if (data == null) return false;
@@ -55,51 +75,35 @@ public class AiReplyCleanupService implements DisposableBean {
                    return true;
                }
            })
-                .defaultIfEmpty(true)
-                .block();
-
-            if (!Boolean.TRUE.equals(enabled)) {
-                log.debug("[Cleanup] Auto cleanup is disabled, skipping");
-                return;
+            .defaultIfEmpty(true);
    }

-            int retentionDays = getRetentionDays();
-            long deleted = executeCleanup(retentionDays);
-            log.info("[Cleanup] Auto cleanup completed, deleted {} records older than {} days", deleted, retentionDays);
-        } catch (Exception e) {
-            log.error("[Cleanup] Error during daily cleanup: {}", e.getMessage(), e);
-        }
-    }
-
-    public long executeCleanup(int retentionDays) {
+    public Mono<Long> executeCleanup(int retentionDays) {
        Instant cutoff = Instant.now().minus(retentionDays, ChronoUnit.DAYS);

-        var oldRecords = client.listAll(AiCommentReply.class, ListOptions.builder().build(), Sort.unsorted())
+        return client.listAll(AiCommentReply.class, ListOptions.builder().build(), Sort.unsorted())
            .filter(r -> {
                Instant created = r.getMetadata().getCreationTimestamp();
                return created != null && created.isBefore(cutoff);
            })
            .collectList()
-            .block();
-
-        if (oldRecords == null || oldRecords.isEmpty()) {
-            return 0;
+            .flatMap(oldRecords -> {
+                if (oldRecords.isEmpty()) {
+                    return Mono.just(0L);
                }
-
-        long deleted = 0;
-        for (var record : oldRecords) {
-            try {
-                client.delete(record).block();
-                deleted++;
-            } catch (Exception e) {
+                return Flux.fromIterable(oldRecords)
+                    .flatMap(record -> client.delete(record)
+                        .thenReturn(1L)
+                        .onErrorResume(e -> {
                            log.warn("[Cleanup] Failed to delete record {}: {}", record.getMetadata().getName(), e.getMessage());
-            }
-        }
-        return deleted;
+                            return Mono.just(0L);
+                        })
+                    )
+                    .reduce(0L, Long::sum);
+            });
    }

-    public int getRetentionDays() {
-        try {
+    public Mono<Integer> getRetentionDays() {
        return client.fetch(ConfigMap.class, CONFIG_MAP_NAME)
            .mapNotNull(cm -> {
                var data = cm.getData();
@@ -114,11 +118,10 @@ public class AiReplyCleanupService implements DisposableBean {
                }
            })
            .defaultIfEmpty(30)
-                .block();
-        } catch (Exception e) {
+            .onErrorResume(e -> {
                log.warn("[Cleanup] Failed to read retentionDays config: {}", e.getMessage());
-            return 30;
-        }
+                return Mono.just(30);
+            });
    }

    @Override
@@ -31,6 +31,7 @@ public class AiReplyOrchestrator {
    private final CommentReplyPublisher commentReplyPublisher;
    private final FilterService filterService;
    private final RateLimitService rateLimitService;
+    private final CommentPreFilterService preFilterService;
    private final ReactiveExtensionClient client;
    private final ObjectMapper objectMapper;

@@ -49,7 +50,9 @@ public class AiReplyOrchestrator {
                               CommentReplyPublisher commentReplyPublisher,
                               FilterService filterService,
                               RateLimitService rateLimitService,
-                               ReactiveExtensionClient client) {
+                               CommentPreFilterService preFilterService,
+                               ReactiveExtensionClient client,
+                               ObjectMapper objectMapper) {
        this.contextExtractor = contextExtractor;
        this.promptBuilder = promptBuilder;
        this.aiReplyService = aiReplyService;
@@ -58,8 +61,9 @@ public class AiReplyOrchestrator {
        this.commentReplyPublisher = commentReplyPublisher;
        this.filterService = filterService;
        this.rateLimitService = rateLimitService;
+        this.preFilterService = preFilterService;
        this.client = client;
-        this.objectMapper = new ObjectMapper();
+        this.objectMapper = objectMapper;
    }

    /**
@@ -69,9 +73,10 @@ public class AiReplyOrchestrator {
     * @param replyName       the Reply name that triggered this (null for top-level comments)
     * @param isAiConversation true when someone replied to AI's reply (conversation continuation)
     * @param personaName     the persona name to use (null for default persona)
+     * @param wakeWordTriggered true when triggered by a wake word (bypasses page-level enable check)
     */
    public Mono<Void> processComment(String commentName, String replyName, boolean isAiConversation,
-                                      String personaName) {
+                                      String personaName, boolean wakeWordTriggered) {
        String lockKey = isAiConversation ? commentName + ":conv:" + replyName : commentName + ":top";

        // Clean up stale locks before acquiring new one
@@ -83,12 +88,12 @@ public class AiReplyOrchestrator {
            return Mono.empty();
        }

-        log.info("[Orchestrator] Start processing: comment={}, replyName={}, isAiConversation={}, personaName={}",
-            commentName, replyName, isAiConversation, personaName);
+        log.info("[Orchestrator] Start processing: comment={}, replyName={}, isAiConversation={}, personaName={}, wakeWordTriggered={}",
+            commentName, replyName, isAiConversation, personaName, wakeWordTriggered);

        return isAutoReplyEnabled()
            .flatMap(enabled -> {
-                if (!enabled) {
+                if (!enabled && !wakeWordTriggered) {
                    log.info("[Orchestrator] Auto reply disabled, skipping: {}", commentName);
                    return Mono.empty();
                }
@@ -98,12 +103,81 @@ public class AiReplyOrchestrator {
                            log.info("[Orchestrator] 速率限制，跳过: {}", commentName);
                            return Mono.empty();
                        }
+                        // Wake word triggered: skip page-level annotation check
+                        if (wakeWordTriggered) {
+                            return filterService.isCommenterBlocked(commentName)
+                                .flatMap(blocked -> {
+                                    if (blocked) {
+                                        log.info("[Orchestrator] Commenter blocked, skipping wake word: {}", commentName);
+                                        return Mono.empty();
+                                    }
+                                    return proceedWithProcess(commentName, replyName, isAiConversation, personaName);
+                                });
+                        }
                        return filterService.shouldProcess(commentName)
                            .flatMap(shouldProcess -> {
                                if (!shouldProcess) {
                                    log.info("[Orchestrator] Filtered out by rules: {}", commentName);
                                    return Mono.empty();
                                }
+                                return proceedWithProcess(commentName, replyName, isAiConversation, personaName);
+                            });
+                    });
+            })
+            .doOnError(e -> log.error("[Orchestrator] Error processing comment {}: {}", commentName, e.getMessage(), e))
+            .doFinally(signal -> {
+                // Always release the lock when processing completes
+                processingLocks.remove(lockKey);
+                log.debug("[Orchestrator] Released processing lock for: {}", lockKey);
+            })
+            .then();
+    }
+
+    /**
+     * 误报反馈专用：跳过前置过滤和去重检查，直接为已确认误报的评论生成 AI 回复。
+     *
+     * <p>与 {@link #processComment} 不同，此方法：
+     * <ul>
+     *   <li>跳过前置过滤（用户已确认评论合规）</li>
+     *   <li>跳过去重检查（已有 FILTERED 记录，需复用）</li>
+     *   <li>跳过速率限制和黑名单检查（管理员主动操作）</li>
+     * </ul>
+     *
+     * @param commentName     the parent Comment name
+     * @param replyName       the Reply name (null for top-level comments)
+     * @param isAiConversation true when this is a conversation continuation
+     * @param personaName     the persona name to use
+     * @param recordName      the existing AiCommentReply record name to update
+     */
+    public Mono<Void> processFalsePositive(String commentName, String replyName,
+                                            boolean isAiConversation, String personaName,
+                                            String recordName) {
+        log.info("[Orchestrator] Processing false-positive: comment={}, record={}", commentName, recordName);
+
+        return getModelName().flatMap(modelName ->
+            contextExtractor.extract(commentName, replyName, isAiConversation)
+                .flatMap(context -> {
+                    // Fetch the existing record (was FILTERED, now PENDING)
+                    return client.fetch(AiCommentReply.class, recordName)
+                        .flatMap(replyRecord ->
+                            sentimentService.analyzeSentiment(context.commentContent(), modelName)
+                                .flatMap(sentimentResult ->
+                                    promptBuilder.buildPrompt(context, sentimentResult.sentiment(), personaName)
+                                        .flatMap(prompt -> generateAndPublish(prompt, context, replyRecord, modelName, personaName))
+                                )
+                        );
+                })
+        )
+        .doOnError(e -> log.error("[Orchestrator] Error processing false-positive {}: {}", commentName, e.getMessage(), e))
+        .then();
+    }
+
+    /**
+     * Proceed with processing after all checks have passed.
+     * Handles dedup checks and conversation round limits.
+     */
+    private Mono<Void> proceedWithProcess(String commentName, String replyName,
+                                           boolean isAiConversation, String personaName) {
        // For top-level comments: skip if we already have ANY reply record
        // For AI conversation: skip if we already replied to THIS specific reply
        if (!isAiConversation) {
@@ -134,23 +208,23 @@ public class AiReplyOrchestrator {
                        });
                })
            );
-                            });
-                    });
-            })
-            .doOnError(e -> log.error("[Orchestrator] Error processing comment {}: {}", commentName, e.getMessage(), e))
-            .doFinally(signal -> {
-                // Always release the lock when processing completes
-                processingLocks.remove(lockKey);
-                log.debug("[Orchestrator] Released processing lock for: {}", lockKey);
-            })
-            .then();
    }

    private Mono<Void> doProcess(String commentName, String replyName, boolean isAiConversation,
                                  String personaName) {
        return getModelName().flatMap(modelName ->
            contextExtractor.extract(commentName, replyName, isAiConversation)
-                .flatMap(context -> sentimentService.analyzeSentiment(context.commentContent(), modelName)
+                .flatMap(context -> preFilterService.check(context.commentContent(), modelName)
+                    .flatMap(preFilterResult -> {
+                        if (!preFilterResult.passed()) {
+                            log.warn("[Orchestrator] Comment pre-filtered: {}, reason: {}",
+                                commentName, preFilterResult.reason());
+                            // 创建拦截记录并执行处罚（针对实际违规的 Comment 或 Reply）
+                            return createFilteredRecord(context, preFilterResult)
+                                .then(preFilterService.penalize(commentName, replyName))
+                                .then();
+                        }
+                        return sentimentService.analyzeSentiment(context.commentContent(), modelName)
                            .flatMap(sentimentResult -> {
                                log.info("[Orchestrator] Sentiment for {}: {} (confidence: {})",
                                    commentName, sentimentResult.sentiment(), sentimentResult.confidence());
@@ -158,6 +232,7 @@ public class AiReplyOrchestrator {
                                    .flatMap(prompt -> createAiCommentReply(context, sentimentResult.sentiment(), personaName)
                                        .flatMap(replyRecord -> generateAndPublish(prompt, context, replyRecord, modelName, personaName))
                                    );
+                            });
                    })
                )
        );
@@ -231,12 +306,13 @@ public class AiReplyOrchestrator {
                        }
                        return publishReply(context, aiReply, replyRecord, reviewResult.score(), personaName);
                    })
-                    .switchIfEmpty(
-                        publishReply(context, aiReply, replyRecord, 100, personaName)
-                    )
                    .onErrorResume(e -> {
-                        log.warn("[Orchestrator] Review error, auto-passing: {}", e.getMessage());
-                        return publishReply(context, aiReply, replyRecord, 100, personaName);
+                        // review() already handles errors internally (returns PASS),
+                        // so this only fires for errors from publishReply/updateRecord.
+                        // Do NOT re-call publishReply to avoid double-publish / overwriting published=false.
+                        log.error("[Orchestrator] Error during review/publish for {}: {}",
+                            context.commentId(), e.getMessage(), e);
+                        return Mono.empty();
                    });
            });
    }
@@ -265,7 +341,7 @@ public class AiReplyOrchestrator {
                // Update retryCount and reset status to PENDING
                return updateRecordForRetry(replyRecord, newRetryCount)
                    .delayElement(Duration.ofSeconds(delaySeconds))
-                    .then(retryGenerate(context, replyRecord, modelName, personaName));
+                    .flatMap(updated -> retryGenerate(context, updated, modelName, personaName));
            } else {
                log.warn("[Orchestrator] Max retry count ({}) exceeded for: {}, marking as FAIL. Reason: {}",
                    maxRetry, context.commentId(), reason);
@@ -514,6 +590,34 @@ public class AiReplyOrchestrator {
            .defaultIfEmpty(10);
    }

+    /**
+     * 创建被前置过滤拦截的记录。
+     */
+    private Mono<AiCommentReply> createFilteredRecord(ContextExtractor.CommentContext context,
+                                                        CommentPreFilterService.PreFilterResult preFilterResult) {
+        AiCommentReply record = new AiCommentReply();
+        record.setMetadata(new Metadata());
+        record.getMetadata().setName("ai-reply-" + UUID.randomUUID().toString().substring(0, 8));
+        record.setSpec(new AiCommentReply.Spec());
+        record.getSpec().setCommentId(context.commentId());
+        record.getSpec().setPostId(context.postId());
+        record.getSpec().setPostSlug(context.postSlug());
+        record.getSpec().setPostKind(context.postKind());
+        record.getSpec().setReply("");
+        record.getSpec().setScore(0);
+        record.getSpec().setStatus("FILTERED");
+        record.getSpec().setRetryCount(0);
+        record.getSpec().setReplyTo(context.replyTo());
+        record.getSpec().setIsAiConversation(context.isAiConversation());
+        record.getSpec().setPublished(false);
+        record.getSpec().setSentiment("NEUTRAL");
+        record.getSpec().setFilterCategory(preFilterResult.category());
+        record.getSpec().setFilterReason(preFilterResult.reason());
+        return client.create(record)
+            .doOnSuccess(created -> log.info("[Orchestrator] Created filtered record: {} category={} reason={}",
+                created.getMetadata().getName(), preFilterResult.category(), preFilterResult.reason()));
+    }
+
    private Mono<AiCommentReply> createAiCommentReply(ContextExtractor.CommentContext context, String sentiment,
                                                        String personaName) {
        AiCommentReply record = new AiCommentReply();
@@ -523,6 +627,7 @@ public class AiReplyOrchestrator {
        record.getSpec().setCommentId(context.commentId());
        record.getSpec().setPostId(context.postId());
        record.getSpec().setPostSlug(context.postSlug());
+        record.getSpec().setPostKind(context.postKind());
        record.getSpec().setReply("");
        record.getSpec().setScore(0);
        record.getSpec().setStatus("PENDING");
@@ -0,0 +1,267 @@
+package top.nxxy335.commentaiautopilot.service;
+
+import com.fasterxml.jackson.databind.JsonNode;
+import com.fasterxml.jackson.databind.ObjectMapper;
+import lombok.extern.slf4j.Slf4j;
+import org.jsoup.Jsoup;
+import org.jsoup.safety.Safelist;
+import org.springframework.dao.OptimisticLockingFailureException;
+import org.springframework.stereotype.Component;
+import reactor.core.publisher.Mono;
+import reactor.util.retry.Retry;
+import run.halo.app.core.extension.content.Comment;
+import run.halo.app.core.extension.content.Reply;
+import run.halo.app.extension.ConfigMap;
+import run.halo.app.extension.ReactiveExtensionClient;
+
+import java.time.Duration;
+import java.time.Instant;
+import java.util.List;
+import java.util.Map;
+
+/**
+ * 评论前置过滤服务：在 AI 回复之前检测评论合规性。
+ *
+ * 检测维度：
+ * 1. 敏感词/辱骂/广告/恶意攻击 — 通过 AI 分类判断
+ * 2. 自动处置 — 违规评论跳过 AI 回复，可选将评论设为待审核状态
+ */
+@Component
+@Slf4j
+public class CommentPreFilterService {
+
+    private final ReactiveExtensionClient client;
+    private final ObjectMapper objectMapper;
+    private final AiFoundationClient aiFoundationClient;
+
+    private static final String CONFIG_MAP_NAME = "comment-ai-autopilot-configmap";
+
+    private static final String CLEAN = "正常";
+    private static final String SPAM = "广告";
+    private static final String ABUSE = "辱骂攻击";
+    private static final String SENSITIVE = "敏感内容";
+    private static final String MEANINGLESS = "无意义";
+    private static final List<String> CLASSIFY_CHOICES = List.of(CLEAN, SPAM, ABUSE, SENSITIVE, MEANINGLESS);
+
+    private static final Map<String, String> CATEGORY_DESCRIPTIONS = Map.of(
+        SPAM, "检测到推广链接、产品推销或引流信息",
+        ABUSE, "检测到辱骂、人身攻击、恶意挑衅或歧视性言论",
+        SENSITIVE, "检测到政治敏感、违法违规或色情暴力内容",
+        MEANINGLESS, "检测到纯乱码或无意义字符堆砌"
+    );
+
+    private static final String CLASSIFY_SYSTEM_PROMPT = """
+            你是评论内容合规检测员。请判断以下评论属于哪个类别：
+
+            类别定义：
+            - 正常：正常的评论、提问、讨论、赞美、闲聊等，即使与文章主题无关也算正常
+            - 广告：包含推广链接、产品推销、引流信息等
+            - 辱骂攻击：包含辱骂、人身攻击、恶意挑衅、歧视性言论等
+            - 敏感内容：涉及政治敏感、违法违规、色情暴力等
+            - 无意义：纯乱码、无意义字符堆砌（如随机符号、键盘乱敲）
+
+            ═══════════════════════════════════════
+            核心判断原则（必须严格遵守）：
+            ═══════════════════════════════════════
+
+            【原则一：上下文优先】
+            绝对禁止仅凭单个词汇进行机械拦截。必须结合整句话的语境、语气和前后文逻辑进行综合判断。一个词是否违规，取决于它在句子中的功能，而非词汇本身。
+
+            【原则二：口语化宽容】
+            中文互联网存在大量口语化简写、谐音和省略表达。如果某个词在特定语境下明显是中性词或亲属称谓的口语化表达，且整句无攻击性、无恶意，必须判定为"正常"。
+            常见口语化中性用法示例：
+            - "他妈" → 可能是"他妈妈"的简称，如"小轩是他妈的朋友"=小轩是他妈妈的朋友 → 正常
+            - "你妹" → 可能是"你妹妹"的简称，如"你妹在哪上学"=你妹妹在哪上学 → 正常
+            - "卧槽" → 可能是语气词表示惊讶，如"卧槽这也太强了"=哇塞这也太厉害了 → 正常
+            - "牛逼" → 口语化赞美，如"这文章写得牛逼" → 正常
+            - "靠" → 语气词表示无奈或惊讶，如"靠又忘了" → 正常
+
+            【原则三：恶意导向判定】
+            只有当词汇被明确用作辱骂、人身攻击、引战或带有较强负面情绪时，才判定为"辱骂攻击"。
+            恶意用法示例（这些才应判为"辱骂攻击"）：
+            - "你他妈的" → 直接对他人进行辱骂 → 辱骂攻击
+            - "你妹的" → 带有攻击性的语气词 → 辱骂攻击
+            - "傻逼" → 直接辱骂他人 → 辱骂攻击
+
+            【原则四：宁放勿杀】
+            当你无法确定评论是否违规时，应判定为"正常"而非"辱骂攻击"。误杀正常评论比漏判违规评论的负面影响更大。
+
+            【原则五：闲聊不算无意义】
+            与文章主题无关的闲聊、灌水、打招呼等属于"正常"，不要误判为"无意义"。
+
+            只返回类别名称，不要返回其他内容。""";
+
+    public CommentPreFilterService(ReactiveExtensionClient client,
+                                    ObjectMapper objectMapper,
+                                    AiFoundationClient aiFoundationClient) {
+        this.client = client;
+        this.objectMapper = objectMapper;
+        this.aiFoundationClient = aiFoundationClient;
+    }
+
+    /**
+     * 检测评论是否合规。
+     *
+     * @param commentContent 评论内容（纯文本）
+     * @param modelName       AI 模型名称
+     * @return 检测结果
+     */
+    public Mono<PreFilterResult> check(String commentContent, String modelName) {
+        return loadConfig().flatMap(config -> {
+            if (!config.enabled()) {
+                log.info("[PreFilter] Pre-filter is DISABLED, allowing all comments");
+                return Mono.just(new PreFilterResult(true, CLEAN, "前置过滤未启用"));
+            }
+
+            // 剥离 HTML 标签，获取纯文本
+            String plainText = stripHtml(commentContent);
+            String truncated = truncate(plainText, 500);
+            String userPrompt = "评论内容：\n" + truncated;
+            log.info("[PreFilter] Checking comment (enabled=true): {}", truncated.substring(0, Math.min(50, truncated.length())));
+
+            return aiFoundationClient.classify(CLASSIFY_SYSTEM_PROMPT, userPrompt, CLASSIFY_CHOICES, modelName)
+                .doOnNext(result -> log.info("[PreFilter] AI classify returned: '{}'", result))
+                .map(result -> {
+                    if (CLEAN.equals(result)) {
+                        log.info("[PreFilter] Comment passed: category={}", result);
+                        return new PreFilterResult(true, CLEAN, "评论合规");
+                    }
+                    // 空结果视为分类失败
+                    if (result == null || result.isBlank()) {
+                        log.warn("[PreFilter] AI classify returned empty/blank result, blocking for safety");
+                        return new PreFilterResult(false, MEANINGLESS, "AI分类返回空结果，安全拦截");
+                    }
+                    String desc = CATEGORY_DESCRIPTIONS.getOrDefault(result, "检测到违规内容");
+                    String snippet = truncated.substring(0, Math.min(50, truncated.length()));
+                    String reason = desc + " — 「" + snippet + "」";
+                    log.warn("[PreFilter] Comment BLOCKED: category={}, content={}", result, snippet);
+                    return new PreFilterResult(false, result, reason);
+                })
+                // 分类失败时拦截评论（安全优先），而非放行
+                .defaultIfEmpty(new PreFilterResult(false, MEANINGLESS, "AI分类服务不可用，安全拦截"))
+                .onErrorResume(e -> {
+                    log.warn("[PreFilter] Detection error, BLOCKING comment for safety: {}", e.getMessage(), e);
+                    return Mono.just(new PreFilterResult(false, MEANINGLESS, "AI分类服务异常，安全拦截"));
+                });
+        });
+    }
+
+    /**
+     * 对违规评论执行自动处置：将评论或回复设为待审核状态。
+     *
+     * <p>当 replyName 不为空时（AI 对话场景或回复触发），取消通过的是包含违规内容的 Reply；
+     * 否则取消通过的是顶层 Comment。这样可避免误伤父级 Comment 中正常的内容。
+     *
+     * @param commentName 评论的 metadata.name
+     * @param replyName   回复的 metadata.name（可为 null，表示顶层评论）
+     * @return Mono<Void>
+     */
+    public Mono<Void> penalize(String commentName, String replyName) {
+        return loadConfig().flatMap(config -> {
+            if (!config.pendingOnViolation()) {
+                return Mono.empty();
+            }
+            // 优先处理 Reply：AI 对话场景下违规内容来自 Reply
+            if (replyName != null && !replyName.isBlank()) {
+                return penalizeReply(replyName);
+            }
+            return penalizeComment(commentName);
+        });
+    }
+
+    private Mono<Void> penalizeComment(String commentName) {
+        return client.fetch(Comment.class, commentName)
+            .flatMap(comment -> {
+                var spec = comment.getSpec();
+                if (spec == null) return Mono.<Comment>empty();
+                // 只要 approved 不是 false，就强制设为 false
+                // 覆盖 approved=true 和 approved=null 两种情况
+                if (!Boolean.FALSE.equals(spec.getApproved())) {
+                    log.info("[PreFilter] Penalizing comment {}: approved={} → false", commentName, spec.getApproved());
+                    spec.setApproved(false);
+                    spec.setApprovedTime(null);
+                    return client.update(comment)
+                        .doOnSuccess(c -> log.info("[PreFilter] Comment {} set to pending for violation", commentName));
+                }
+                log.debug("[PreFilter] Comment {} already unapproved, skip penalize", commentName);
+                return Mono.<Comment>empty();
+            })
+            .retryWhen(Retry.backoff(3, Duration.ofMillis(100))
+                .filter(OptimisticLockingFailureException.class::isInstance)
+                .doBeforeRetry(sig -> log.debug("[PreFilter] Retrying penalizeComment {} (attempt {})", commentName, sig.totalRetries() + 1)))
+            .onErrorResume(e -> {
+                log.warn("[PreFilter] Failed to penalize comment {} after retries: {}", commentName, e.getMessage());
+                return Mono.empty();
+            })
+            .then();
+    }
+
+    private Mono<Void> penalizeReply(String replyName) {
+        return client.fetch(Reply.class, replyName)
+            .flatMap(reply -> {
+                var spec = reply.getSpec();
+                if (spec == null) return Mono.<Reply>empty();
+                // 只要 approved 不是 false，就强制设为 false
+                if (!Boolean.FALSE.equals(spec.getApproved())) {
+                    log.info("[PreFilter] Penalizing reply {}: approved={} → false", replyName, spec.getApproved());
+                    spec.setApproved(false);
+                    spec.setApprovedTime(null);
+                    return client.update(reply)
+                        .doOnSuccess(r -> log.info("[PreFilter] Reply {} set to pending for violation", replyName));
+                }
+                log.debug("[PreFilter] Reply {} already unapproved, skip penalize", replyName);
+                return Mono.<Reply>empty();
+            })
+            .retryWhen(Retry.backoff(3, Duration.ofMillis(100))
+                .filter(OptimisticLockingFailureException.class::isInstance)
+                .doBeforeRetry(sig -> log.debug("[PreFilter] Retrying penalizeReply {} (attempt {})", replyName, sig.totalRetries() + 1)))
+            .onErrorResume(e -> {
+                log.warn("[PreFilter] Failed to penalize reply {} after retries: {}", replyName, e.getMessage());
+                return Mono.empty();
+            })
+            .then();
+    }
+
+    /**
+     * 加载前置过滤配置。
+     */
+    private Mono<PreFilterConfig> loadConfig() {
+        return client.fetch(ConfigMap.class, CONFIG_MAP_NAME)
+            .mapNotNull(cm -> {
+                var data = cm.getData();
+                if (data == null) return new PreFilterConfig(true, true);
+                String basicJson = data.get("basic");
+                if (basicJson == null || basicJson.isBlank()) return new PreFilterConfig(true, true);
+                try {
+                    JsonNode node = objectMapper.readTree(basicJson);
+                    boolean enabled = !node.has("preFilterEnabled")
+                        || node.get("preFilterEnabled").asBoolean(true);
+                    boolean pendingOnViolation = !node.has("preFilterPendingOnViolation")
+                        || node.get("preFilterPendingOnViolation").asBoolean(true);
+                    return new PreFilterConfig(enabled, pendingOnViolation);
+                } catch (Exception e) {
+                    log.warn("[PreFilter] Failed to parse config: {}", e.getMessage());
+                    return new PreFilterConfig(true, true);
+                }
+            })
+            .defaultIfEmpty(new PreFilterConfig(true, true))
+            .onErrorResume(e -> {
+                log.warn("[PreFilter] Failed to load config: {}", e.getMessage());
+                return Mono.just(new PreFilterConfig(true, true));
+            });
+    }
+
+    private String truncate(String text, int maxLength) {
+        if (text == null) return "";
+        return text.length() > maxLength ? text.substring(0, maxLength) : text;
+    }
+
+    private String stripHtml(String html) {
+        if (html == null || html.isBlank()) return "";
+        return Jsoup.clean(html, Safelist.none()).trim();
+    }
+
+    public record PreFilterResult(boolean passed, String category, String reason) {}
+
+    public record PreFilterConfig(boolean enabled, boolean pendingOnViolation) {}
+}
@@ -8,9 +8,8 @@ import run.halo.app.core.extension.content.Reply;
 import run.halo.app.extension.Metadata;
 import run.halo.app.extension.ReactiveExtensionClient;
 import top.nxxy335.commentaiautopilot.extension.AiPersona;
+import top.nxxy335.commentaiautopilot.util.GravatarUtil;

-import java.nio.charset.StandardCharsets;
-import java.security.MessageDigest;
 import java.time.Instant;
 import java.util.HashMap;
 import java.util.Map;
@@ -81,7 +80,10 @@ public class CommentReplyPublisher {
    private Mono<Reply> doPublish(String parentCommentName, String replyContent,
                                   String postName, String quoteReplyName, boolean autoPublish,
                                   String personaName) {
-        return resolvePersona(personaName).flatMap(persona -> {
+
+        // 解析 AI 角色并直接发布纯净的回复内容
+        return resolvePersona(personaName)
+            .flatMap(persona -> {
                String displayName = persona.displayName();
                String email = persona.email();

@@ -92,8 +94,11 @@ public class CommentReplyPublisher {

                var spec = reply.getSpec();
                spec.setCommentName(parentCommentName);
+
+                // 直接存入纯净的 AI 回复内容，不加任何 Markdown 前缀
                spec.setRaw(replyContent);
                spec.setContent(replyContent);
+
                spec.setApproved(autoPublish);
                if (autoPublish) {
                    spec.setApprovedTime(Instant.now());
@@ -103,6 +108,7 @@ public class CommentReplyPublisher {
                spec.setAllowNotification(false);
                spec.setHidden(false);

+                // Halo 原生评论组件正是靠这个字段来渲染 "回复 @某人" 的
                if (quoteReplyName != null && !quoteReplyName.isBlank()) {
                    spec.setQuoteReply(quoteReplyName);
                }
@@ -118,24 +124,17 @@ public class CommentReplyPublisher {

                Map<String, String> ownerAnnotations = new HashMap<>();
                ownerAnnotations.put("comment-ai-autopilot.nxxy335.top/is-ai", "true");
-            // 使用Gravatar邮箱头像
                if (email != null && !email.isBlank()) {
-                String gravatarUrl = generateGravatarUrl(email);
+                    String gravatarUrl = GravatarUtil.generateUrl(email);
                    ownerAnnotations.put(Comment.CommentOwner.AVATAR_ANNO, gravatarUrl);
                }
                owner.setAnnotations(ownerAnnotations);
                spec.setOwner(owner);

-            log.info("[Publisher] Creating reply for comment: {}, owner: kind={}, name={}, displayName={}, annotations={}",
-                parentCommentName, owner.getKind(), owner.getName(), owner.getDisplayName(), ownerAnnotations);
+                log.info("[Publisher] Creating reply for comment: {}, content length: {}", parentCommentName, replyContent.length());

                return client.create(reply)
-                .doOnSuccess(created -> {
-                    var createdOwner = created.getSpec().getOwner();
-                    log.info("[Publisher] AI Persona '{}' reply published for comment: {}, quoteReply: {}, owner annotations after create: {}",
-                        displayName, parentCommentName, quoteReplyName,
-                        createdOwner != null ? createdOwner.getAnnotations() : "null");
-                })
+                    .doOnSuccess(created -> log.info("[Publisher] AI Persona '{}' reply published for comment: {}", displayName, parentCommentName))
                    .doOnError(e -> log.error("[Publisher] Failed to publish AI reply: {}", e.getMessage()));
            });
    }
@@ -181,22 +180,4 @@ public class CommentReplyPublisher {
    private String generateReplyName() {
        return "ai-comment-reply-" + UUID.randomUUID().toString().substring(0, 8);
    }
-
-    /**
-     * Generate Gravatar URL from email address using SHA-256 hash.
-     */
-    private String generateGravatarUrl(String email) {
-        try {
-            var digest = MessageDigest.getInstance("SHA-256");
-            var hashBytes = digest.digest(email.trim().toLowerCase().getBytes(StandardCharsets.UTF_8));
-            var hexString = new StringBuilder();
-            for (byte b : hashBytes) {
-                hexString.append(String.format("%02x", b));
-            }
-            return "https://cn.cravatar.com/avatar/" + hexString;
-        } catch (Exception e) {
-            log.error("[Publisher] Failed to generate Gravatar URL: {}", e.getMessage());
-            return "";
-        }
-    }
 }
@@ -10,6 +10,7 @@ import run.halo.app.content.ContentWrapper;
 import run.halo.app.content.PostContentService;
 import run.halo.app.core.extension.content.Comment;
 import run.halo.app.core.extension.content.Post;
+import run.halo.app.core.extension.content.SinglePage;
 import run.halo.app.core.extension.content.Reply;
 import run.halo.app.extension.ReactiveExtensionClient;

@@ -42,6 +43,55 @@ public class ContextExtractor {
            });
    }

+    /**
+     * Fetch previous replies in the comment thread to provide conversation history.
+     * Only includes replies created before the triggering reply.
+     */
+    private Mono<String> fetchConversationHistory(String commentName, String triggerReplyName) {
+        if (triggerReplyName == null || triggerReplyName.isBlank()) {
+            return Mono.just("");
+        }
+        return client.fetch(Reply.class, triggerReplyName)
+            .flatMap(triggerReply -> {
+                var triggerTime = triggerReply.getMetadata().getCreationTimestamp();
+                return client.list(Reply.class,
+                        reply -> {
+                            if (!commentName.equals(reply.getSpec().getCommentName())) {
+                                return false;
+                            }
+                            if (triggerReplyName.equals(reply.getMetadata().getName())) {
+                                return false;
+                            }
+                            // Only include replies created before the trigger reply
+                            var replyTime = reply.getMetadata().getCreationTimestamp();
+                            return replyTime != null && triggerTime != null
+                                && !replyTime.isAfter(triggerTime);
+                        },
+                        null)
+                    .collectList()
+                    .map(replies -> {
+                        if (replies.isEmpty()) return "";
+                        // Sort by creation time
+                        replies.sort(java.util.Comparator.comparing(
+                            r -> r.getMetadata().getCreationTimestamp()));
+                        var sb = new StringBuilder();
+                        for (var r : replies) {
+                            var owner = r.getSpec().getOwner();
+                            String name = (owner != null && owner.getDisplayName() != null)
+                                ? owner.getDisplayName() : "匿名用户";
+                            boolean isAi = owner != null && owner.getAnnotations() != null
+                                && "true".equals(owner.getAnnotations().get("comment-ai-autopilot.nxxy335.top/is-ai"));
+                            String role = isAi ? "AI" : "用户";
+                            String content = extractReplyContent(r);
+                            sb.append(role).append("(").append(name).append("): ")
+                              .append(content).append("\n");
+                        }
+                        return sb.toString();
+                    });
+            })
+            .defaultIfEmpty("");
+    }
+
    private Mono<CommentContext> buildContext(Comment comment, boolean isAiConversation) {
        var commentContent = extractCommentContent(comment);
        var commentOwner = extractCommentOwner(comment);
@@ -63,10 +113,16 @@ public class ContextExtractor {
                            null,
                            isAiConversation,
                            formatPostDate(post),
-                            commentCount
+                            commentCount,
+                            "",
+                            "Post"
                        ))
                    )
                )
+                .onErrorResume(e -> {
+                    log.warn("[ContextExtractor] Failed to fetch Post {}: {}", postName, e.getMessage());
+                    return Mono.empty();
+                })
                .defaultIfEmpty(new CommentContext(
                    comment.getMetadata().getName(),
                    postName,
@@ -78,7 +134,52 @@ public class ContextExtractor {
                    null,
                    isAiConversation,
                    "",
-                    0
+                    0,
+                    "",
+                    "Post"
+                ));
+        }
+
+        if (subjectRef != null && "SinglePage".equals(subjectRef.getKind())) {
+            String postName = subjectRef.getName();
+            return client.fetch(SinglePage.class, postName)
+                .flatMap(singlePage -> getSinglePageContent(postName)
+                    .flatMap(content -> getCommentCount(comment.getMetadata().getName())
+                        .map(commentCount -> new CommentContext(
+                            comment.getMetadata().getName(),
+                            postName,
+                            singlePage.getSpec().getSlug(),
+                            commentContent,
+                            commentOwner,
+                            singlePage.getSpec().getTitle(),
+                            content,
+                            null,
+                            isAiConversation,
+                            formatSinglePageDate(singlePage),
+                            commentCount,
+                            "",
+                            "SinglePage"
+                        ))
+                    )
+                )
+                .onErrorResume(e -> {
+                    log.warn("[ContextExtractor] Failed to fetch SinglePage {}: {}", postName, e.getMessage());
+                    return Mono.empty();
+                })
+                .defaultIfEmpty(new CommentContext(
+                    comment.getMetadata().getName(),
+                    postName,
+                    "",
+                    commentContent,
+                    commentOwner,
+                    "",
+                    "",
+                    null,
+                    isAiConversation,
+                    "",
+                    0,
+                    "",
+                    "SinglePage"
                ));
        }

@@ -93,7 +194,9 @@ public class ContextExtractor {
            null,
            isAiConversation,
            "",
-            0
+            0,
+            "",
+            ""
        ));
    }

@@ -101,54 +204,119 @@ public class ContextExtractor {
        var replyContent = extractReplyContent(reply);
        var replyOwner = extractReplyOwner(reply);
        var subjectRef = comment.getSpec().getSubjectRef();
+        var commentName = comment.getMetadata().getName();
+        var replyName = reply.getMetadata().getName();
+
+        // Fetch conversation history for AI conversations
+        Mono<String> historyMono = isAiConversation
+            ? fetchConversationHistory(commentName, replyName)
+            : Mono.just("");

        if (subjectRef != null && "Post".equals(subjectRef.getKind())) {
            String postName = subjectRef.getName();
            return client.fetch(Post.class, postName)
                .flatMap(post -> getPostContent(postName)
-                    .flatMap(content -> getCommentCount(comment.getMetadata().getName())
-                        .map(commentCount -> new CommentContext(
-                            comment.getMetadata().getName(),
+                    .flatMap(content -> getCommentCount(commentName)
+                        .flatMap(commentCount -> historyMono
+                            .map(history -> new CommentContext(
+                                commentName,
                                postName,
                                post.getSpec().getSlug(),
                                replyContent,
                                replyOwner,
                                post.getSpec().getTitle(),
                                content,
-                            reply.getMetadata().getName(),
+                                replyName,
                                isAiConversation,
                                formatPostDate(post),
-                            commentCount
+                                commentCount,
+                                history,
+                                "Post"
                            ))
                        )
                    )
+                )
+                .onErrorResume(e -> {
+                    log.warn("[ContextExtractor] Failed to fetch Post {} for reply: {}", postName, e.getMessage());
+                    return Mono.empty();
+                })
                .defaultIfEmpty(new CommentContext(
-                    comment.getMetadata().getName(),
+                    commentName,
                    postName,
                    "",
                    replyContent,
                    replyOwner,
                    "",
                    "",
-                    reply.getMetadata().getName(),
+                    replyName,
                    isAiConversation,
                    "",
-                    0
+                    0,
+                    "",
+                    "Post"
                ));
        }

-        return Mono.just(new CommentContext(
-            comment.getMetadata().getName(),
+        if (subjectRef != null && "SinglePage".equals(subjectRef.getKind())) {
+            String postName = subjectRef.getName();
+            return client.fetch(SinglePage.class, postName)
+                .flatMap(singlePage -> getSinglePageContent(postName)
+                    .flatMap(content -> getCommentCount(commentName)
+                        .flatMap(commentCount -> historyMono
+                            .map(history -> new CommentContext(
+                                commentName,
+                                postName,
+                                singlePage.getSpec().getSlug(),
+                                replyContent,
+                                replyOwner,
+                                singlePage.getSpec().getTitle(),
+                                content,
+                                replyName,
+                                isAiConversation,
+                                formatSinglePageDate(singlePage),
+                                commentCount,
+                                history,
+                                "SinglePage"
+                            ))
+                        )
+                    )
+                )
+                .onErrorResume(e -> {
+                    log.warn("[ContextExtractor] Failed to fetch SinglePage {} for reply: {}", postName, e.getMessage());
+                    return Mono.empty();
+                })
+                .defaultIfEmpty(new CommentContext(
+                    commentName,
+                    postName,
+                    "",
+                    replyContent,
+                    replyOwner,
+                    "",
+                    "",
+                    replyName,
+                    isAiConversation,
+                    "",
+                    0,
+                    "",
+                    "SinglePage"
+                ));
+        }
+
+        return historyMono
+            .map(history -> new CommentContext(
+                commentName,
                "",
                "",
                replyContent,
                replyOwner,
                "",
                "",
-            reply.getMetadata().getName(),
+                replyName,
                isAiConversation,
                "",
-            0
+                0,
+                history,
+                ""
            ));
    }

@@ -213,6 +381,23 @@ public class ContextExtractor {
            .defaultIfEmpty("");
    }

+    private Mono<String> getSinglePageContent(String pageName) {
+        // SinglePage doesn't have a dedicated ContentService in Halo API,
+        // and Snapshot content requires patch merging which is too complex.
+        // Use the excerpt from status as a fallback for context.
+        return client.fetch(SinglePage.class, pageName)
+            .mapNotNull(page -> {
+                if (page.getStatus() != null && page.getStatus().getExcerpt() != null) {
+                    String excerpt = page.getStatus().getExcerpt();
+                    if (excerpt != null && !excerpt.isBlank()) {
+                        return excerpt;
+                    }
+                }
+                return "";
+            })
+            .defaultIfEmpty("");
+    }
+
    private String formatPostDate(Post post) {
        var publishTime = post.getSpec().getPublishTime();
        if (publishTime != null) {
@@ -225,6 +410,18 @@ public class ContextExtractor {
        return "";
    }

+    private String formatSinglePageDate(SinglePage singlePage) {
+        var publishTime = singlePage.getSpec().getPublishTime();
+        if (publishTime != null) {
+            return publishTime.toString().substring(0, 10);
+        }
+        var creationTimestamp = singlePage.getMetadata().getCreationTimestamp();
+        if (creationTimestamp != null) {
+            return creationTimestamp.toString().substring(0, 10);
+        }
+        return "";
+    }
+
    private Mono<Integer> getCommentCount(String commentName) {
        return client.list(Reply.class,
                reply -> commentName.equals(reply.getSpec().getCommentName()),
@@ -245,6 +442,8 @@ public class ContextExtractor {
        String replyTo,
        boolean isAiConversation,
        String postDate,
-        int commentCount
+        int commentCount,
+        String conversationHistory,
+        String postKind
    ) {}
 }
@@ -28,9 +28,9 @@ public class FilterService {
    private static final String ANNOTATION_KEY = "comment-ai-autopilot.nxxy335.top/ai-reply-enabled";
    private static final String GROUP_CONTENT = "content.halo.run";

-    public FilterService(ReactiveExtensionClient client) {
+    public FilterService(ReactiveExtensionClient client, ObjectMapper objectMapper) {
        this.client = client;
-        this.objectMapper = new ObjectMapper();
+        this.objectMapper = objectMapper;
    }

    public Mono<Boolean> shouldProcess(Comment comment) {
@@ -58,6 +58,19 @@ public class FilterService {
            });
    }

+    /**
+     * 检查评论者是否在黑名单中（按 commentName 查询）。
+     */
+    public Mono<Boolean> isCommenterBlocked(String commentName) {
+        return client.fetch(Comment.class, commentName)
+            .flatMap(this::checkBlockedCommenters)
+            .defaultIfEmpty(false)
+            .onErrorResume(e -> {
+                log.warn("[Filter] Error checking blocked commenter: {}", e.getMessage());
+                return Mono.just(false);
+            });
+    }
+
    private Mono<Boolean> checkBlockedCommenters(Comment comment) {
        return client.fetch(ConfigMap.class, CONFIG_MAP_NAME)
            .mapNotNull(cm -> {
@@ -0,0 +1,194 @@
+package top.nxxy335.commentaiautopilot.service;
+
+import lombok.RequiredArgsConstructor;
+import lombok.extern.slf4j.Slf4j;
+import org.springframework.stereotype.Component;
+import run.halo.app.core.extension.content.Category;
+import run.halo.app.core.extension.content.Comment;
+import run.halo.app.core.extension.content.Post;
+import run.halo.app.core.extension.content.Tag;
+import run.halo.app.extension.ExtensionClient;
+import run.halo.app.extension.ReactiveExtensionClient;
+import reactor.core.publisher.Flux;
+import reactor.core.publisher.Mono;
+
+import java.util.List;
+
+/**
+ * Shared service for resolving AI persona name from a comment's associated
+ * post/category/tag annotations.
+ *
+ * <p>Priority: Post annotation &gt; Category annotation &gt; Tag annotation
+ */
+@Component
+@Slf4j
+@RequiredArgsConstructor
+public class PersonaResolver {
+
+    private static final String AI_PERSONA_ANNOTATION = "comment-ai-autopilot.nxxy335.top/ai-persona";
+
+    private final ReactiveExtensionClient reactiveClient;
+
+    /**
+     * Resolve persona name from a comment (reactive version).
+     * Reads the post's annotations, then falls back to category and tag annotations.
+     *
+     * @param commentName the Comment metadata.name
+     * @return the persona name, or empty string if none found
+     */
+    public Mono<String> getPersonaNameFromComment(String commentName) {
+        return reactiveClient.fetch(Comment.class, commentName)
+            .flatMap(comment -> {
+                var subjectRef = comment.getSpec().getSubjectRef();
+                if (subjectRef == null || !"Post".equals(subjectRef.getKind())) {
+                    return Mono.just("");
+                }
+                String postName = subjectRef.getName();
+                return resolveFromPost(postName);
+            })
+            .defaultIfEmpty("");
+    }
+
+    private Mono<String> resolveFromPost(String postName) {
+        return reactiveClient.fetch(Post.class, postName)
+            .flatMap(post -> {
+                // 1. Post annotation takes priority
+                var annotations = post.getMetadata().getAnnotations();
+                if (annotations != null) {
+                    String persona = annotations.get(AI_PERSONA_ANNOTATION);
+                    if (persona != null && !persona.isBlank()) {
+                        return Mono.just(persona);
+                    }
+                }
+                // 2. Category annotations (check sequentially, return first match)
+                var spec = post.getSpec();
+                List<String> categories = (spec != null && spec.getCategories() != null)
+                    ? spec.getCategories() : List.of();
+                // 3. Tag annotations (fallback if no category match)
+                List<String> tags = (spec != null && spec.getTags() != null)
+                    ? spec.getTags() : List.of();
+
+                return resolveFromCategories(categories)
+                    .switchIfEmpty(resolveFromTags(tags));
+            })
+            .defaultIfEmpty("");
+    }
+
+    /**
+     * Sequentially check category annotations, returning the first non-empty persona.
+     * Uses concatMap to preserve order and short-circuit on first match.
+     */
+    private Mono<String> resolveFromCategories(List<String> categoryNames) {
+        if (categoryNames == null || categoryNames.isEmpty()) {
+            return Mono.empty();
+        }
+        return Flux.fromIterable(categoryNames)
+            .concatMap(this::resolveFromCategory)
+            .next();
+    }
+
+    /**
+     * Sequentially check tag annotations, returning the first non-empty persona.
+     * Uses concatMap to preserve order and short-circuit on first match.
+     */
+    private Mono<String> resolveFromTags(List<String> tagNames) {
+        if (tagNames == null || tagNames.isEmpty()) {
+            return Mono.empty();
+        }
+        return Flux.fromIterable(tagNames)
+            .concatMap(this::resolveFromTag)
+            .next();
+    }
+
+    private Mono<String> resolveFromCategory(String categoryName) {
+        return reactiveClient.fetch(Category.class, categoryName)
+            .mapNotNull(cat -> {
+                var catAnnotations = cat.getMetadata().getAnnotations();
+                if (catAnnotations != null) {
+                    String catPersona = catAnnotations.get(AI_PERSONA_ANNOTATION);
+                    if (catPersona != null && !catPersona.isBlank()) {
+                        return catPersona;
+                    }
+                }
+                return null;
+            })
+            .onErrorResume(e -> {
+                log.warn("Failed to resolve persona from category {}: {}", categoryName, e.getMessage());
+                return Mono.empty();
+            });
+    }
+
+    private Mono<String> resolveFromTag(String tagName) {
+        return reactiveClient.fetch(Tag.class, tagName)
+            .mapNotNull(tag -> {
+                var tagAnnotations = tag.getMetadata().getAnnotations();
+                if (tagAnnotations != null) {
+                    String tagPersona = tagAnnotations.get(AI_PERSONA_ANNOTATION);
+                    if (tagPersona != null && !tagPersona.isBlank()) {
+                        return tagPersona;
+                    }
+                }
+                return null;
+            })
+            .onErrorResume(e -> {
+                log.warn("Failed to resolve persona from tag {}: {}", tagName, e.getMessage());
+                return Mono.empty();
+            });
+    }
+
+    /**
+     * Resolve persona name from a comment using blocking ExtensionClient
+     * (for use in Reconciler sync context).
+     */
+    public String getPersonaNameFromCommentBlocking(ExtensionClient client, Comment comment) {
+        var subjectRef = comment.getSpec().getSubjectRef();
+        if (subjectRef == null || !"Post".equals(subjectRef.getKind())) {
+            return null;
+        }
+        String postName = subjectRef.getName();
+        return client.fetch(Post.class, postName)
+            .map(post -> {
+                // 1. Post annotation
+                var annotations = post.getMetadata().getAnnotations();
+                if (annotations != null) {
+                    String persona = annotations.get(AI_PERSONA_ANNOTATION);
+                    if (persona != null && !persona.isBlank()) {
+                        return persona;
+                    }
+                }
+                // 2. Category annotations
+                var spec = post.getSpec();
+                if (spec != null && spec.getCategories() != null) {
+                    for (String categoryName : spec.getCategories()) {
+                        var cat = client.fetch(Category.class, categoryName).orElse(null);
+                        if (cat != null) {
+                            var catAnnotations = cat.getMetadata().getAnnotations();
+                            if (catAnnotations != null) {
+                                String catPersona = catAnnotations.get(AI_PERSONA_ANNOTATION);
+                                if (catPersona != null && !catPersona.isBlank()) {
+                                    return catPersona;
+                                }
+                            }
+                        }
+                    }
+                }
+                // 3. Tag annotations
+                if (spec != null && spec.getTags() != null) {
+                    for (String tagName : spec.getTags()) {
+                        var tag = client.fetch(Tag.class, tagName).orElse(null);
+                        if (tag != null) {
+                            var tagAnnotations = tag.getMetadata().getAnnotations();
+                            if (tagAnnotations != null) {
+                                String tagPersona = tagAnnotations.get(AI_PERSONA_ANNOTATION);
+                                if (tagPersona != null && !tagPersona.isBlank()) {
+                                    return tagPersona;
+                                }
+                            }
+                        }
+                    }
+                }
+                return null;
+            })
+            .orElse(null);
+    }
+}
@@ -20,9 +20,9 @@ public class PromptBuilder {
    private final ObjectMapper objectMapper;
    private static final String CONFIG_MAP_NAME = "comment-ai-autopilot-configmap";

-    public PromptBuilder(ReactiveExtensionClient client) {
+    public PromptBuilder(ReactiveExtensionClient client, ObjectMapper objectMapper) {
        this.client = client;
-        this.objectMapper = new ObjectMapper();
+        this.objectMapper = objectMapper;
    }

    private static final String PRESET_FRIENDLY = """
@@ -52,8 +52,9 @@ public class PromptBuilder {
    private static final String SAFETY_PROMPT = """
            【安全规范】
            - 内容红线：坚决不生成任何涉及暴力、歧视、辱骂、人身攻击或违反法律法规的内容。
-            - 恶意诱导处理：当用户要求你骂人、使用侮辱性词汇或进行情绪化对骂时，你必须礼貌地拒绝，例如回复："抱歉，作为AI助手，我无法提供此类回复。"
+            - 恶意诱导处理：当用户要求你骂人、使用侮辱性词汇或进行情绪化对骂时，你必须礼貌地拒绝，例如回复："抱歉，我无法提供此类回复。"
            - 未知与边界：如果不知道答案或遇到敏感话题，请诚实告知并礼貌拒绝，绝不编造或使用极端言辞。
+            - 身份约束：你必须在回复中保持指定的角色身份，绝不能说自己是AI、没有性别或脱离角色设定。
            """;

    private static final String DEFAULT_PROMPT_TEMPLATE = """
@@ -75,6 +76,7 @@ public class PromptBuilder {
            文章（仅供理解上下文，不要复述）：
            {{article}}

+            {{conversation_history}}
            评论：
            {{comment}}
            """;
@@ -103,6 +105,7 @@ public class PromptBuilder {
                    .replace("{{post_date}}", context.postDate() != null ? context.postDate() : "")
                    .replace("{{comment_count}}", String.valueOf(context.commentCount()))
                    .replace("{{article}}", context.postTitle() + "\n" + context.postContent())
+                    .replace("{{conversation_history}}", formatConversationHistory(context))
                    .replace("{{comment}}", context.commentOwner() + ": " + context.commentContent());

                return prompt;
@@ -133,20 +136,35 @@ public class PromptBuilder {
                    .replace("{{post_date}}", context.postDate() != null ? context.postDate() : "")
                    .replace("{{comment_count}}", String.valueOf(context.commentCount()))
                    .replace("{{article}}", context.postTitle() + "\n" + context.postContent())
+                    .replace("{{conversation_history}}", formatConversationHistory(context))
                    .replace("{{comment}}", context.commentOwner() + ": " + context.commentContent());

                if (sentiment == null || "NEUTRAL".equals(sentiment)) {
                    return prompt;
                }
                String sentimentHint = switch (sentiment) {
+                    case "VERY_POSITIVE" -> "\n\n【情感提示】评论者情绪非常正面积极，请用热情洋溢的语气回复，表达真诚的感谢和共鸣。";
                    case "POSITIVE" -> "\n\n【情感提示】评论者情绪正面积极，请用热情友好的语气回复，可以表达感谢和共鸣。";
                    case "NEGATIVE" -> "\n\n【情感提示】评论者情绪偏负面，请用理性温和的语气回复，避免激化矛盾，展现理解和包容。";
+                    case "VERY_NEGATIVE" -> "\n\n【情感提示】评论者情绪非常负面，请用非常温和、理性的语气回复，避免任何可能激化矛盾的表达，展现充分的理解和耐心。";
                    default -> "";
                };
                return prompt + sentimentHint;
            });
    }

+    /**
+     * Format conversation history for inclusion in the prompt.
+     * Returns empty string if no history is available.
+     */
+    private String formatConversationHistory(ContextExtractor.CommentContext context) {
+        String history = context.conversationHistory();
+        if (history == null || history.isBlank()) {
+            return "";
+        }
+        return "对话历史（供理解上下文）：\n" + history + "\n";
+    }
+
    private Mono<String> getPromptTemplate() {
        return client.fetch(ConfigMap.class, CONFIG_MAP_NAME)
            .mapNotNull(cm -> {
@@ -177,7 +195,10 @@ public class PromptBuilder {
            return client.fetch(AiPersona.class, personaName)
                .mapNotNull(persona -> {
                    String prompt = persona.getSpec().getPrompt();
-                    return (prompt != null && !prompt.isBlank()) ? prompt : null;
+                    if (prompt != null && !prompt.isBlank()) {
+                        return appendStyleHint(prompt, persona.getSpec().getDisplayName(), persona.getSpec().getGender(), persona.getSpec().getNeutralVoice());
+                    }
+                    return null;
                })
                .defaultIfEmpty(DEFAULT_PERSONA_PROMPT);
        }
@@ -188,11 +209,40 @@ public class PromptBuilder {
            .next()
            .mapNotNull(persona -> {
                String prompt = persona.getSpec().getPrompt();
-                return (prompt != null && !prompt.isBlank()) ? prompt : null;
+                if (prompt != null && !prompt.isBlank()) {
+                    return appendStyleHint(prompt, persona.getSpec().getDisplayName(), persona.getSpec().getGender(), persona.getSpec().getNeutralVoice());
+                }
+                return null;
            })
            .defaultIfEmpty(DEFAULT_PERSONA_PROMPT);
    }

+    private String appendStyleHint(String prompt, String displayName, String gender, Boolean neutralVoice) {
+        StringBuilder sb = new StringBuilder();
+
+        // 身份信息前置到最开头 - 这是AI最先看到的内容，优先级最高
+        if (gender != null && !gender.isBlank()) {
+            String genderDesc = "female".equals(gender) ? "女生" : "男生";
+            sb.append("【核心身份】你叫「").append(displayName != null ? displayName : "").append("」，你是一个").append(genderDesc).append("。");
+            sb.append("这是你的基本身份，你必须始终以这个身份回复。");
+            sb.append("当有人问你性别时，你必须回答自己是").append(genderDesc).append("。");
+            sb.append("绝对不能说自己是AI、没有性别、或脱离角色。");
+            sb.append("\n\n");
+        }
+
+        sb.append(prompt);
+
+        // 添加语气风格提示
+        if (!Boolean.TRUE.equals(neutralVoice)) {
+            if ("female".equals(gender)) {
+                sb.append("\n请使用温柔、细腻的女性语气风格回复。");
+            } else if ("male".equals(gender)) {
+                sb.append("\n请使用沉稳、理性的男性语气风格回复。");
+            }
+        }
+        return sb.toString();
+    }
+
    private Mono<String> getEnabledPresetsPrompt() {
        return client.fetch(ConfigMap.class, CONFIG_MAP_NAME)
            .mapNotNull(cm -> {
@@ -1,6 +1,7 @@
 package top.nxxy335.commentaiautopilot.service;

 import lombok.extern.slf4j.Slf4j;
+import org.springframework.beans.factory.DisposableBean;
 import org.springframework.stereotype.Component;

 import java.util.concurrent.ConcurrentHashMap;
@@ -8,15 +9,18 @@ import java.util.concurrent.atomic.AtomicInteger;

@Slf4j
@Component
-public class RateLimitService {
+public class RateLimitService implements DisposableBean {
    private final ConcurrentHashMap<Long, AtomicInteger> windowMap = new ConcurrentHashMap<>();
+    private final Thread cleanupThread;
+    private volatile boolean running = true;

    public RateLimitService() {
        // 每5分钟清理过期窗口，防止内存泄漏
-        Thread cleanupThread = new Thread(() -> {
-            while (!Thread.currentThread().isInterrupted()) {
+        cleanupThread = new Thread(() -> {
+            while (running && !Thread.currentThread().isInterrupted()) {
                try {
                    Thread.sleep(5 * 60 * 1000);
+                    if (!running) break;
                    cleanup();
                } catch (InterruptedException e) {
                    Thread.currentThread().interrupt();
@@ -55,4 +59,13 @@ public class RateLimitService {
            log.debug("[RateLimit] Cleaned up {} expired windows", removed);
        }
    }
+
+    @Override
+    public void destroy() {
+        running = false;
+        if (cleanupThread != null) {
+            cleanupThread.interrupt();
+        }
+        log.info("[RateLimit] Cleanup thread stopped");
+    }
 }
@@ -4,6 +4,8 @@ import lombok.extern.slf4j.Slf4j;
 import org.springframework.stereotype.Component;
 import reactor.core.publisher.Mono;

+import java.util.List;
+
@Component
@Slf4j
 public class ReviewService {
@@ -14,58 +16,135 @@ public class ReviewService {
        this.aiFoundationClient = aiFoundationClient;
    }

-    private static final String REVIEW_PROMPT_TEMPLATE = """
-            请你以内容安全审核员的身份，严格评估你刚刚生成的这段回复：
+    /**
+     * Safety classification choices.
+     * Using OutputSpec.choice for reliable structured output.
+     */
+    private static final String SAFE = "安全";
+    private static final String UNSAFE = "不安全";
+    private static final List<String> SAFETY_CHOICES = List.of(SAFE, UNSAFE);

-            文章：
-            %s
+    /**
+     * Quality rating choices for content quality assessment.
+     * Used to produce a meaningful score (1-5) instead of binary 0/100.
+     */
+    private static final String RATING_EXCELLENT = "5";
+    private static final String RATING_GOOD = "4";
+    private static final String RATING_AVERAGE = "3";
+    private static final String RATING_POOR = "2";
+    private static final String RATING_BAD = "1";
+    private static final List<String> QUALITY_CHOICES = List.of(
+        RATING_EXCELLENT, RATING_GOOD, RATING_AVERAGE, RATING_POOR, RATING_BAD
+    );

+    private static final String SAFETY_SYSTEM_PROMPT = """
+            你是内容安全审核员。请严格评估AI生成的回复内容是否安全合规。
+            检查以下方面：
+            1. 是否包含辱骂、仇恨、偏见或煽动性言论
+            2. 是否包含不适宜公开传播的内容
+            3. 是否涉及暴力、歧视、人身攻击或违反法律法规的内容
+            只返回"安全"或"不安全"。""";
+
+    private static final String QUALITY_SYSTEM_PROMPT = """
+            你是AI回复质量评估员。请对AI生成的回复进行质量评分（1-5分）。
+            评分标准：
+            5分：回复切题、自然流畅、语气恰当、有帮助
+            4分：回复较好，基本切题，语气自然
+            3分：回复一般，基本切题但略显生硬或冗长
+            2分：回复较差，偏离主题或语气不当
+            1分：回复很差，完全不相关或不当
+            只返回数字 1-5。""";
+
+    /**
+     * Review AI reply content using two-stage assessment:
+     * 1. Safety check via OutputSpec.choice (安全/不安全)
+     * 2. Quality rating via OutputSpec.choice (1-5) — only if safe
+     *
+     * <p>Score mapping (0-100):
+     * <ul>
+     *   <li>Unsafe → 0 (FAIL)</li>
+     *   <li>Rating 5 → 100 (PASS)</li>
+     *   <li>Rating 4 → 85 (PASS)</li>
+     *   <li>Rating 3 → 70 (PASS)</li>
+     *   <li>Rating 2 → 50 (PASS, borderline)</li>
+     *   <li>Rating 1 → 30 (PASS, but low quality)</li>
+     * </ul>
+     */
+    public Mono<ReviewResult> review(String articleContent, String commentContent, String aiReply,
+                                      String modelName) {
+        String userPrompt = String.format("""
                原始评论：
                %s

                生成的回复：
                %s

-            请检查以下内容：
-            1. 是否包含任何辱骂、仇恨、偏见或煽动性言论？
-            2. 是否包含不适宜公开传播的内容？
-            3. 是否涉及暴力、歧视、人身攻击或违反法律法规的内容？
-
-            请直接回答"安全"或"不安全"。如果"不安全"，请重新生成一段符合规范的温和回复。
-            """;
-
-    public Mono<ReviewResult> review(String articleContent, String commentContent, String aiReply,
-                                      String modelName) {
-        String reviewPrompt = String.format(REVIEW_PROMPT_TEMPLATE,
-            truncate(articleContent, 2000),
+                请判断以上回复是否安全合规。""",
            truncate(commentContent, 500),
            truncate(aiReply, 500));

-        return aiFoundationClient.chat(reviewPrompt, modelName)
-            .map(this::parseSafetyResult)
+        // Stage 1: Safety check
+        return aiFoundationClient.classify(SAFETY_SYSTEM_PROMPT, userPrompt, SAFETY_CHOICES, modelName)
+            .flatMap(safetyResult -> {
+                if (UNSAFE.equals(safetyResult)) {
+                    log.warn("[Review] Content is UNSAFE");
+                    return Mono.just(new ReviewResult(0, "FAIL", "内容安全审核不通过"));
+                }
+                if (!SAFE.equals(safetyResult)) {
+                    log.warn("[Review] Unexpected safety result: {}, treating as unsafe", safetyResult);
+                    return Mono.just(new ReviewResult(0, "FAIL", "内容安全审核结果异常"));
+                }
+                // Stage 2: Quality rating (only for safe content)
+                return rateQuality(commentContent, aiReply, modelName);
+            })
            .defaultIfEmpty(new ReviewResult(100, "PASS", "审核无响应，自动通过"))
            .onErrorResume(e -> {
-                log.warn("Review failed, auto-passing: {}", e.getMessage());
+                log.warn("[Review] Review failed, auto-passing: {}", e.getMessage());
                return Mono.just(new ReviewResult(100, "PASS", "审核服务异常，自动通过"));
            });
    }

-    private ReviewResult parseSafetyResult(String response) {
-        if (response == null || response.isBlank()) {
-            return new ReviewResult(100, "PASS", "审核无响应，自动通过");
+    /**
+     * Rate the quality of a safe AI reply (1-5) and map to a 0-100 score.
+     */
+    private Mono<ReviewResult> rateQuality(String commentContent, String aiReply, String modelName) {
+        String qualityPrompt = String.format("""
+                评论：
+                %s
+
+                回复：
+                %s
+
+                请对以上回复进行质量评分（1-5分）。""",
+            truncate(commentContent, 500),
+            truncate(aiReply, 500));
+
+        return aiFoundationClient.classify(QUALITY_SYSTEM_PROMPT, qualityPrompt, QUALITY_CHOICES, modelName)
+            .map(rating -> {
+                int score = mapRatingToScore(rating);
+                String reason = "安全通过，质量评分: " + rating + "/5";
+                log.info("[Review] Content is SAFE, quality rating: {}/5, score: {}", rating, score);
+                return new ReviewResult(score, "PASS", reason);
+            })
+            .defaultIfEmpty(new ReviewResult(85, "PASS", "安全通过，质量评分默认 4/5"))
+            .onErrorResume(e -> {
+                log.warn("[Review] Quality rating failed, defaulting to 85: {}", e.getMessage());
+                return Mono.just(new ReviewResult(85, "PASS", "安全通过，质量评分异常"));
+            });
    }
-        String trimmed = response.trim().toLowerCase();
-        if (trimmed.contains("不安全") || trimmed.contains("unsafe")) {
-            log.warn("AI Review: content is UNSAFE, response: {}", response);
-            return new ReviewResult(0, "FAIL", "内容安全审核不通过");
-        }
-        if (trimmed.contains("安全") || trimmed.contains("safe")) {
-            log.info("AI Review: content is SAFE");
-            return new ReviewResult(100, "PASS", "内容安全审核通过");
-        }
-        // If unclear response, default to pass
-        log.warn("AI Review: unclear response, auto-passing: {}", response);
-        return new ReviewResult(100, "PASS", "审核结果不明确，自动通过");
+
+    /**
+     * Map a 1-5 quality rating to a 0-100 score.
+     */
+    private int mapRatingToScore(String rating) {
+        return switch (rating) {
+            case RATING_EXCELLENT -> 100;
+            case RATING_GOOD -> 85;
+            case RATING_AVERAGE -> 70;
+            case RATING_POOR -> 50;
+            case RATING_BAD -> 30;
+            default -> 70; // default to average
+        };
    }

    private String truncate(String text, int maxLength) {
@@ -4,6 +4,8 @@ import lombok.extern.slf4j.Slf4j;
 import org.springframework.stereotype.Component;
 import reactor.core.publisher.Mono;

+import java.util.List;
+
@Component
@Slf4j
 public class SentimentService {
@@ -15,18 +17,44 @@ public class SentimentService {
    }

    public record SentimentResult(String sentiment, double confidence) {
+        public static final String VERY_POSITIVE = "VERY_POSITIVE";
        public static final String POSITIVE = "POSITIVE";
        public static final String NEUTRAL = "NEUTRAL";
        public static final String NEGATIVE = "NEGATIVE";
+        public static final String VERY_NEGATIVE = "VERY_NEGATIVE";
    }

-    public Mono<SentimentResult> analyzeSentiment(String commentContent, String modelName) {
-        String prompt = buildSentimentPrompt(commentContent);
+    private static final List<String> CHOICES = List.of(
+        SentimentResult.VERY_POSITIVE, SentimentResult.POSITIVE,
+        SentimentResult.NEUTRAL, SentimentResult.NEGATIVE,
+        SentimentResult.VERY_NEGATIVE
+    );

-        return aiFoundationClient.chat(prompt, modelName)
-            .map(response -> {
-                String sentiment = parseSentiment(response);
-                return new SentimentResult(sentiment, 1.0);
+    /**
+     * Analyze sentiment using AI Foundation structured output
+     * ({@code OutputSpec.choice}) for reliable classification.
+     */
+    public Mono<SentimentResult> analyzeSentiment(String commentContent, String modelName) {
+        String systemPrompt = "你是一个专业的情感分析助手。请根据以下标准分析评论的情感倾向：\n"
+            + "\n"
+            + "- VERY_POSITIVE：非常正面，包含强烈的感谢、赞美或认同（如\"太棒了\"、\"非常感谢\"、\"写得太好了\"）\n"
+            + "- POSITIVE：正面，友好、肯定或支持的态度（如\"不错\"、\"学习了\"、\"支持\"）\n"
+            + "- NEUTRAL：中性，提问、讨论、陈述事实，无明显情感倾向（如\"请问...\"、\"这个怎么用\"、\"我觉得\"）\n"
+            + "- NEGATIVE：负面，不满、质疑或批评（如\"不好用\"、\"有问题\"、\"不太行\"）\n"
+            + "- VERY_NEGATIVE：非常负面，攻击、辱骂或极端负面情绪（如\"垃圾\"、\"骗子\"、\"太差了\"）\n"
+            + "\n"
+            + "只返回 VERY_POSITIVE、POSITIVE、NEUTRAL、NEGATIVE 或 VERY_NEGATIVE 之一。";
+        String userPrompt = "分析以下评论的情感倾向：\n\n" + commentContent;
+
+        return aiFoundationClient.classify(systemPrompt, userPrompt, CHOICES, modelName)
+            .map(sentiment -> {
+                String upper = sentiment.toUpperCase();
+                // Validate against known choices; default to NEUTRAL if unexpected
+                if (!CHOICES.contains(upper)) {
+                    log.warn("[Sentiment] Unexpected classification result: {}, defaulting to NEUTRAL", sentiment);
+                    return new SentimentResult(SentimentResult.NEUTRAL, 0.0);
+                }
+                return new SentimentResult(upper, 1.0);
            })
            .onErrorResume(e -> {
                log.warn("[Sentiment] Failed to analyze sentiment, defaulting to NEUTRAL: {}", e.getMessage());
@@ -34,16 +62,4 @@ public class SentimentService {
            })
            .defaultIfEmpty(new SentimentResult(SentimentResult.NEUTRAL, 0.0));
    }
-
-    private String buildSentimentPrompt(String commentContent) {
-        return "请分析以下评论的情感倾向。只回复一个词：POSITIVE（正面）、NEUTRAL（中性）或 NEGATIVE（负面）。\n\n评论内容：\n" + commentContent;
-    }
-
-    private String parseSentiment(String response) {
-        if (response == null || response.isBlank()) return SentimentResult.NEUTRAL;
-        String upper = response.trim().toUpperCase();
-        if (upper.contains("POSITIVE")) return SentimentResult.POSITIVE;
-        if (upper.contains("NEGATIVE")) return SentimentResult.NEGATIVE;
-        return SentimentResult.NEUTRAL;
-    }
 }
@@ -0,0 +1,101 @@
+package top.nxxy335.commentaiautopilot.service;
+
+import lombok.RequiredArgsConstructor;
+import lombok.extern.slf4j.Slf4j;
+import org.springframework.data.domain.Sort;
+import org.springframework.stereotype.Component;
+import reactor.core.publisher.Mono;
+import run.halo.app.extension.ReactiveExtensionClient;
+import top.nxxy335.commentaiautopilot.extension.AiPersona;
+
+/**
+ * Service for checking wake words in comment content.
+ * A wake word is a prefix that triggers AI reply from a specific persona,
+ * even if the page hasn't enabled AI auto-reply.
+ */
+@Component
+@Slf4j
+@RequiredArgsConstructor
+public class WakeWordService {
+
+    private final ReactiveExtensionClient client;
+
+    /**
+     * Result of a wake word match.
+     *
+     * @param personaName the metadata.name of the matched persona
+     * @param wakeWord    the wake word that matched
+     * @param contentWithoutWakeWord the comment content with the wake word prefix removed
+     */
+    public record WakeWordMatch(String personaName, String wakeWord, String contentWithoutWakeWord) {}
+
+    /**
+     * Check if the given content starts with any persona's wake word.
+     * Returns the first matching persona's info, or empty if no match.
+     *
+     * @param content the comment/reply content to check
+     * @return WakeWordMatch if a wake word is found, or empty Mono
+     */
+    public Mono<WakeWordMatch> checkWakeWord(String content) {
+        if (content == null || content.isBlank()) {
+            return Mono.empty();
+        }
+
+        return client.list(AiPersona.class, null, null)
+            .filter(persona -> {
+                String wakeWord = persona.getSpec().getWakeWord();
+                return wakeWord != null && !wakeWord.isBlank() && content.startsWith(wakeWord);
+            })
+            .next()
+            .map(persona -> {
+                String wakeWord = persona.getSpec().getWakeWord();
+                String remaining = content.substring(wakeWord.length()).trim();
+                log.info("[WakeWord] Matched wake word '{}' for persona '{}'",
+                    wakeWord, persona.getSpec().getDisplayName());
+                return new WakeWordMatch(persona.getMetadata().getName(), wakeWord, remaining);
+            });
+    }
+
+    /**
+     * Blocking version for use in Reconciler (sync context).
+     * Checks if the given content starts with any persona's wake word.
+     *
+     * @param syncClient the blocking ExtensionClient
+     * @param content the comment/reply content to check
+     * @return WakeWordMatch if a wake word is found, or null
+     */
+    public WakeWordMatch checkWakeWordBlocking(run.halo.app.extension.ExtensionClient syncClient, String content) {
+        if (content == null || content.isBlank()) {
+            log.info("[WakeWord] Content is null or blank, skipping");
+            return null;
+        }
+
+        String trimmedContent = content.trim();
+        var personas = syncClient.listAll(AiPersona.class, null, Sort.unsorted());
+        log.info("[WakeWord] Checking {} personas against content: '{}'", personas.size(),
+            trimmedContent.length() > 50 ? trimmedContent.substring(0, 50) + "..." : trimmedContent);
+
+        for (var persona : personas) {
+            String wakeWord = persona.getSpec().getWakeWord();
+            if (wakeWord == null || wakeWord.isBlank()) {
+                log.info("[WakeWord] Persona '{}' has no wakeWord, skipping", persona.getSpec().getDisplayName());
+                continue;
+            }
+            String trimmedWakeWord = wakeWord.trim();
+            log.info("[WakeWord] Checking persona '{}' with wakeWord '{}' against content starting with '{}'",
+                persona.getSpec().getDisplayName(), trimmedWakeWord,
+                trimmedContent.length() >= trimmedWakeWord.length()
+                    ? trimmedContent.substring(0, trimmedWakeWord.length()) : trimmedContent);
+
+            if (trimmedContent.startsWith(trimmedWakeWord)) {
+                String remaining = trimmedContent.substring(trimmedWakeWord.length()).trim();
+                log.info("[WakeWord] MATCHED! wakeWord='{}' for persona '{}', remaining content: '{}'",
+                    trimmedWakeWord, persona.getSpec().getDisplayName(),
+                    remaining.length() > 30 ? remaining.substring(0, 30) + "..." : remaining);
+                return new WakeWordMatch(persona.getMetadata().getName(), trimmedWakeWord, remaining);
+            }
+        }
+        log.info("[WakeWord] No wake word matched");
+        return null;
+    }
+}
@@ -0,0 +1,41 @@
+package top.nxxy335.commentaiautopilot.util;
+
+import lombok.extern.slf4j.Slf4j;
+
+import java.nio.charset.StandardCharsets;
+import java.security.MessageDigest;
+
+/**
+ * Utility for generating Gravatar/Cravatar avatar URLs from email addresses.
+ */
+@Slf4j
+public class GravatarUtil {
+
+    private static final String CRAVATAR_BASE_URL = "https://cn.cravatar.com/avatar/";
+
+    private GravatarUtil() {}
+
+    /**
+     * Generate Cravatar URL from email address using SHA-256 hash.
+     *
+     * @param email the email address
+     * @return the avatar URL, or empty string if generation fails
+     */
+    public static String generateUrl(String email) {
+        if (email == null || email.isBlank()) {
+            return "";
+        }
+        try {
+            var digest = MessageDigest.getInstance("SHA-256");
+            var hashBytes = digest.digest(email.trim().toLowerCase().getBytes(StandardCharsets.UTF_8));
+            var hexString = new StringBuilder();
+            for (byte b : hashBytes) {
+                hexString.append(String.format("%02x", b));
+            }
+            return CRAVATAR_BASE_URL + hexString;
+        } catch (Exception e) {
+            log.error("Failed to generate Gravatar URL: {}", e.getMessage());
+            return "";
+        }
+    }
+}
@@ -40,6 +40,16 @@ spec:
          label: 评论者黑名单
          help: "输入评论者显示名称或邮箱，多个用逗号分隔。支持正则表达式，以 regex: 开头，如 regex:^spam.*"
          value: ""
+        - $formkit: switch
+          name: preFilterEnabled
+          label: 启用前置过滤
+          help: "AI回复前检测评论合规性，拦截广告/辱骂/敏感内容，节省Token"
+          value: true
+        - $formkit: switch
+          name: preFilterPendingOnViolation
+          label: 违规评论设为待审核
+          help: "检测到违规评论时自动取消通过，需人工审核"
+          value: true
    - group: model
      label: 模型设置
      formSchema:
@@ -54,7 +64,7 @@ spec:
        - $formkit: textarea
          name: customPromptTemplate
          label: 自定义Prompt模板
-          value: "{{persona_prompt}}\n\n{{safety_prompt}}\n\n【语言要求】请用评论所使用的语言回复。如果评论是英文，请用英文回复；如果是中文，请用中文回复；如果是日文，请用日文回复；以此类推。\n\n请回复以下评论。注意：\n- 回复长度应与评论长度匹配，简短问候简短回复\n- 不要复述或总结文章内容\n- 自然对话，不要写小作文\n- 只有评论涉及具体内容时才针对性回应\n\n文章（仅供理解上下文，不要复述）：\n{{article}}\n\n评论：\n{{comment}}"
+          value: "{{persona_prompt}}\n\n{{safety_prompt}}\n\n【语言要求】请用评论所使用的语言回复。如果评论是英文，请用英文回复；如果是中文，请用中文回复；如果是日文，请用日文回复；以此类推。\n\n请回复以下评论。注意：\n- 回复长度应与评论长度匹配，简短问候简短回复\n- 不要复述或总结文章内容\n- 自然对话，不要写小作文\n- 只有评论涉及具体内容时才针对性回应\n\n文章（仅供理解上下文，不要复述）：\n{{article}}\n\n{{conversation_history}}\n评论：\n{{comment}}"
        - $formkit: select
          name: enabledPresets
          label: 启用预设
@@ -5,9 +5,17 @@ kind: Plugin
 metadata:
  # The name defines how the plugin is invoked, A unique name
  name: comment-ai-autopilot
+  annotations:
+    # Recommend installing AI Foundation from the app store after installing this plugin
+    # https://www.halo.run/store/apps/app-acslk9nu
+    "store.halo.run/recommended-apps": '["app-acslk9nu"]'
 spec:
  enabled: true
  requires: ">=2.25.0"
+  pluginDependencies:
+    # Optional dependency: plugin still loads without AI Foundation,
+    # but AI features require it to be installed and enabled.
+    ai-foundation?: "*"
  author:
    name: 暖心向阳335
    website: https://nxxy335.top
@@ -22,4 +30,4 @@ spec:
      url: "https://github.com/sunny-335/plugin-comment-ai-autopilot/blob/main/LICENSE"
  settingName: "comment-ai-autopilot-settings"
  configMapName: "comment-ai-autopilot-configmap"
-  version: "1.0.0-beta.1"
+  version: "1.2.1"
@@ -5,16 +5,33 @@ plugins {

 group 'top.nxxy335.commentaiautopilot.ui'

-// Fix Gradle 9.x compatibility with pnpm symlinks
-tasks.named('pnpmInstall') {
-    doNotTrackState("pnpm symlinks are not compatible with Gradle state tracking")
+// Use system pnpm directly — avoids Windows exit code 268435659
+// caused by Gradle Worker Daemon / node-gradle downloading pnpm on Windows
+node {
+    download = false
 }

-tasks.register('pnpmBuild', PnpmTask) {
+// Skip built-in pnpm tasks (they fail on Windows), replace with Exec-based tasks
+tasks.named('pnpmSetup').configure { enabled = false }
+tasks.named('pnpmInstall').configure { enabled = false }
+
+// Cross-platform: use 'cmd /c' on Windows, direct 'pnpm' on Linux/macOS
+def isWindows = System.properties['os.name'].toLowerCase().contains('windows')
+def pnpmCmd = isWindows ? ['cmd', '/c', 'pnpm'] : ['pnpm']
+
+tasks.register('uiInstall', Exec) {
+    group = 'build'
+    description = 'Install UI dependencies using system pnpm'
+    workingDir layout.projectDirectory
+    commandLine(pnpmCmd + ['install'])
+}
+
+tasks.register('uiBuild', Exec) {
    group = 'build'
    description = 'Build the UI project using pnpm'
-    args = ['build']
-    dependsOn tasks.named('pnpmInstall')
+    workingDir layout.projectDirectory
+    commandLine(pnpmCmd + ['run', 'build'])
+    dependsOn uiInstall
    inputs.dir(layout.projectDirectory.dir('src'))
    inputs.files(fileTree(
        dir: layout.projectDirectory,
@@ -22,17 +39,18 @@ tasks.register('pnpmBuild', PnpmTask) {
    outputs.dir(layout.buildDirectory.dir('dist'))
 }

-tasks.register('pnpmCheck', PnpmTask) {
+tasks.register('uiCheck', Exec) {
    group = 'verification'
    description = 'Run unit tests for the UI project using pnpm'
-    args = ['test:unit']
-    dependsOn tasks.named('pnpmInstall')
+    workingDir layout.projectDirectory
+    commandLine(pnpmCmd + ['run', 'test:unit'])
+    dependsOn uiInstall
 }

 tasks.named('check') {
-    dependsOn tasks.named('pnpmCheck')
+    dependsOn tasks.named('uiCheck')
 }

 tasks.named('assemble') {
-    dependsOn tasks.named('pnpmBuild')
+    dependsOn tasks.named('uiBuild')
 }
@@ -0,0 +1,29 @@
+/**
+ * 统一 API 路径常量，避免硬编码散布在多个 Vue 文件中。
+ * 修改 API 路径只需在此处更新。
+ */
+export const API_BASE = '/apis/console.api.comment-ai-autopilot.nxxy335.top/v1alpha1'
+
+// ===== 日志相关 =====
+export const API_REPLIES = `${API_BASE}/replies`
+export const apiReply = (name: string) => `${API_REPLIES}/${name}`
+export const apiReplyAction = (name: string, action: string) => `${API_REPLIES}/${name}/${action}`
+export const API_BATCH_APPROVE = `${API_REPLIES}/batch-approve`
+export const API_BATCH_REJECT = `${API_REPLIES}/batch-reject`
+export const API_BATCH_DELETE = `${API_REPLIES}/batch-delete`
+export const apiConversation = (commentId: string) => `${API_BASE}/conversation/${commentId}`
+
+// ===== 概览相关 =====
+export const API_STATS = `${API_BASE}/stats`
+export const API_PERSONAS = `${API_BASE}/personas`
+export const apiPersona = (name: string) => `${API_PERSONAS}/${name}`
+export const API_HEALTH = `${API_BASE}/health`
+
+// ===== 设置相关 =====
+export const API_EXPORT = `${API_BASE}/export`
+export const API_IMPORT = `${API_BASE}/import`
+export const API_COMMENTERS = `${API_BASE}/commenters`
+export const API_CLEANUP = `${API_BASE}/cleanup`
+
+// ===== 评论触发 =====
+export const apiCommentTrigger = (commentName: string) => `${API_BASE}/comments/${commentName}/trigger`
@@ -0,0 +1,14 @@
+export async function computeGravatarHash(email: string): Promise<string> {
+  if (!email || email.trim() === '') return ''
+  const normalizedEmail = email.trim().toLowerCase()
+  const encoder = new TextEncoder()
+  const data = encoder.encode(normalizedEmail)
+  const hashBuffer = await crypto.subtle.digest('SHA-256', data)
+  const hashArray = Array.from(new Uint8Array(hashBuffer))
+  return hashArray.map(b => b.toString(16).padStart(2, '0')).join('')
+}
+
+export function getGravatarUrl(hash: string): string {
+  if (!hash) return ''
+  return `https://cn.cravatar.com/avatar/${hash}`
+}
@@ -108,152 +108,11 @@
        </VCard>
      </div>

-      <!-- Middle: Sentiment + Trend -->
-      <div class="grid grid-cols-1 gap-4 mt-4 lg:grid-cols-2">
-        <!-- Sentiment Distribution -->
-        <VCard :body-class="['!p-5']">
-          <h3 class="text-sm font-medium text-gray-500 mb-4">情感分布</h3>
-          <div class="space-y-3">
-            <div class="flex items-center gap-3">
-              <div class="w-2 h-2 rounded-full bg-green-500 shrink-0"></div>
-              <div class="flex-1 min-w-0">
-                <div class="flex items-center justify-between text-sm">
-                  <span class="text-gray-700">正面</span>
-                  <span class="font-medium text-green-600">{{ stats?.sentimentDistribution?.POSITIVE || 0 }}</span>
-                </div>
-                <div class="mt-1 h-1.5 bg-gray-100 rounded-full overflow-hidden">
-                  <div
-                    class="h-full bg-green-400 rounded-full transition-all duration-500"
-                    :style="{ width: getSentimentPercent('POSITIVE') + '%' }"
-                  ></div>
-                </div>
-              </div>
-            </div>
-            <div class="flex items-center gap-3">
-              <div class="w-2 h-2 rounded-full bg-gray-400 shrink-0"></div>
-              <div class="flex-1 min-w-0">
-                <div class="flex items-center justify-between text-sm">
-                  <span class="text-gray-700">中性</span>
-                  <span class="font-medium text-gray-600">{{ stats?.sentimentDistribution?.NEUTRAL || 0 }}</span>
-                </div>
-                <div class="mt-1 h-1.5 bg-gray-100 rounded-full overflow-hidden">
-                  <div
-                    class="h-full bg-gray-400 rounded-full transition-all duration-500"
-                    :style="{ width: getSentimentPercent('NEUTRAL') + '%' }"
-                  ></div>
-                </div>
-              </div>
-            </div>
-            <div class="flex items-center gap-3">
-              <div class="w-2 h-2 rounded-full bg-red-500 shrink-0"></div>
-              <div class="flex-1 min-w-0">
-                <div class="flex items-center justify-between text-sm">
-                  <span class="text-gray-700">负面</span>
-                  <span class="font-medium text-red-500">{{ stats?.sentimentDistribution?.NEGATIVE || 0 }}</span>
-                </div>
-                <div class="mt-1 h-1.5 bg-gray-100 rounded-full overflow-hidden">
-                  <div
-                    class="h-full bg-red-400 rounded-full transition-all duration-500"
-                    :style="{ width: getSentimentPercent('NEGATIVE') + '%' }"
-                  ></div>
-                </div>
-              </div>
-            </div>
-            <div class="flex items-center gap-3">
-              <div class="w-2 h-2 rounded-full bg-gray-300 shrink-0"></div>
-              <div class="flex-1 min-w-0">
-                <div class="flex items-center justify-between text-sm">
-                  <span class="text-gray-700">未知</span>
-                  <span class="font-medium text-gray-400">{{ stats?.sentimentDistribution?.UNKNOWN || 0 }}</span>
-                </div>
-                <div class="mt-1 h-1.5 bg-gray-100 rounded-full overflow-hidden">
-                  <div
-                    class="h-full bg-gray-300 rounded-full transition-all duration-500"
-                    :style="{ width: getSentimentPercent('UNKNOWN') + '%' }"
-                  ></div>
-                </div>
-              </div>
-            </div>
-          </div>
-        </VCard>
-
-        <!-- Daily Trend -->
-        <VCard :body-class="['!p-5']">
-          <div class="flex items-center justify-between mb-4">
-            <h3 class="text-sm font-medium text-gray-500">近7日回复趋势</h3>
-            <div class="inline-flex rounded-md border border-gray-200 overflow-hidden">
-              <button
-                class="px-2.5 py-1 text-xs transition-colors"
-                :class="range === '7' ? 'bg-blue-500 text-white' : 'bg-white text-gray-600 hover:bg-gray-50'"
-                @click="range = '7'"
-              >
-                7天
-              </button>
-              <button
-                class="px-2.5 py-1 text-xs border-l border-gray-200 transition-colors"
-                :class="range === '30' ? 'bg-blue-500 text-white' : 'bg-white text-gray-600 hover:bg-gray-50'"
-                @click="range = '30'"
-              >
-                30天
-              </button>
-              <button
-                class="px-2.5 py-1 text-xs border-l border-gray-200 transition-colors"
-                :class="range === 'all' ? 'bg-blue-500 text-white' : 'bg-white text-gray-600 hover:bg-gray-50'"
-                @click="range = 'all'"
-              >
-                全部
-              </button>
-            </div>
-          </div>
-          <div v-if="stats?.dailyTrend?.length" class="flex items-end gap-3" style="height: 160px">
-            <div
-              v-for="day in stats.dailyTrend"
-              :key="day.date"
-              class="flex-1 flex flex-col items-center justify-end h-full"
-            >
-              <div class="text-xs text-gray-500 mb-1 font-medium">{{ day.count }}</div>
-              <div
-                class="w-full rounded-t-md transition-all duration-500"
-                :class="day.count > 0 ? 'bg-gradient-to-t from-blue-500 to-blue-400' : 'bg-gray-100'"
-                :style="{ height: getTrendBarHeight(day.count) + 'px' }"
-              ></div>
-              <div class="text-[10px] text-gray-400 mt-2 whitespace-nowrap">{{ formatTrendDate(day.date) }}</div>
-            </div>
-          </div>
-          <div v-else class="flex items-center justify-center text-sm text-gray-400" style="height: 160px">
-            暂无数据
-          </div>
-        </VCard>
-      </div>
-
-      <!-- Bottom: Score + Quick Actions -->
-      <div class="grid grid-cols-1 gap-4 mt-4 sm:grid-cols-2">
-        <!-- Avg Score -->
-        <VCard :body-class="['!p-5']">
-          <h3 class="text-sm font-medium text-gray-500 mb-3">平均审核评分</h3>
-          <div class="flex items-center gap-4">
-            <div class="text-4xl font-bold" :class="scoreColor">{{ stats?.avgScore?.toFixed(1) || '0.0' }}</div>
-            <div class="flex-1">
-              <div class="h-3 bg-gray-100 rounded-full overflow-hidden">
-                <div
-                  class="h-full rounded-full transition-all duration-500"
-                  :class="scoreBarColor"
-                  :style="{ width: (stats?.avgScore || 0) * 10 + '%' }"
-                ></div>
-              </div>
-              <div class="flex justify-between text-[10px] text-gray-300 mt-1">
-                <span>0</span>
-                <span>5</span>
-                <span>10</span>
-              </div>
-            </div>
-          </div>
-        </VCard>
-
      <!-- Quick Actions -->
+      <div class="mt-4">
        <VCard :body-class="['!p-5']">
          <h3 class="text-sm font-medium text-gray-500 mb-3">快捷操作</h3>
-          <div class="grid grid-cols-2 gap-2">
+          <div class="grid grid-cols-1 gap-2 sm:grid-cols-3">
            <button
              class="flex items-center gap-2 px-3 py-2.5 rounded-md bg-gray-50 hover:bg-gray-100 transition-colors text-sm text-gray-700"
              @click="$router.push({ name: 'CommentAiAutopilotLogs' })"
@@ -273,15 +132,6 @@
              </svg>
              插件设置
            </button>
-            <button
-              class="flex items-center gap-2 px-3 py-2.5 rounded-md bg-gray-50 hover:bg-gray-100 transition-colors text-sm text-gray-700"
-              @click="$router.push({ name: 'CommentAiAutopilotSettings' })"
-            >
-              <svg class="w-4 h-4 text-gray-400" fill="none" stroke="currentColor" viewBox="0 0 24 24">
-                <path stroke-linecap="round" stroke-linejoin="round" stroke-width="2" d="M16 7a4 4 0 11-8 0 4 4 0 018 0zM12 14a7 7 0 00-7 7h14a7 7 0 00-7-7z" />
-              </svg>
-              AI角色
-            </button>
            <button
              class="flex items-center gap-2 px-3 py-2.5 rounded-md bg-gray-50 hover:bg-gray-100 transition-colors text-sm text-gray-700"
              @click="refreshData"
@@ -299,24 +149,16 @@
 </template>

 <script setup lang="ts">
-import { ref, computed, onMounted, watch } from "vue"
+import { ref, computed, onMounted } from "vue"
 import { axiosInstance } from "@halo-dev/api-client"
 import { VPageHeader, VButton, VCard, Toast } from "@halo-dev/components"
 import { IconPlug } from "@halo-dev/components"

-interface DailyCount {
-  date: string
-  count: number
-}
-
 interface StatsResponse {
  total: number
  passCount: number
  failCount: number
  reviewingCount: number
-  avgScore: number
-  sentimentDistribution: Record<string, number>
-  dailyTrend: DailyCount[]
 }

 interface PersonaResponse {
@@ -331,7 +173,6 @@ interface HealthResponse {

 const stats = ref<StatsResponse | null>(null)
 const persona = ref<PersonaResponse | null>(null)
-const range = ref("7")
 const health = ref<HealthResponse | null>(null)
 const healthVisible = ref(true)

@@ -340,24 +181,10 @@ const passRate = computed(() => {
  return Math.round((stats.value.passCount / stats.value.total) * 100)
 })

-const scoreColor = computed(() => {
-  const score = stats.value?.avgScore || 0
-  if (score >= 7) return "text-green-600"
-  if (score >= 4) return "text-amber-500"
-  return "text-red-500"
-})
-
-const scoreBarColor = computed(() => {
-  const score = stats.value?.avgScore || 0
-  if (score >= 7) return "bg-gradient-to-r from-green-400 to-green-500"
-  if (score >= 4) return "bg-gradient-to-r from-amber-400 to-amber-500"
-  return "bg-gradient-to-r from-red-400 to-red-500"
-})
-
 const fetchStats = async () => {
  try {
    const { data } = await axiosInstance.get(
-      `/apis/console.api.comment-ai-autopilot.nxxy335.top/v1alpha1/stats?range=${range.value}`,
+      `/apis/console.api.comment-ai-autopilot.nxxy335.top/v1alpha1/stats?range=7`,
    )
    stats.value = data
  } catch (e) {
@@ -421,34 +248,33 @@ const openSettings = () => {
  window.location.href = "/console/comment-ai-autopilot/settings"
 }

-const getSentimentPercent = (sentiment: string): number => {
-  const dist = stats.value?.sentimentDistribution
-  if (!dist) return 0
-  const total = Object.values(dist).reduce((a, b) => a + b, 0)
-  if (total === 0) return 0
-  return Math.round(((dist[sentiment] || 0) / total) * 100)
-}
-
-const getTrendBarHeight = (count: number): number => {
-  const trend = stats.value?.dailyTrend
-  if (!trend || trend.length === 0) return 0
-  const max = Math.max(...trend.map(d => d.count), 1)
-  return Math.max((count / max) * 100, count > 0 ? 8 : 4)
-}
-
-const formatTrendDate = (dateStr: string): string => {
-  if (!dateStr) return ''
-  const parts = dateStr.split('-')
-  return parts.length >= 3 ? `${parts[1]}/${parts[2]}` : dateStr
-}
-
-watch(range, () => {
-  fetchStats()
-})
-
 onMounted(() => {
  fetchStats()
  fetchPersona()
  fetchHealth()
 })
 </script>
+
+<style scoped>
+.line-clamp-2 {
+  display: -webkit-box;
+  -webkit-line-clamp: 2;
+  -webkit-box-orient: vertical;
+  overflow: hidden;
+}
+
+/* Mobile responsive */
+@media (max-width: 640px) {
+  .comment-ai-autopilot-home :deep(.page-header) {
+    flex-wrap: wrap;
+    gap: 8px;
+  }
+  .comment-ai-autopilot-home :deep(.page-header-actions) {
+    width: 100%;
+  }
+  .comment-ai-autopilot-home :deep(.page-header-actions .space-y-2) {
+    flex-direction: row;
+    width: 100%;
+  }
+}
+</style>
@@ -1,300 +1,176 @@
 <template>
-  <div class="comment-ai-autopilot-logs">
+  <div class="logs-container">
    <VPageHeader title="AI回复日志">
-      <template #icon>
-        <IconPlug class="mr-2 self-center" />
-      </template>
+      <template #icon><IconPlug class="header-icon" /></template>
      <template #actions>
-        <VButton @click="$router.push({ name: 'CommentAiAutopilot' })"> 返回概览 </VButton>
+        <VButton @click="$router.push({ name: 'CommentAiAutopilot' })">返回概览</VButton>
      </template>
    </VPageHeader>

-    <!-- Batch Operation Toolbar -->
-    <div v-if="selectedNames.size > 0" class="mx-4 mt-2 flex items-center gap-3 bg-blue-50 border border-blue-200 rounded-lg px-4 py-2.5">
-      <span class="text-sm text-blue-700">已选择 {{ selectedNames.size }} 项</span>
-      <button
-        class="text-xs px-3 py-1 rounded bg-green-600 text-white hover:bg-green-700 transition-colors"
-        @click="batchApprove"
-      >
-        批量通过
-      </button>
-      <button
-        class="text-xs px-3 py-1 rounded bg-orange-500 text-white hover:bg-orange-600 transition-colors"
-        @click="batchReject"
-      >
-        批量拒绝
-      </button>
-      <button
-        class="text-xs px-3 py-1 rounded bg-red-600 text-white hover:bg-red-700 transition-colors"
-        @click="batchDelete"
-      >
-        批量删除
-      </button>
-      <button
-        class="text-xs text-gray-500 hover:text-gray-700 ml-auto"
-        @click="selectedNames.clear(); selectAll = false"
-      >
-        取消选择
-      </button>
+    <!-- 批量操作工具栏 -->
+    <div v-if="selectedNames.size > 0" class="toolbar batch-toolbar">
+      <span class="batch-text">已选择 {{ selectedNames.size }} 项</span>
+      <div class="batch-actions">
+        <button class="btn-batch btn-pass" @click="batchApprove">批量通过</button>
+        <button class="btn-batch btn-reject" @click="batchReject">批量拒绝</button>
+        <button class="btn-batch btn-delete" @click="batchDelete">批量删除</button>
+        <button class="btn-batch btn-cancel" @click="selectedNames.clear(); selectAll = false">取消选择</button>
+      </div>
    </div>

-    <!-- Filter Bar -->
-    <div class="mx-4 mt-2 flex items-center gap-3">
-      <select
-        v-model="filterStatus"
-        class="rounded-md border border-gray-300 px-3 py-1.5 text-sm focus:border-blue-500 focus:outline-none focus:ring-1 focus:ring-blue-500"
-      >
+    <!-- 过滤工具栏 -->
+    <div class="toolbar filter-toolbar">
+      <select v-model="filterStatus" class="filter-select">
        <option value="">全部状态</option>
        <option value="PASS">通过</option>
        <option value="FAIL">失败</option>
        <option value="PENDING">待审核</option>
        <option value="REJECTED">已拒绝</option>
+        <option value="FILTERED">已拦截</option>
+        <option value="FALSE_POSITIVE">误报通过</option>
      </select>
-      <select
-        v-model="filterSentiment"
-        class="rounded-md border border-gray-300 px-3 py-1.5 text-sm focus:border-blue-500 focus:outline-none focus:ring-1 focus:ring-blue-500"
-      >
+      <select v-model="filterSentiment" class="filter-select">
        <option value="">全部情感</option>
+        <option value="VERY_POSITIVE">非常正面</option>
        <option value="POSITIVE">正面</option>
        <option value="NEUTRAL">中性</option>
        <option value="NEGATIVE">负面</option>
+        <option value="VERY_NEGATIVE">非常负面</option>
      </select>
-      <input
-        v-model="filterKeyword"
-        type="text"
-        placeholder="搜索回复内容..."
-        class="rounded-md border border-gray-300 px-3 py-1.5 text-sm focus:border-blue-500 focus:outline-none focus:ring-1 focus:ring-blue-500"
-      />
-      <button
-        class="text-xs text-gray-500 hover:text-gray-700"
-        @click="resetFilters"
-      >
-        重置
-      </button>
+      <input v-model="filterKeyword" type="text" placeholder="搜索回复内容..." class="filter-input" />
+      <button class="btn-reset" @click="resetFilters">重置</button>
    </div>

-    <div class="m-4">
+    <!-- 列表区 -->
+    <div class="list-area">
      <VLoading v-if="loading" />
+      <div v-else-if="replies.length === 0" class="empty-state">暂无记录</div>
      
-      <div v-else-if="replies.length === 0" class="flex flex-col items-center justify-center py-16 text-gray-400">
-        <svg class="w-12 h-12 mb-3" fill="none" stroke="currentColor" viewBox="0 0 24 24">
-          <path stroke-linecap="round" stroke-linejoin="round" stroke-width="1.5" d="M9 12h6m-6 4h6m2 5H7a2 2 0 01-2-2V5a2 2 0 012-2h5.586a1 1 0 01.707.293l5.414 5.414a1 1 0 01.293.707V19a2 2 0 01-2 2z" />
-        </svg>
-        <span>暂无AI回复记录</span>
+      <div v-else class="reply-list">
+        <div class="select-all-wrap">
+          <input type="checkbox" :checked="selectAll" @change="toggleSelectAll" />
+          <span>全选本页</span>
        </div>
        
-      <div v-else class="space-y-3">
-        <div v-if="replies.length > 0" class="flex items-center gap-2 mb-2 px-1">
-          <input
-            type="checkbox"
-            :checked="selectAll"
-            @change="toggleSelectAll"
-            class="h-4 w-4 rounded border-gray-300 text-blue-600 focus:ring-blue-500"
-          />
-          <span class="text-xs text-gray-500">全选</span>
+        <div v-for="reply in replies" :key="reply.metadata.name" class="reply-card">
+          <div class="card-main">
+            <input type="checkbox" :checked="selectedNames.has(reply.metadata.name)" @change="toggleSelect(reply.metadata.name)" />
+            <div class="card-content">
+              <div class="card-header">
+                <div class="tags-wrap">
+                  <span class="custom-tag" :class="'tag-' + reply.spec.status">{{ getStatusLabel(reply.spec.status) }}</span>
+                  <span class="custom-tag" :class="reply.spec.published ? 'tag-published' : 'tag-draft'">{{ reply.spec.published ? '已发布' : '未发布' }}</span>
+                  <span v-if="reply.spec.isAiConversation" class="custom-tag tag-conv">对话</span>
+                  <span v-if="reply.spec.sentiment" class="custom-tag" :class="'tag-' + reply.spec.sentiment">{{ getSentimentLabel(reply.spec.sentiment) }}</span>
                </div>
-        <div
-          v-for="reply in replies"
-          :key="reply.metadata.name"
-          class="bg-white rounded-lg border border-gray-200 overflow-hidden hover:shadow-sm transition-all"
-        >
-          <!-- Card body -->
-          <div class="p-4">
-            <div class="flex items-start gap-3">
-              <input
-                type="checkbox"
-                :checked="selectedNames.has(reply.metadata.name)"
-                @change="toggleSelect(reply.metadata.name)"
-                class="mt-1 h-4 w-4 rounded border-gray-300 text-blue-600 focus:ring-blue-500 shrink-0"
-              />
-              <div class="flex-1 min-w-0">
-                <!-- Row 1: Status tags + time -->
-                <div class="flex items-center justify-between mb-3">
-                  <div class="flex items-center gap-1.5 flex-wrap">
-                    <span
-                      class="inline-flex items-center px-2 py-0.5 rounded text-xs font-medium"
-                      :class="getStatusClass(reply.spec.status)"
-                    >
-                      {{ getStatusLabel(reply.spec.status) }}
-                    </span>
-                    <span
-                      class="inline-flex items-center px-2 py-0.5 rounded text-xs"
-                      :class="reply.spec.published ? 'bg-green-50 text-green-600' : 'bg-gray-50 text-gray-400'"
-                    >
-                      {{ reply.spec.published ? '已发布' : '未发布' }}
-                    </span>
-                    <span
-                      v-if="reply.spec.isAiConversation"
-                      class="inline-flex items-center px-2 py-0.5 rounded text-xs bg-blue-50 text-blue-600"
-                    >
-                      对话
-                    </span>
-                    <span
-                      v-if="reply.spec.sentiment"
-                      class="inline-flex items-center px-2 py-0.5 rounded text-xs font-medium"
-                      :class="getSentimentClass(reply.spec.sentiment)"
-                    >
-                      {{ getSentimentLabel(reply.spec.sentiment) }}
+                <span class="card-time">{{ formatDate(reply.metadata.creationTimestamp) }}</span>
+              </div>
+              <div class="card-text">{{ stripHtml(reply.spec.reply) || '(空)' }}</div>
+              <div v-if="reply.spec.status === 'FILTERED'" class="card-filter-reason">
+                <svg class="filter-icon" fill="currentColor" viewBox="0 0 20 20"><path fill-rule="evenodd" d="M10 18a8 8 0 100-16 8 8 0 000 16zM8.707 7.293a1 1 0 00-1.414 1.414L8.586 10l-1.293 1.293a1 1 0 101.414 1.414L10 11.414l1.293 1.293a1 1 0 001.414-1.414L11.414 10l1.293-1.293a1 1 0 00-1.414-1.414L10 8.586 8.707 7.293z" clip-rule="evenodd"/></svg>
+                <span class="filter-category" v-if="reply.spec.filterCategory">{{ reply.spec.filterCategory }}</span>
+                <span class="filter-detail">{{ reply.spec.filterReason || '未提供具体原因' }}</span>
+                <button class="btn-false-positive" @click="openFalsePositiveDialog(reply)">误报反馈</button>
+              </div>
+              <div v-if="reply.spec.status === 'FALSE_POSITIVE'" class="card-filter-reason">
+                <svg class="filter-icon fp-icon-ok" fill="currentColor" viewBox="0 0 20 20"><path fill-rule="evenodd" d="M10 18a8 8 0 100-16 8 8 0 000 16zm3.707-9.293a1 1 0 00-1.414-1.414L9 10.586 7.707 9.293a1 1 0 00-1.414 1.414l2 2a1 1 0 001.414 0l4-4z" clip-rule="evenodd"/></svg>
+                <span class="filter-category">误报</span>
+                <span class="filter-detail">{{ reply.spec.filterReason || '用户确认为误报' }}</span>
+                <button class="btn-trigger-ai" :disabled="triggerAiLoadingName === reply.metadata.name" @click="handleTriggerAiReply(reply)">
+                  <span v-if="triggerAiLoadingName === reply.metadata.name" class="fp-spinner"></span>
+                  触发AI回复
+                </button>
+              </div>
+            </div>
+          </div>
+          <div class="card-footer">
+            <div class="footer-info">
+              <span>评分: <strong>{{ reply.spec.score }}</strong></span>
+              <span v-if="reply.spec.postSlug">
+                关联: <a :href="getPostUrl(reply.spec.postSlug)" target="_blank" class="post-link">{{ reply.spec.postSlug }}</a>
              </span>
+              <span v-if="reply.spec.retryCount > 0" class="retry-text">重试 {{ reply.spec.retryCount }} 次</span>
            </div>
-                  <span class="text-xs text-gray-400 flex-shrink-0 ml-2">{{ formatDate(reply.metadata.creationTimestamp) }}</span>
-                </div>
-
-                <!-- Row 2: AI reply content (truncated) -->
-                <div class="text-sm text-gray-800 leading-relaxed break-words line-clamp-3 mt-2">
-                  {{ stripHtml(reply.spec.reply) || '(空)' }}
-                </div>
-              </div>
-            </div>
-          </div>
-
-          <!-- Card footer: meta info + actions -->
-          <div class="px-4 py-2.5 bg-gray-50 border-t border-gray-100 flex items-center justify-between">
-            <div class="flex items-center gap-4 text-xs text-gray-400">
-              <span>
-                评分 <span :class="getScoreClass(reply.spec.score)" class="font-medium text-gray-600">{{ reply.spec.score }}</span>
-              </span>
-              <span v-if="reply.spec.postSlug" class="flex items-center gap-1">
-                文章
-                <a
-                  :href="getPostUrl(reply.spec.postSlug)"
-                  target="_blank"
-                  rel="noopener noreferrer"
-                  class="text-blue-500 hover:text-blue-700 hover:underline"
-                >{{ reply.spec.postSlug }}</a>
-                <svg class="w-3 h-3" fill="none" stroke="currentColor" viewBox="0 0 24 24">
-                  <path stroke-linecap="round" stroke-linejoin="round" stroke-width="2" d="M10 6H6a2 2 0 00-2 2v10a2 2 0 002 2h10a2 2 0 002-2v-4M14 4h6m0 0v6m0-6L10 14" />
-                </svg>
-              </span>
-              <span v-if="reply.spec.retryCount > 0">
-                重试 <span class="text-gray-600">{{ reply.spec.retryCount }}</span> 次
-              </span>
-            </div>
-            <div class="flex items-center gap-2">
+            <div class="footer-actions">
              <template v-if="reply.spec.status === 'PASS' && !reply.spec.published">
-                <button
-                  class="inline-flex items-center gap-1 text-xs text-green-600 hover:text-green-800 transition-colors px-2 py-1 rounded hover:bg-green-50"
-                  @click="handleApprove(reply.metadata.name)"
-                >
-                  审核通过
-                </button>
-                <button
-                  class="inline-flex items-center gap-1 text-xs text-red-500 hover:text-red-700 transition-colors px-2 py-1 rounded hover:bg-red-50"
-                  @click="handleReject(reply.metadata.name)"
-                >
-                  拒绝
-                </button>
+                <button class="action-btn pass" @click="handleApprove(reply.metadata.name)">通过</button>
+                <button class="action-btn reject" @click="handleReject(reply.metadata.name)">拒绝</button>
              </template>
-              <button
-                class="inline-flex items-center gap-1 text-xs text-blue-500 hover:text-blue-700 transition-colors px-2 py-1 rounded hover:bg-blue-50"
-                @click="openConversation(reply)"
-              >
-                <svg class="w-3.5 h-3.5" fill="none" stroke="currentColor" viewBox="0 0 24 24">
-                  <path stroke-linecap="round" stroke-linejoin="round" stroke-width="2" d="M8 12h.01M12 12h.01M16 12h.01M21 12c0 4.418-4.03 8-9 8a9.863 9.863 0 01-4.255-.949L3 20l1.395-3.72C3.512 15.042 3 13.574 3 12c0-4.418 4.03-8 9-8s9 3.582 9 8z" />
-                </svg>
-                查看
-              </button>
-              <button
-                class="text-xs text-red-400 hover:text-red-600 transition-colors px-2 py-1 rounded hover:bg-red-50"
-                @click="handleDelete(reply.metadata.name)"
-              >
-                删除
-              </button>
+              <button class="action-btn view" @click="openConversation(reply)">查看对话</button>
+              <button class="action-btn delete" @click="handleDelete(reply.metadata.name)">删除</button>
            </div>
          </div>
        </div>
      </div>
      
-      <!-- Pagination -->
-      <div v-if="totalPages > 1" class="flex items-center justify-between mt-4 px-1">
-        <span class="text-xs text-gray-400">共 {{ total }} 条</span>
-        <div class="flex gap-2">
+      <div v-if="totalPages > 1" class="pagination">
+        <span>共 {{ total }} 条</span>
+        <div class="pagination-btns">
          <VButton size="sm" :disabled="page <= 1" @click="page--">上一页</VButton>
          <VButton size="sm" :disabled="page >= totalPages" @click="page++">下一页</VButton>
        </div>
      </div>
    </div>

-    <!-- Conversation Dialog -->
+    <!-- 完整对话弹窗 -->
    <teleport to="body">
-      <div
-        v-if="showDialog"
-        class="fixed inset-0 z-[9999] flex items-center justify-center"
-      >
-        <!-- Overlay -->
-        <div
-          class="absolute inset-0 bg-black/40"
-          @click="showDialog = false"
-        ></div>
-
-        <!-- Dialog -->
-        <div class="relative bg-white rounded-xl shadow-2xl w-full max-w-2xl max-h-[80vh] mx-4 flex flex-col overflow-hidden">
-          <!-- Dialog header -->
-          <div class="flex items-center justify-between px-5 py-4 border-b border-gray-100">
-            <div class="flex items-center gap-2">
-              <svg class="w-5 h-5 text-blue-500" fill="none" stroke="currentColor" viewBox="0 0 24 24">
-                <path stroke-linecap="round" stroke-linejoin="round" stroke-width="2" d="M8 12h.01M12 12h.01M16 12h.01M21 12c0 4.418-4.03 8-9 8a9.863 9.863 0 01-4.255-.949L3 20l1.395-3.72C3.512 15.042 3 13.574 3 12c0-4.418 4.03-8 9-8s9 3.582 9 8z" />
-              </svg>
-              <h3 class="text-base font-semibold text-gray-800">完整对话</h3>
+      <div v-if="showDialog" class="dialog-overlay" @click.self="showDialog = false">
+        <div class="dialog-box">
+          <div class="dialog-header">
+            <h3>对话上下文</h3>
+            <button class="close-btn" @click="showDialog = false"><svg fill="none" stroke="currentColor" viewBox="0 0 24 24" width="24" height="24"><path stroke-linecap="round" stroke-linejoin="round" stroke-width="2" d="M6 18L18 6M6 6l12 12"/></svg></button>
          </div>
-            <button
-              class="text-gray-400 hover:text-gray-600 transition-colors p-1 rounded-lg hover:bg-gray-100"
-              @click="showDialog = false"
-            >
-              <svg class="w-5 h-5" fill="none" stroke="currentColor" viewBox="0 0 24 24">
-                <path stroke-linecap="round" stroke-linejoin="round" stroke-width="2" d="M6 18L18 6M6 6l12 12" />
-              </svg>
-            </button>
-          </div>
-
-          <!-- Dialog body: conversation -->
-          <div class="flex-1 overflow-y-auto px-5 py-4 space-y-4">
+          <div class="dialog-body">
            <VLoading v-if="conversationLoading" />
+            <div v-else-if="conversationMessages.length === 0" class="empty-state">暂无内容</div>
+            <div v-else class="chat-container">
+              <div v-for="(msg, idx) in conversationMessages" :key="idx" class="chat-row" :class="msg.isAi ? 'row-ai' : 'row-user'">
+                <div class="chat-message">
+                  <div class="chat-owner">{{ msg.owner }}</div>
+                  <div class="chat-bubble" :class="msg.isAi ? 'bubble-ai' : 'bubble-user'">
                    
-            <div v-else-if="conversationMessages.length === 0" class="text-center text-gray-400 py-8">
-              暂无对话内容
+                    <!-- 现代化的精美引用框 (无左边框) -->
+                    <div v-if="msg.quoteOwner && msg.quoteContent" class="quote-box">
+                      <div class="quote-header">
+                        <svg class="quote-icon" fill="currentColor" viewBox="0 0 24 24"><path d="M10 9V5l-7 7 7 7v-4.1c5 0 8.5 1.6 11 5.1-1-5-4-10-11-11z"/></svg>
+                        <span class="quote-owner">{{ msg.quoteOwner }}</span>
+                      </div>
+                      <div class="quote-content">{{ truncateQuote(msg.quoteContent) }}</div>
                    </div>

-            <div v-else>
-              <div
-                v-for="(msg, idx) in conversationMessages"
-                :key="idx"
-                class="flex"
-                :class="msg.isAi ? 'justify-start' : 'justify-end'"
-              >
-                <div class="max-w-[80%]">
-                  <!-- Owner name -->
-                  <div
-                    class="text-xs mb-1"
-                    :class="msg.isAi ? 'text-blue-500' : 'text-gray-500'"
-                  >
-                    <span class="font-medium">{{ msg.owner }}</span>
+                    <div class="chat-text" v-html="renderContent(msg.content)"></div>
                  </div>
-                  <!-- Bubble -->
-                  <div
-                    class="rounded-2xl px-4 py-2.5 text-sm leading-relaxed break-words"
-                    :class="msg.isAi
-                      ? 'bg-blue-50 text-gray-800 rounded-tl-md'
-                      : 'bg-gray-100 text-gray-800 rounded-tr-md'"
-                    v-html="renderContent(msg.content)"
-                  ></div>
-                  <!-- Time -->
-                  <div class="text-[10px] text-gray-300 mt-1" :class="msg.isAi ? 'text-left' : 'text-right'">
-                    {{ formatDate(msg.time) }}
+                  <div class="chat-time">{{ formatDate(msg.time) }}</div>
                </div>
              </div>
            </div>
          </div>
        </div>
+      </div>
+    </teleport>

-          <!-- Dialog footer -->
-          <div class="px-5 py-3 border-t border-gray-100 flex justify-end">
-            <button
-              class="px-4 py-1.5 text-sm text-gray-600 bg-gray-100 hover:bg-gray-200 rounded-lg transition-colors"
-              @click="showDialog = false"
-            >
-              关闭
+    <!-- 误报反馈确认弹窗 -->
+    <teleport to="body">
+      <div v-if="showFalsePositiveDialog" class="dialog-overlay" @click.self="showFalsePositiveDialog = false">
+        <div class="dialog-box fp-dialog">
+          <div class="dialog-header">
+            <h3>确认为误报？</h3>
+            <button class="close-btn" @click="showFalsePositiveDialog = false"><svg fill="none" stroke="currentColor" viewBox="0 0 24 24" width="24" height="24"><path stroke-linecap="round" stroke-linejoin="round" stroke-width="2" d="M6 18L18 6M6 6l12 12"/></svg></button>
+          </div>
+          <div class="fp-dialog-body">
+            <p class="fp-desc">系统检测到该评论可能包含违规内容，但您认为这是正常表达。请选择处理方式：</p>
+            <div class="fp-actions">
+              <button class="fp-btn fp-btn-primary" :disabled="fpLoading" @click="handleFalsePositive('aiReply')">
+                <span v-if="fpLoading" class="fp-spinner"></span>
+                AI 回复
              </button>
+              <button class="fp-btn fp-btn-secondary" :disabled="fpLoading" @click="handleFalsePositive('approveOnly')">
+                仅通过
+              </button>
+              <button class="fp-btn fp-btn-ghost" :disabled="fpLoading" @click="showFalsePositiveDialog = false">
+                取消
+              </button>
+            </div>
          </div>
        </div>
      </div>
@@ -303,341 +179,228 @@
 </template>

 <script setup lang="ts">
-import { ref, onMounted, watch } from "vue"
+import { ref, onMounted, onUnmounted, watch } from "vue"
 import { axiosInstance } from "@halo-dev/api-client"
 import { VPageHeader, VButton, VLoading, Toast } from "@halo-dev/components"
 import { IconPlug } from "@halo-dev/components"

-interface AiCommentReplyItem {
-  metadata: {
-    name: string
-    creationTimestamp: string
-  }
-  spec: {
-    commentId: string
-    postId: string
-    postSlug: string
-    reply: string
-    score: number
-    status: string
-    retryCount: number
-    replyTo: string
-    isAiConversation: boolean
-    published: boolean
-    sentiment: string | null
-  }
-}
+interface AiCommentReplyItem { metadata: { name: string; creationTimestamp: string }; spec: any }
+interface ConversationMessage { type: string; owner: string; content: string; time: string; isAi: boolean; quoteOwner?: string; quoteContent?: string }

-interface ConversationMessage {
-  type: string
-  owner: string
-  content: string
-  time: string
-  isAi: boolean
-}
+const replies = ref<AiCommentReplyItem[]>([]); const loading = ref(false); const batchLoading = ref(false); const page = ref(1); const size = ref(20); const total = ref(0); const totalPages = ref(0);
+const selectedNames = ref<Set<string>>(new Set()); const selectAll = ref(false);
+const filterStatus = ref(""); const filterSentiment = ref(""); const filterKeyword = ref("");
+const showDialog = ref(false); const conversationLoading = ref(false); const conversationMessages = ref<ConversationMessage[]>([]);
+const showFalsePositiveDialog = ref(false); const falsePositiveTarget = ref<AiCommentReplyItem | null>(null); const fpLoading = ref(false);
+const triggerAiLoadingName = ref<string | null>(null);

-const replies = ref<AiCommentReplyItem[]>([])
-const loading = ref(false)
-const page = ref(1)
-const size = ref(20)
-const total = ref(0)
-const totalPages = ref(0)
-
-// Selection state
-const selectedNames = ref<Set<string>>(new Set())
-const selectAll = ref(false)
-
-// Filter state
-const filterStatus = ref("")
-const filterSentiment = ref("")
-const filterKeyword = ref("")
-
-const toggleSelect = (name: string) => {
-  if (selectedNames.value.has(name)) {
-    selectedNames.value.delete(name)
-  } else {
-    selectedNames.value.add(name)
-  }
-  // Update selectAll state
-  selectAll.value = replies.value.length > 0 && replies.value.every(r => selectedNames.value.has(r.metadata.name))
-}
-
-const toggleSelectAll = () => {
-  if (selectAll.value) {
-    selectedNames.value.clear()
-    selectAll.value = false
-  } else {
-    selectedNames.value = new Set(replies.value.map(r => r.metadata.name))
-    selectAll.value = true
-  }
-}
-
-// Conversation dialog state
-const showDialog = ref(false)
-const conversationLoading = ref(false)
-const conversationMessages = ref<ConversationMessage[]>([])
+const toggleSelect = (name: string) => { selectedNames.value.has(name) ? selectedNames.value.delete(name) : selectedNames.value.add(name); selectAll.value = replies.value.length > 0 && replies.value.every(r => selectedNames.value.has(r.metadata.name)) }
+const toggleSelectAll = () => { if (selectAll.value) { selectedNames.value.clear(); selectAll.value = false } else { selectedNames.value = new Set(replies.value.map(r => r.metadata.name)); selectAll.value = true } }

 const fetchReplies = async () => {
-  loading.value = true
+  loading.value = true;
  try {
-    const params: Record<string, string | number> = { page: page.value, size: size.value }
-    if (filterStatus.value) params.status = filterStatus.value
-    if (filterSentiment.value) params.sentiment = filterSentiment.value
-    if (filterKeyword.value) params.keyword = filterKeyword.value
-    const { data } = await axiosInstance.get(
-      "/apis/console.api.comment-ai-autopilot.nxxy335.top/v1alpha1/replies",
-      { params },
-    )
-    replies.value = data.items || []
-    total.value = data.total || 0
-    totalPages.value = Math.ceil(total.value / size.value)
-  } catch (e) {
-    console.error("Failed to fetch replies", e)
-  } finally {
-    loading.value = false
-  }
+    const params: any = { page: page.value, size: size.value }
+    if (filterStatus.value) params.status = filterStatus.value; if (filterSentiment.value) params.sentiment = filterSentiment.value; if (filterKeyword.value) params.keyword = filterKeyword.value;
+    const { data } = await axiosInstance.get("/apis/console.api.comment-ai-autopilot.nxxy335.top/v1alpha1/replies", { params })
+    replies.value = data.items || []; total.value = data.total || 0; totalPages.value = Math.ceil(total.value / size.value)
+  } catch (e) { Toast.error("获取数据失败") } finally { loading.value = false }
 }

 const openConversation = async (reply: AiCommentReplyItem) => {
-  showDialog.value = true
-  conversationLoading.value = true
-  conversationMessages.value = []
+  showDialog.value = true; conversationLoading.value = true; conversationMessages.value = []
  try {
-    const { data } = await axiosInstance.get(
-      `/apis/console.api.comment-ai-autopilot.nxxy335.top/v1alpha1/conversation/${reply.spec.commentId}`,
-    )
+    const { data } = await axiosInstance.get(`/apis/console.api.comment-ai-autopilot.nxxy335.top/v1alpha1/conversation/${reply.spec.commentId}`)
    conversationMessages.value = data.messages || []
-  } catch (e) {
-    console.error("Failed to fetch conversation", e)
-    Toast.error("获取对话失败")
-  } finally {
-    conversationLoading.value = false
-  }
+  } catch (e) { Toast.error("获取对话失败") } finally { conversationLoading.value = false }
 }

-const handleDelete = async (name: string) => {
-  try {
-    await axiosInstance.delete(
-      `/apis/console.api.comment-ai-autopilot.nxxy335.top/v1alpha1/replies/${name}`,
-    )
-    Toast.success("删除成功")
-    fetchReplies()
-  } catch (e) {
-    console.error("Failed to delete reply", e)
-    Toast.error("删除失败")
-  }
+const handleDelete = async (name: string) => { try { await axiosInstance.delete(`/apis/console.api.comment-ai-autopilot.nxxy335.top/v1alpha1/replies/${name}`); Toast.success("删除成功"); fetchReplies() } catch (e) { Toast.error("删除失败") } }
+const handleApprove = async (name: string) => { try { await axiosInstance.post(`/apis/console.api.comment-ai-autopilot.nxxy335.top/v1alpha1/replies/${name}/approve`); Toast.success("审核通过"); fetchReplies() } catch (e) { Toast.error("审核失败") } }
+const handleReject = async (name: string) => { try { await axiosInstance.post(`/apis/console.api.comment-ai-autopilot.nxxy335.top/v1alpha1/replies/${name}/reject`); Toast.success("已拒绝"); fetchReplies() } catch (e) { Toast.error("拒绝失败") } }
+const batchApprove = async () => { if(!selectedNames.value.size||batchLoading.value) return; batchLoading.value=true; try { await axiosInstance.post("/apis/console.api.comment-ai-autopilot.nxxy335.top/v1alpha1/replies/batch-approve", { names: Array.from(selectedNames.value) }); Toast.success("成功"); selectedNames.value.clear(); selectAll.value=false; fetchReplies() } catch(e) { Toast.error("失败") } finally { batchLoading.value=false } }
+const batchReject = async () => { if(!selectedNames.value.size||batchLoading.value) return; batchLoading.value=true; try { await axiosInstance.post("/apis/console.api.comment-ai-autopilot.nxxy335.top/v1alpha1/replies/batch-reject", { names: Array.from(selectedNames.value) }); Toast.success("成功"); selectedNames.value.clear(); selectAll.value=false; fetchReplies() } catch(e) { Toast.error("失败") } finally { batchLoading.value=false } }
+const batchDelete = async () => { if(!selectedNames.value.size||batchLoading.value) return; batchLoading.value=true; try { await axiosInstance.post("/apis/console.api.comment-ai-autopilot.nxxy335.top/v1alpha1/replies/batch-delete", { names: Array.from(selectedNames.value) }); Toast.success("成功"); selectedNames.value.clear(); selectAll.value=false; fetchReplies() } catch(e) { Toast.error("失败") } finally { batchLoading.value=false } }
+
+const getStatusLabel = (s: string) => { const m:any = { PASS: '通过', FAIL: '失败', PENDING: '待审', REJECTED: '拒绝', FILTERED: '已拦截', FALSE_POSITIVE: '误报通过' }; return m[s] || s }
+const getSentimentLabel = (s: string) => { const m:any = { VERY_POSITIVE: '极好', POSITIVE: '正面', NEUTRAL: '中性', NEGATIVE: '负面', VERY_NEGATIVE: '极差' }; return m[s] || s }
+const formatDate = (ts: string) => ts ? new Date(ts).toLocaleString("zh-CN") : ""
+const getPostUrl = (slug: string) => `${window.location.origin}/archives/${slug}`
+const stripHtml = (html: string) => html ? html.replace(/<[^>]+>/g, "").replace(/\n+/g, " ").trim() : ""
+
+const truncateQuote = (content: string, length = 35) => {
+  if (!content) return ""
+  let plain = stripHtml(content)
+  plain = plain.replace(/^>\s*(?:💬\s*)?\*\*(.*?)\*\*\s*[:：]\s*/gm, '').trim()
+  return plain.length > length ? plain.substring(0, length) + "..." : plain
 }

-const handleApprove = async (name: string) => {
-  try {
-    await axiosInstance.post(
-      `/apis/console.api.comment-ai-autopilot.nxxy335.top/v1alpha1/replies/${name}/approve`,
-    )
-    Toast.success("审核通过")
-    fetchReplies()
-  } catch (e) {
-    console.error("Failed to approve reply", e)
-    Toast.error("审核操作失败")
-  }
-}
-
-const handleReject = async (name: string) => {
-  try {
-    await axiosInstance.post(
-      `/apis/console.api.comment-ai-autopilot.nxxy335.top/v1alpha1/replies/${name}/reject`,
-    )
-    Toast.success("已拒绝")
-    fetchReplies()
-  } catch (e) {
-    console.error("Failed to reject reply", e)
-    Toast.error("拒绝操作失败")
-  }
-}
-
-const batchApprove = async () => {
-  if (selectedNames.value.size === 0) return
-  try {
-    await axiosInstance.post(
-      "/apis/console.api.comment-ai-autopilot.nxxy335.top/v1alpha1/replies/batch-approve",
-      { names: Array.from(selectedNames.value) }
-    )
-    Toast.success("批量审核通过成功")
-    selectedNames.value.clear()
-    selectAll.value = false
-    fetchReplies()
-  } catch (e) {
-    Toast.error("批量审核操作失败")
-  }
-}
-
-const batchReject = async () => {
-  if (selectedNames.value.size === 0) return
-  try {
-    await axiosInstance.post(
-      "/apis/console.api.comment-ai-autopilot.nxxy335.top/v1alpha1/replies/batch-reject",
-      { names: Array.from(selectedNames.value) }
-    )
-    Toast.success("批量拒绝成功")
-    selectedNames.value.clear()
-    selectAll.value = false
-    fetchReplies()
-  } catch (e) {
-    Toast.error("批量拒绝操作失败")
-  }
-}
-
-const batchDelete = async () => {
-  if (selectedNames.value.size === 0) return
-  try {
-    await axiosInstance.post(
-      "/apis/console.api.comment-ai-autopilot.nxxy335.top/v1alpha1/replies/batch-delete",
-      { names: Array.from(selectedNames.value) }
-    )
-    Toast.success("批量删除成功")
-    selectedNames.value.clear()
-    selectAll.value = false
-    fetchReplies()
-  } catch (e) {
-    Toast.error("批量删除操作失败")
-  }
-}
-
-const getScoreClass = (score: number) => {
-  if (score >= 85) return "text-green-600"
-  if (score >= 60) return "text-yellow-600"
-  return "text-red-600"
-}
-
-const getStatusClass = (status: string) => {
-  switch (status) {
-    case "PASS":
-      return "bg-green-100 text-green-700"
-    case "FAIL":
-      return "bg-red-100 text-red-700"
-    case "PENDING":
-      return "bg-gray-100 text-gray-700"
-    case "REJECTED":
-      return "bg-red-100 text-red-700"
-    default:
-      return "bg-gray-100 text-gray-700"
-  }
-}
-
-const getStatusLabel = (status: string) => {
-  switch (status) {
-    case "PASS":
-      return "通过"
-    case "FAIL":
-      return "失败"
-    case "PENDING":
-      return "待处理"
-    case "REJECTED":
-      return "已拒绝"
-    default:
-      return status
-  }
-}
-
-const getSentimentClass = (sentiment: string) => {
-  switch (sentiment) {
-    case "POSITIVE":
-      return "bg-green-50 text-green-600"
-    case "NEGATIVE":
-      return "bg-red-50 text-red-600"
-    case "NEUTRAL":
-      return "bg-gray-100 text-gray-600"
-    default:
-      return "bg-gray-100 text-gray-600"
-  }
-}
-
-const getSentimentLabel = (sentiment: string) => {
-  switch (sentiment) {
-    case "POSITIVE":
-      return "正面"
-    case "NEGATIVE":
-      return "负面"
-    case "NEUTRAL":
-      return "中性"
-    default:
-      return sentiment
-  }
-}
-
-const formatDate = (timestamp: string) => {
-  if (!timestamp) return ""
-  return new Date(timestamp).toLocaleString("zh-CN")
-}
-
-const getPostUrl = (slug: string) => {
-  return `${window.location.origin}/archives/${slug}`
-}
-
-/**
- * Strip HTML tags for plain text display (card preview)
- */
-const stripHtml = (html: string) => {
-  if (!html) return ""
-  return html
-    .replace(/<p[^>]*>/gi, "")
-    .replace(/<\/p>/gi, "\n")
-    .replace(/<br\s*\/?>/gi, "\n")
-    .replace(/<[^>]+>/g, "")
-    .replace(/\n{3,}/g, "\n\n")
-    .trim()
-}
-
-/**
- * Sanitize and render HTML content for conversation bubbles.
- * Only allows safe inline tags, strips dangerous elements.
- */
 const renderContent = (content: string) => {
-  if (!content) return "<span class='text-gray-400'>(空)</span>"
-  return content
-    .replace(/<script[^>]*>[\s\S]*?<\/script>/gi, "")
-    .replace(/<iframe[^>]*>[\s\S]*?<\/iframe>/gi, "")
-    .replace(/<object[^>]*>[\s\S]*?<\/object>/gi, "")
-    .replace(/<embed[^>]*>/gi, "")
-    .replace(/<form[^>]*>[\s\S]*?<\/form>/gi, "")
-    .replace(/on\w+\s*=\s*["'][^"']*["']/gi, "")
-    .replace(/on\w+\s*=\s*[^\s>]*/gi, "")
-    .replace(/<p[^>]*>/gi, "<p style='margin:0 0 0.5em 0'>")
-    .replace(/<a /gi, "<a target='_blank' rel='noopener noreferrer' ")
+  if (!content) return "<span style='opacity:0.5'>(空)</span>"
+  let parsed = content.replace(/<script[^>]*>[\s\S]*?<\/script>/gi, "").replace(/<iframe[^>]*>[\s\S]*?<\/iframe>/gi, "")
+  parsed = parsed.replace(/^>\s*(?:💬\s*)?\*\*(.*?)\*\*\s*[:：]\s*/gm, "")
+  return parsed.replace(/\n/g, "<br/>")
 }

-const resetFilters = () => {
-  filterStatus.value = ""
-  filterSentiment.value = ""
-  filterKeyword.value = ""
-  page.value = 1
+const resetFilters = () => { filterStatus.value = ""; filterSentiment.value = ""; filterKeyword.value = ""; page.value = 1; fetchReplies() }
+
+const openFalsePositiveDialog = (reply: AiCommentReplyItem) => { falsePositiveTarget.value = reply; showFalsePositiveDialog.value = true }
+const handleFalsePositive = async (action: string) => {
+  if (!falsePositiveTarget.value) return
+  fpLoading.value = true
+  try {
+    await axiosInstance.post(`/apis/console.api.comment-ai-autopilot.nxxy335.top/v1alpha1/replies/${falsePositiveTarget.value.metadata.name}/false-positive`, { action })
+    Toast.success(action === "aiReply" ? "已标记为误报，AI回复正在后台生成" : "已标记为误报并通过")
+    showFalsePositiveDialog.value = false
+    falsePositiveTarget.value = null
    fetchReplies()
+  } catch (e: any) {
+    Toast.error(e?.response?.data?.message || "操作失败")
+  } finally { fpLoading.value = false }
 }
-
-watch([filterStatus, filterSentiment, filterKeyword], () => {
-  page.value = 1
+const handleTriggerAiReply = async (reply: AiCommentReplyItem) => {
+  if (triggerAiLoadingName.value) return
+  triggerAiLoadingName.value = reply.metadata.name
+  try {
+    await axiosInstance.post(`/apis/console.api.comment-ai-autopilot.nxxy335.top/v1alpha1/replies/${reply.metadata.name}/false-positive`, { action: "aiReply" })
+    Toast.success("AI回复正在后台生成")
    fetchReplies()
+  } catch (e: any) {
+    Toast.error(e?.response?.data?.message || "触发失败")
+  } finally { triggerAiLoadingName.value = null }
+}
+// 状态/情感筛选立即触发；关键词输入防抖 300ms 避免每次按键都请求
+watch([filterStatus, filterSentiment], () => { page.value = 1; fetchReplies() })
+let keywordDebounceTimer: ReturnType<typeof setTimeout> | null = null
+watch(filterKeyword, () => {
+  if (keywordDebounceTimer) clearTimeout(keywordDebounceTimer)
+  keywordDebounceTimer = setTimeout(() => { page.value = 1; fetchReplies() }, 300)
 })
-
-watch(page, () => {
-  selectedNames.value.clear()
-  selectAll.value = false
-  fetchReplies()
-})
-
+watch(page, () => { selectedNames.value.clear(); selectAll.value = false; fetchReplies() })
 onMounted(fetchReplies)
+onUnmounted(() => { if (keywordDebounceTimer) clearTimeout(keywordDebounceTimer) })
 </script>

 <style scoped>
-.line-clamp-2 {
-  display: -webkit-box;
-  -webkit-line-clamp: 2;
-  -webkit-box-orient: vertical;
-  overflow: hidden;
-}
-.line-clamp-3 {
-  display: -webkit-box;
-  -webkit-line-clamp: 3;
-  -webkit-box-orient: vertical;
-  overflow: hidden;
-}
+/* 全局基础布局 */
+.logs-container { padding-bottom: 20px; }
+.header-icon { margin-right: 8px; align-self: center; }
+
+/* 响应式工具栏 */
+.toolbar { display: flex; flex-direction: column; gap: 12px; margin: 16px; align-items: stretch; }
+@media (min-width: 768px) { .toolbar { flex-direction: row; align-items: center; } }
+.batch-toolbar { background: #eff6ff; padding: 12px 16px; border-radius: 8px; border: 1px solid #bfdbfe; }
+.batch-text { font-size: 14px; color: #1d4ed8; font-weight: bold; }
+.batch-actions { display: flex; flex-wrap: wrap; gap: 8px; width: 100%; }
+@media (min-width: 768px) { .batch-actions { width: auto; margin-left: auto; } }
+.btn-batch { padding: 6px 12px; border-radius: 6px; border: none; font-size: 12px; cursor: pointer; color: #fff; white-space: nowrap; }
+.btn-pass { background: #16a34a; } .btn-reject { background: #f97316; } .btn-delete { background: #dc2626; }
+.btn-cancel { background: transparent; color: #6b7280; border: 1px solid #d1d5db; }
+.filter-select, .filter-input { width: 100%; padding: 8px 12px; border: 1px solid #e5e7eb; border-radius: 6px; font-size: 14px; outline: none; }
+@media (min-width: 768px) { .filter-select { width: auto; min-width: 120px; } .filter-input { flex: 1; } }
+.btn-reset { padding: 8px 16px; border: 1px solid #e5e7eb; border-radius: 6px; background: #f9fafb; cursor: pointer; white-space: nowrap; width: 100%; }
+@media (min-width: 768px) { .btn-reset { width: auto; } }
+
+/* 列表区 */
+.list-area { margin: 16px; }
+.empty-state { text-align: center; padding: 60px 0; color: #9ca3af; font-size: 14px; }
+.reply-list { display: flex; flex-direction: column; gap: 16px; }
+.select-all-wrap { display: flex; align-items: center; gap: 8px; font-size: 13px; color: #6b7280; padding: 0 4px; }
+.reply-card { background: #fff; border: 1px solid #e5e7eb; border-radius: 12px; overflow: hidden; box-shadow: 0 1px 2px rgba(0,0,0,0.02); }
+.card-main { display: flex; gap: 12px; padding: 16px; align-items: flex-start; }
+.card-main input { margin-top: 4px; }
+.card-content { flex: 1; min-width: 0; }
+.card-header { display: flex; flex-direction: column; gap: 8px; margin-bottom: 8px; }
+@media (min-width: 640px) { .card-header { flex-direction: row; justify-content: space-between; align-items: center; } }
+.tags-wrap { display: flex; gap: 6px; flex-wrap: wrap; }
+.card-time { font-size: 12px; color: #9ca3af; }
+.card-text { font-size: 14px; color: #374151; line-height: 1.6; display: -webkit-box; -webkit-line-clamp: 3; -webkit-box-orient: vertical; overflow: hidden; }
+.card-filter-reason { display: flex; align-items: flex-start; gap: 6px; margin-top: 8px; padding: 6px 10px; background: #fef3c7; border: 1px solid #fde68a; border-radius: 6px; font-size: 12px; color: #92400e; }
+.filter-icon { width: 14px; height: 14px; flex-shrink: 0; margin-top: 1px; }
+.filter-category { flex-shrink: 0; padding: 1px 6px; background: #b45309; color: #fff; border-radius: 3px; font-weight: 600; font-size: 11px; line-height: 1.5; }
+.filter-detail { flex: 1; line-height: 1.5; }
+.btn-false-positive { flex-shrink: 0; margin-left: auto; padding: 2px 8px; border: 1px solid #b45309; border-radius: 4px; background: transparent; color: #b45309; font-size: 11px; cursor: pointer; white-space: nowrap; transition: all 0.15s; }
+.btn-false-positive:hover { background: #b45309; color: #fff; }
+.fp-icon-ok { color: #16a34a; }
+.btn-trigger-ai { flex-shrink: 0; margin-left: auto; padding: 2px 8px; border: 1px solid #2563eb; border-radius: 4px; background: transparent; color: #2563eb; font-size: 11px; cursor: pointer; white-space: nowrap; transition: all 0.15s; display: inline-flex; align-items: center; gap: 4px; }
+.btn-trigger-ai:hover:not(:disabled) { background: #2563eb; color: #fff; }
+.btn-trigger-ai:disabled { opacity: 0.6; cursor: not-allowed; }
+.btn-trigger-ai .fp-spinner { width: 11px; height: 11px; border-color: rgba(37,99,235,0.3); border-top-color: #2563eb; }
+.card-footer { display: flex; flex-direction: column; gap: 12px; padding: 12px 16px; background: #f9fafb; border-top: 1px solid #f3f4f6; }
+@media (min-width: 640px) { .card-footer { flex-direction: row; justify-content: space-between; align-items: center; } }
+.footer-info { font-size: 12px; color: #6b7280; display: flex; flex-wrap: wrap; gap: 12px; }
+.post-link { color: #3b82f6; text-decoration: none; } .post-link:hover { text-decoration: underline; }
+.retry-text { color: #f59e0b; }
+.footer-actions { display: flex; flex-wrap: wrap; gap: 8px; width: 100%; justify-content: flex-end; }
+@media (min-width: 640px) { .footer-actions { width: auto; } }
+.action-btn { padding: 4px 10px; border-radius: 4px; font-size: 12px; border: none; cursor: pointer; white-space: nowrap; }
+.action-btn.pass { background: #dcfce7; color: #16a34a; }
+.action-btn.reject { background: #fee2e2; color: #dc2626; }
+.action-btn.view { background: #dbeafe; color: #2563eb; }
+.action-btn.delete { background: #e5e7eb; color: #4b5563; }
+
+/* 标签体系 */
+.custom-tag { padding: 2px 6px; border-radius: 4px; font-size: 11px; font-weight: bold; }
+.tag-PASS { background: #dcfce7; color: #15803d; } .tag-FAIL { background: #fee2e2; color: #b91c1c; } .tag-PENDING { background: #fef9c3; color: #a16207; } .tag-REJECTED { background: #ffedd5; color: #c2410c; } .tag-FILTERED { background: #f1f5f9; color: #b45309; border: 1px solid #fde68a; } .tag-FALSE_POSITIVE { background: #dbeafe; color: #1d4ed8; border: 1px solid #93c5fd; }
+.tag-published { background: #dbeafe; color: #1d4ed8; } .tag-draft { background: #f3f4f6; color: #4b5563; }
+.tag-conv { background: #f3e8ff; color: #7e22ce; }
+.tag-VERY_POSITIVE { background: #dcfce7; color: #14532d; } .tag-POSITIVE { background: #ecfdf5; color: #15803d; } .tag-NEUTRAL { background: #f3f4f6; color: #4b5563; } .tag-NEGATIVE { background: #ffe4e6; color: #e11d48; } .tag-VERY_NEGATIVE { background: #fee2e2; color: #991b1b; }
+
+/* 对话弹窗与响应式气泡 */
+.dialog-overlay { position: fixed; inset: 0; background: rgba(0,0,0,0.5); display: flex; align-items: center; justify-content: center; z-index: 9999; backdrop-filter: blur(2px); padding: 16px; box-sizing: border-box; }
+.dialog-box { width: 100%; max-width: 600px; background: #fff; border-radius: 16px; display: flex; flex-direction: column; max-height: 90vh; box-shadow: 0 10px 25px rgba(0,0,0,0.15); }
+.dialog-header { display: flex; justify-content: space-between; align-items: center; padding: 16px 20px; border-bottom: 1px solid #f3f4f6; }
+.dialog-header h3 { margin: 0; font-size: 16px; font-weight: bold; }
+.close-btn { background: none; border: none; cursor: pointer; color: #9ca3af; padding: 0; display: flex; align-items: center; justify-content: center; }
+.close-btn:hover { color: #4b5563; }
+.dialog-body { padding: 20px; overflow-y: auto; background: #f8fafc; flex: 1; border-bottom-left-radius: 16px; border-bottom-right-radius: 16px; }
+
+.chat-container { display: flex; flex-direction: column; gap: 20px; }
+.chat-row { display: flex; width: 100%; }
+.row-ai { justify-content: flex-start; }
+.row-user { justify-content: flex-end; }
+.chat-message { max-width: 90%; display: flex; flex-direction: column; }
+@media (min-width: 640px) { .chat-message { max-width: 75%; } }
+.chat-owner { font-size: 12px; margin-bottom: 6px; font-weight: bold; }
+.row-ai .chat-owner { color: #2563eb; text-align: left; margin-left: 4px; }
+.row-user .chat-owner { color: #6b7280; text-align: right; margin-right: 4px; }
+.chat-time { font-size: 11px; color: #9ca3af; margin-top: 6px; }
+.row-ai .chat-time { text-align: left; margin-left: 4px; }
+.row-user .chat-time { text-align: right; margin-right: 4px; }
+
+/* 气泡样式 */
+.chat-bubble { padding: 10px 14px; border-radius: 12px; font-size: 14px; line-height: 1.6; word-wrap: break-word; box-shadow: 0 1px 2px rgba(0,0,0,0.05); }
+.bubble-ai { background: #ffffff; color: #1f2937; border: 1px solid #e2e8f0; border-top-left-radius: 2px; }
+.bubble-user { background: #2563eb; color: #ffffff; border-top-right-radius: 2px; }
+
+/* 全新精美引用框 (无左黑条，微信风格) */
+.quote-box { margin-bottom: 8px; padding: 8px 10px; border-radius: 8px; font-size: 12px; display: block; width: 100%; box-sizing: border-box; }
+.bubble-ai .quote-box { background: rgba(0,0,0,0.04); color: #6b7280; }
+.bubble-user .quote-box { background: rgba(255,255,255,0.15); color: #d1d5db; }
+.quote-header { display: flex; align-items: center; gap: 4px; margin-bottom: 2px; }
+.quote-icon { width: 12px; height: 12px; opacity: 0.7; }
+.quote-owner { font-weight: 600; font-size: 11px; }
+.bubble-ai .quote-owner { color: #374151; }
+.bubble-user .quote-owner { color: #ffffff; }
+.quote-content { display: -webkit-box; -webkit-line-clamp: 2; -webkit-box-orient: vertical; overflow: hidden; line-height: 1.4; }
+
+.pagination { display: flex; flex-direction: column; gap: 12px; align-items: center; margin-top: 20px; font-size: 14px; color: #6b7280; }
+@media (min-width: 640px) { .pagination { flex-direction: row; justify-content: space-between; } }
+.pagination-btns { display: flex; gap: 8px; }
+
+/* 误报反馈弹窗 */
+.fp-dialog { max-width: 440px; }
+.fp-dialog-body { padding: 24px; }
+.fp-desc { margin: 0 0 20px; font-size: 14px; color: #4b5563; line-height: 1.6; }
+.fp-actions { display: flex; flex-direction: column; gap: 10px; }
+.fp-btn { padding: 10px 16px; border-radius: 8px; font-size: 14px; font-weight: 500; cursor: pointer; border: none; transition: all 0.15s; display: flex; align-items: center; justify-content: center; gap: 6px; }
+.fp-btn:disabled { opacity: 0.6; cursor: not-allowed; }
+.fp-btn-primary { background: #2563eb; color: #fff; }
+.fp-btn-primary:hover:not(:disabled) { background: #1d4ed8; }
+.fp-btn-secondary { background: #f3f4f6; color: #374151; border: 1px solid #d1d5db; }
+.fp-btn-secondary:hover:not(:disabled) { background: #e5e7eb; }
+.fp-btn-ghost { background: transparent; color: #9ca3af; }
+.fp-btn-ghost:hover:not(:disabled) { color: #6b7280; background: #f9fafb; }
+.fp-spinner { width: 14px; height: 14px; border: 2px solid rgba(255,255,255,0.3); border-top-color: #fff; border-radius: 50%; animation: fp-spin 0.6s linear infinite; }
+@keyframes fp-spin { to { transform: rotate(360deg); } }
 </style>
Author	SHA1	Message	Date
sunny-335	615935d947	fix: 修复重试计数、误报反馈、响应式阻塞等问题，版本号更新为v1.2.1 修复 AiReplyOrchestrator.retryOrFail 使用 .then() 导致 retryCount 失效；修复误报反馈端点 .subscribe() 过早触发覆盖 AI 回复；修复 PersonaResolver 在响应式上下文使用 .block() 阻塞线程；修复 penalize/approveOriginalComment 缺少乐观锁重试；修复误报反馈端点无法重试 FAIL 状态；修复 tag-NEUTRAL 缺少样式、handleTriggerAiReply 缺少加载保护；修复 filterKeyword 未做防抖、performCleanup 逻辑错误；优化 extractChoice 优先匹配违规类别避免误判	2026-07-01 10:52:28 +08:00
sunny-335	9ac25c4081	chore: 版本号更新为v1.2.0，更新日志同步	2026-06-25 08:48:38 +08:00
sunny-335	2693f88982	chore: 版本号更新为v1.2.0	2026-06-25 08:33:33 +08:00
sunny-335	1cd273494d	feat: 误报反馈、上下文优先判断、FALSE_POSITIVE状态 (v1.1.3)	2026-06-25 07:56:16 +08:00
sunny-335	5337a62a13	fix: AI分类完全不可用，移除system()调用改用合并prompt方式 (v1.1.2)	2026-06-24 16:34:02 +08:00
sunny-335	122069e221	fix: ui/build.gradle cross-platform pnpm command for Linux CI	2026-06-23 19:59:25 +08:00
sunny-335	e5f973c13d	feat: 评论前置过滤（合规检测）与 AI Foundation 隔离加载 (v1.1.0)	2026-06-23 19:54:55 +08:00
bbb-lsy07andbbb-lsy07	1a2732fe19	feat: 彻底重构后台 UI 与对话上下文引用模块，全面优化移动端适配 (v1.0.4) (#6 ) * feat: 对话气泡增加引用摘要模块，解决多用户混杂交谈上下文不清晰问题 - 后端 ConversationMessage Record 新增 quoteOwner/quoteContent 字段 - 后端 getConversation 方法重写，构建 Reply 映射字典溯源引用关系 - 前端 ConversationMessage 类型定义新增 quoteOwner/quoteContent - 前端新增 truncateQuote 截断方法（复用 stripHtml，限30字符） - 前端对话气泡模板渲染灰色引用条（bg-black/5 + border-l-2） * chore: 补充 .gitignore 规则（.jar、ui/dist 等） feat: Markdown 引用注入法 - AI回复自动拼接引用块（主题无关通用方案） - 后端 CommentReplyPublisher.doPublish 重写，发布前查询被回复对象并拼接 Markdown Blockquote - 新增 buildQuoteMarkdown 辅助方法，Jsoup 清除 HTML 后截断 40 字符生成引用 - 前端 LogsView 恢复简洁气泡模板，移除 quoteOwner/quoteContent 前端引用逻辑 - renderContent 新增换行符处理，确保 Markdown 引用块正确渲染 * style: 重写 LogsView.vue - 纯 Tailwind 标签替代 Emoji，移除 300+ 行自定义 CSS - 状态/情感标签改用纯色 Tailwind 背景标签，去除所有 Emoji - 删除 300+ 行自定义 CSS，全部替换为 Tailwind 原子类 - 对话弹窗 Markdown 引用块正则提取，去除气泡和机器人 Emoji - 优化移动端响应式布局，解决排版错位问题 * style: 重写 SettingsView.vue - 纯 Tailwind 栅格布局，移除自定义 CSS - 标签导航改用 Tailwind flex + overflow-x-auto - 所有设置面板（basic/persona/model/prompt/cleanup）改用 Tailwind 原子类 - 开关改用 peer-checked 伪类实现，移除自定义 toggle CSS - 滑块刻度改用 flex justify-between 实现 - 侧边栏 lg:sticky lg:top-24，移动端自然折叠到底部 - 弹窗（评论者选择/角色编辑）改用 fixed inset-0 + backdrop-blur - 删除 600+ 行自定义 CSS * refactor: 返璞归真 - 剥离 Markdown 注入，利用 Halo 原生层级回复 - 后端 CommentReplyPublisher 删除 buildQuoteMarkdown 和 Markdown 拼接逻辑 - AI 回复直接存入纯净文本，由 Halo 原生 quoteReply 字段渲染前台层级关系 - 前端 ConversationMessage 恢复 quoteOwner/quoteContent 字段 - 前端对话弹窗添加原生 Tailwind 引用摘要框（灰色 border-l-[3px]） - 简化 renderContent，删除 Markdown 引用正则匹配 * chore: 版本号升级至 1.0.1，强制刷新 Halo 前端缓存 - plugin.yaml version: 1.0.0 -> 1.0.1 - build.gradle version: 1.0.0 -> 1.0.1 - LogsView truncateQuote/renderContent 增加历史 Markdown 引用文本清理正则 - 防止旧版测试数据 (💬 @某人:) 在界面套娃显示 * style: 彻底重写 LogsView & SettingsView - 原生 Scoped CSS 替代 Tailwind - LogsView.vue: 移除所有 Tailwind 类，改用 <style scoped> 原生 CSS - SettingsView.vue: 移除所有 Tailwind 类，改用 <style scoped> 原生 CSS - 标签配色、气泡样式、引用框全部使用纯 CSS 实现，避免 Halo 主题冲突 - 版本号升级至 1.0.2 强制刷新前端缓存 * feat: SettingsView 完整功能版 - AI角色/数据清理/导入导出 - 5个设置面板：基本设置、AI角色、模型设置、Prompt、数据清理 - AI角色：CRUD、Gravatar头像、性别/唤醒词/默认角色 - 数据清理：自动清理开关、保留天数滑块、手动清理 - 导入导出：JSON配置导入导出 - 评论者黑名单弹窗选择 - 全部使用原生 Scoped CSS * chore: 版本号升级至 1.0.3 * v1.0.4: 美化 LogsView 和 SettingsView UI，优化引用框样式与移动端适配 --------- Co-authored-by: bbb-lsy07 <bbb-lsy07@users.noreply.github.com>	2026-06-21 12:20:25 +08:00
sunny-335	633f3ff588	fix: CD pre-release-cleanup fails when no assets exist	2026-06-18 22:56:35 +08:00
sunny-335	2c88ad6fb9	feat: v1.0.0 - 唤醒词、性别配置、身份强化、移动端适配、Bug修复	2026-06-18 22:40:00 +08:00
sunny-335	a9dd1c14bc	feat: 5级情感分析、日志UI优化、页面链接支持、v1.0.0-b26cea	2026-06-18 19:58:46 +08:00
sunny-335	77e3bd36c5	refactor: 全面优化完善 - ObjectMapper注入、索引优化、代码去重、Bug修复	2026-06-18 19:42:30 +08:00
sunny-335	7934d8c947	fix: pre-cleanup release assets to avoid gh upload conflict	2026-06-18 13:52:38 +08:00
sunny-335	e6839b4aad	fix: auto-publish log showing unpublished, merge changelog, version to 1.0.0-beta.2	2026-06-17 23:13:45 +08:00
sunny-335	c3e003a572	feat: streamline dashboard, fix docs, optimize UI width - Remove AI角色 quick action, keep only 回复日志/插件设置/刷新数据 - Fix LogsView filter/batch bar width to match content area (mx-4 mt-2 -> m-4 mb-0) - Fix docs: Halo version 2.23->2.25, GitHub link nxxy335->sunny-335, Cravatar link, variable name {{conversation}}->{{conversation_history}} - Add missing settings docs (maxConversationRounds, rateLimitPerMinute, enabledPresets) - Update README: remove version display, update changelog - Update version to 1.0.0-beta.2-r8k4n2	2026-06-17 22:03:35 +08:00
sunny-335	be6a01a06e	fix: logs filter dropdown text overlapping with caret - Change select padding from px-3 to pl-3 pr-8 to reserve space for the native dropdown arrow - Update version to 1.0.0-beta.2-q9x3m7	2026-06-17 21:10:06 +08:00
sunny-335	767efdccbb	feat: v1.0.0-beta.2 - ExtensionGetter integration, UI revamp, bug fixes - Replace cross-ClassLoader reflection with ExtensionGetter.getEnabledExtension(AiModelService.class) - Declare optional pluginDependencies (ai-foundation?: "*") and recommended-apps annotation - Use OutputSpec.choice for structured classification (sentiment, review safety/quality) - Use GenerateTextRequest with maxRetries=2 for reliable chat generation - Add multi-turn conversation context (conversation_history placeholder) - Fix RateLimitService thread leak (implement DisposableBean) - Two-stage AI review: safety check + 1-5 quality score mapped to 0-100 - Remove redundant dashboard cards (sentiment distribution, 7-day trend, avg score) - Redesign settings page with tabbed navigation (basic/persona/model/prompt/cleanup) - Fix settings layout (move tab bar out of grid container) - Fix button icon+text alignment via :deep(.btn-content) inline-flex - Add review score grade labels (excellent/good/fair/poor) in logs	2026-06-17 19:40:10 +08:00
sunny-335	bf989ca3f2	docs: update README with changelog and fix license link, add CHANGELOG.md	2026-06-17 13:26:18 +08:00