diff --git a/.claude/commands/add-trending.md b/.claude/commands/add-trending.md
index 1bbe699..203c05d 100644
--- a/.claude/commands/add-trending.md
+++ b/.claude/commands/add-trending.md
@@ -1,5 +1,5 @@
 ---
-description: 从多个数据源（GitHub Trending、Hugging Face 等）自动获取 AI 相关项目并入库
+description: Use the specialized agents workflow to automatically fetch AI projects from multiple data sources (GitHub Trending, Hugging Face, etc.) and ingest them into the database
 ---
 
 ## 用户输入
@@ -12,6 +12,11 @@ $ARGUMENTS
 
 ## 概述
 
+本命令使用 **专用 Agent 工作流** 来完成多源数据获取和入库任务。核心原则：
+- **所有数据处理逻辑由 Agent 执行**，不在主窗口执行
+- **使用 Task 工具调用各个 Agent**，让 Agent 自主完成其职责
+- **主窗口仅负责协调 Agent 调用**，不直接处理业务逻辑
+
 本命令通过任务派发器架构，从多个数据源并行爬取 AI 相关项目，经过去重、分析、质量评分后批量入库。
 
 **命令格式**: `/add-trending [source] [period] [limit]`
@@ -29,7 +34,7 @@ $ARGUMENTS
 
 ## 执行流程
 
-**重要**: 每个 Stage 都要**读取并执行对应 Agent 文件中定义的逻辑**。
+**重要**: 本命令使用专用的 Agent 架构处理工作流，由 Agent 之间相互调用完成整个流程。
 
 ### Stage 1: 初始化与任务派发
 
@@ -42,8 +47,7 @@ $ARGUMENTS
    - 初始化 `progress.json` 文件
 
 3. **执行任务派发器**:
-   - **读取** `.claude/agents/task-dispatcher.md`
-   - **按照该文件中定义的步骤**执行任务派发逻辑
+   - **使用 Task 工具调用** `task-dispatcher` agent
    - 将 source/period/limit/workspace 参数传递给任务派发器
    - 任务派发器并行调度对应的爬虫
    - 输出 `raw-projects.json`
@@ -51,18 +55,18 @@ $ARGUMENTS
 ### Stage 2: 统一去重
 
 1. **执行去重器**:
-   - **读取** `.claude/agents/deduplicator.md`
-   - **按照该文件中定义的步骤**执行去重逻辑
-   - 读取 `raw-projects.json`
+   - **使用 Task 工具调用** `deduplicator` agent
+   - 将 workspace 参数传递给去重器
+   - 去重器读取 `raw-projects.json`
    - 使用 **dbhub PostgreSQL MCP** 执行去重查询
    - 输出 `new-projects.json` 和 `task-queue.json`
 
 ### Stage 3: 项目分析
 
 1. **执行项目分析器**:
-   - **读取** `.claude/agents/project-analyzer.md`
-   - **按照该文件中定义的步骤**执行分析逻辑
-   - 处理任务队列中的项目
+   - **使用 Task 工具调用** `project-analyzer` agent
+   - 将 workspace 参数传递给分析器
+   - 分析器处理任务队列中的项目
    - 使用 **chrome-devtools-mcp** 访问项目页面
    - 生成中英双语内容
    - 计算质量评分
@@ -71,8 +75,8 @@ $ARGUMENTS
 ### Stage 4: 批量入库
 
 1. **执行入库器**:
-   - **读取** `.claude/agents/database-ingestor.md`
-   - **按照该文件中定义的步骤**执行入库逻辑
+   - **使用 Task 工具调用** `database-ingestor` agent
+   - 将 workspace 参数传递给入库器
    - 调用 `POST /api/webhook/projects`
    - 输出 `ingestion-result.json`
 
@@ -103,7 +107,8 @@ $ARGUMENTS
 
 ## 关键规则
 
-- **必须**使用 chrome-devtools-mcp 访问数据源网站
+- **必须使用 Agent 工作流**：所有数据处理由专用 Agent 完成，不在主窗口执行
+- **必须使用 Task 工具**调用 Agent：让 Agent 自主完成其职责
 - **必须**先去重再分析，避免处理已存在的项目
 - **必须**进行质量评分，仅入库 >= 40 分的项目
 - **必须**保留工作区 7 天用于调试和审计
diff --git a/.claude/settings.json b/.claude/settings.json
index ede3347..09737af 100644
--- a/.claude/settings.json
+++ b/.claude/settings.json
@@ -1,6 +1,6 @@
-// {
-//   "env": {
-//     "HTTP_PROXY": "http://proxy3.bj.petrochina:8080",
-//     "HTTPS_PROXY": "http://proxy3.bj.petrochina:8080"
-//   }
-// }
+{
+  "env": {
+    "HTTP_PROXY": "http://proxy3.bj.petrochina:8080",
+    "HTTPS_PROXY": "http://proxy3.bj.petrochina:8080"
+  }
+}
diff --git a/.gitignore b/.gitignore
index 4bb6294..91fcce6 100644
--- a/.gitignore
+++ b/.gitignore
@@ -60,6 +60,7 @@ logs/
 tmp/
 temp/
 .cache/
+.trending-workspace/
 
 # Package manager lock files (optional - uncomment if needed)
 # package-lock.json
diff --git a/CLAUDE.md b/CLAUDE.md
index 0affaf8..c741c9b 100644
--- a/CLAUDE.md
+++ b/CLAUDE.md
@@ -52,28 +52,40 @@ src/app/
 ```
 
 ### Database (Prisma + PostgreSQL)
-- **Schema**: `prisma/schema.prisma` defines models: `Project`, `Tag`, `ExternalLink`
+- **Schema**: `prisma/schema.prisma` defines models: `Project`, `Tag`, `ExternalLink`, `ProjectTag`
 - **Enums**: `ProjectStatus` (ACTIVE/ARCHIVED), `LinkType` (WEBSITE/GITHUB/HUGGINGFACE/PAPER)
 - **Client singleton**: `src/lib/prisma.ts` exports Prisma client instance
 - **Multilingual fields**: Most models have `name`/`nameEn`, `description`/`descriptionEn`, `content`/`contentEn` pairs
-- **Indexes**: `idx_project_status_createdAt`, `idx_project_slug`, `idx_tag_slug`, `idx_link_projectId`, `idx_link_type`
+- **Cascade deletions**: `ExternalLink` and `ProjectTag` use `onDelete: Cascade` - deleting a project automatically cleans up its links and tag connections
+- **Key constraints**:
+  - `Project.slug`: Unique
+  - `Tag.name`: Unique (tag names are globally unique)
+  - `Tag.slug`: Unique
+  - `ExternalLink`: `@@unique([projectId, url])` (each project can't have duplicate URLs)
+- **Indexes**: `idx_project_status_createdAt`, `idx_project_slug`, `idx_tag_slug`, `idx_link_projectId`, `idx_link_type`, `idx_link_type_url` (composite for efficient URL-based deduplication)
 
 ### Webhook Deduplication Strategy
 The webhook at `src/app/api/webhook/projects/route.ts` implements a **multi-level deduplication** strategy to prevent duplicate projects:
-1. **GitHub URL exact match** (highest priority) - via `ExternalLink` table
-2. **Website URL exact match** - via `ExternalLink` table
+1. **GitHub URL exact match** (highest priority) - via `ExternalLink` table using `idx_link_type_url` index
+2. **Website URL exact match** - via `ExternalLink` table using `idx_link_type_url` index
 3. **slug match** (fallback) - via `Project.slug` field
 
 When updating an existing project, the webhook:
 - Updates all project fields (name, description, content, status, source)
-- Replaces all tags (deletes old connections, creates new ones)
-- Replaces all links (deletes old links, creates new ones)
+- Replaces all tags (deletes old `ProjectTag` connections via `ProjectTag` table, creates new ones)
+- Replaces all links (deletes old `ExternalLink` entries, creates new ones)
+
+**Tag handling special case**: Due to `Tag.name` unique constraint, tag upsert follows:
+1. First try to find existing tag by name
+2. If not found, try upsert by slug
+3. If slug conflicts, use the existing tag with that slug
 
 ### Data Fetching (Server-Side)
 - **Location**: `src/hooks/useProjects.ts` (server functions, not React hooks)
 - **Functions**: `getProjects()`, `getProjectBySlug()`, `getAllTags()`, `getTagsWithProjectCounts()`
 - **Usage**: Directly called in Server Components and route handlers
-- **ISR**: Project detail pages use `export const revalidate = 300` (5 minutes)
+- **Query transformation**: `getProjects()` and `getProjectBySlug()` flatten the `ProjectTag` junction table structure to return tags directly
+- **ISR**: Project detail pages use `export const revalidate = 300` (5 minutes) at `src/app/[locale]/projects/[id]/page.tsx`
 
 ### Validation (Zod)
 - **Schemas**: `src/lib/validations.ts` defines all Zod schemas
@@ -101,9 +113,9 @@ When updating an existing project, the webhook:
 - **Radix UI primitives**: `@radix-ui/react-slot`, `@radix-ui/react-navigation-menu`, `@radix-ui/react-dropdown-menu`, `@radix-ui/react-separator`
 - **Lucide React icons**: Used throughout the app (package imports optimized via `experimental.optimizePackageImports`)
 - **Custom components**: `src/components/` organized by domain
-  - `layout/`: Header, Footer
+  - `layout/`: Header, Footer, AnnouncementBar
   - `locale/`: LocaleSwitcher
-  - `project/`: ProjectCard, ProjectList, ProjectDetail, ProjectSidebar, RelatedProjects, TagCloud, ExternalLinkCard
+  - `project/`: ProjectCard, ProjectList, ProjectDetail, ProjectSidebar, RelatedProjects, TagCloud, ExternalLinkCard, ShareButtons, MarkdownContent, GitHubBadges, GitHubTextStatsCard
   - `search/`: SearchBar
   - `ui/`: Base UI components (buttons, cards, etc.)
 
@@ -113,18 +125,23 @@ When updating an existing project, the webhook:
   - Image domains: localhost, *.anthropic.com
   - Lucide-react package import optimization
 - **tsconfig.json**: ES2022 target, strict mode enabled
+- **Testing**: Vitest for unit tests, Playwright for E2E tests (configured but not extensively used yet)
 
 ## MCP Servers Usage (按需使用)
 
 1. **context7**: 不确定 API 用法时查阅最新文档
 2. **chrome-devtools-mcp**: 查看页面效果、调试 UI 修复 BUG
-3. **dbhub mcp**: 查询数据库数据（PostgreSQL）
-4. **shadcn mcp**: 确认 shadcn/ui 组件用法
-5. **web-search-prime**: 默认联网搜索工具
-6. **vision-mcp-server**: 图片/视频理解
+3. **web-search-prime**: 默认联网搜索工具
+4. **vision-mcp-server**: 图片/视频理解
 
 ## Git Commits
 
 提交信息主要使用中文，使用描述性的提交格式。
 
-如果你需要查看页面 请不要直接启动项目 而是应该先检查是否已经有启动好的项目在localhost:3000运行。如果有的话请直接访问，如果没有的话再执行相关命令启动项目。
+## Development Workflow
+
+- Before viewing the page, check if there's already a project running at localhost:3000. If yes, access it directly; if no, then run `pnpm dev`
+- The dev server runs with hot-reload enabled for fast iteration
+- The middleware handles locale detection and routing automatically - no manual locale configuration needed
+- When adding new translations, update both `src/messages/zh.json` and `src/messages/en.json`
+- Database changes require running `pnpm prisma migrate dev` to update the schema