Add screenshot content analysis using VLM

Features:
- ScreenshotAnalyzer class for VLM-based image analysis
- Real-time analysis during video recording
- Extract likes, comments, tags, category from screenshots
- Frontend display for category badges and tags
- Batch analysis API endpoint

Co-Authored-By: Claude <noreply@anthropic.com>

phone_agent/video_learning.py

@@ -101,6 +101,75 @@ class LearningSession:
         }
 
 
+class ScreenshotAnalyzer:
+    """分析视频截图，提取内容信息"""
+
+    ANALYSIS_PROMPT = """分析这张短视频截图，提取以下信息并以JSON格式返回：
+{
+    "description": "视频描述文案（屏幕上显示的文字，如果有的话）",
+    "likes": 点赞数（纯数字，如12000，没有则为null）,
+    "comments": 评论数（纯数字，没有则为null）,
+    "shares": 分享数（纯数字，没有则为null）,
+    "tags": ["标签1", "标签2"],
+    "category": "视频类型（美食/旅行/搞笑/知识/生活/音乐/舞蹈/其他）",
+    "elements": ["画面中的主要元素，如：人物、食物、风景等"]
+}
+注意：
+1. 只返回JSON，不要其他文字
+2. 数字不要带单位，如"1.2万"应转为12000
+3. 如果无法识别某项，设为null或空数组"""
+
+    def __init__(self, model_config: ModelConfig):
+        """初始化分析器"""
+        from phone_agent.model.client import ModelClient
+        self.model_client = ModelClient(model_config)
+
+    def analyze(self, screenshot_base64: str) -> Dict[str, Any]:
+        """分析截图并返回提取的信息"""
+        from phone_agent.model.client import MessageBuilder
+
+        # 构建消息
+        messages = [
+            MessageBuilder.create_user_message(
+                text=self.ANALYSIS_PROMPT,
+                image_base64=screenshot_base64
+            )
+        ]
+
+        try:
+            # 调用 VLM
+            response = self.model_client.request(messages)
+            result_text = response.content.strip()
+
+            # 解析 JSON
+            return self._parse_result(result_text)
+        except Exception as e:
+            print(f"[ScreenshotAnalyzer] Error: {e}")
+            return {}
+
+    def _parse_result(self, text: str) -> Dict[str, Any]:
+        """解析 VLM 返回的 JSON 结果"""
+        import re
+
+        # 尝试提取 JSON
+        json_match = re.search(r'\{[\s\S]*\}', text)
+        if not json_match:
+            return {}
+
+        try:
+            result = json.loads(json_match.group())
+            # 确保数字字段是整数
+            for field in ['likes', 'comments', 'shares']:
+                if field in result and result[field] is not None:
+                    try:
+                        result[field] = int(result[field])
+                    except (ValueError, TypeError):
+                        result[field] = None
+            return result
+        except json.JSONDecodeError:
+            return {}
+
+
 class VideoLearningAgent:
     """
     Agent for learning from short video platforms.
@@ -174,6 +243,14 @@ class VideoLearningAgent:
         # Video detection: track screenshot changes (simplified)
         self._last_screenshot_hash: Optional[str] = None
 
+        # Screenshot analyzer for content extraction
+        self._analyzer: Optional[ScreenshotAnalyzer] = None
+        try:
+            self._analyzer = ScreenshotAnalyzer(model_config)
+            print("[VideoLearning] Screenshot analyzer initialized")
+        except Exception as e:
+            print(f"[VideoLearning] Analyzer init failed: {e}")
+
     def start_session(
         self,
         device_id: str,
@@ -371,12 +448,31 @@ class VideoLearningAgent:
         return None
 
     def _record_video_from_screenshot(self, screenshot):
-        """Helper method to record video from screenshot."""
+        """Helper method to record video from screenshot with analysis."""
         import base64
         screenshot_bytes = base64.b64decode(screenshot.base64_data)
+
+        # 分析截图内容
+        analysis_result = {}
+        if self._analyzer:
+            try:
+                print(f"[VideoLearning] Analyzing screenshot...")
+                analysis_result = self._analyzer.analyze(screenshot.base64_data)
+                if analysis_result:
+                    print(f"[VideoLearning] Analysis: {analysis_result.get('category', 'N/A')}")
+            except Exception as e:
+                print(f"[VideoLearning] Analysis failed: {e}")
+
+        # 记录视频
         self.record_video(
             screenshot=screenshot_bytes,
-            description=f"Video #{self.video_counter + 1}",
+            description=analysis_result.get('description', f"Video #{self.video_counter + 1}"),
+            likes=analysis_result.get('likes'),
+            comments=analysis_result.get('comments'),
+            shares=analysis_result.get('shares'),
+            tags=analysis_result.get('tags', []),
+            category=analysis_result.get('category'),
+            elements=analysis_result.get('elements', []),
         )
 
     def _before_action(self, action: Dict[str, Any]) -> Optional[Dict[str, Any]]: