diff --git a/it0_app/android/app/src/main/AndroidManifest.xml b/it0_app/android/app/src/main/AndroidManifest.xml
index b97d576..2ca542c 100644
--- a/it0_app/android/app/src/main/AndroidManifest.xml
+++ b/it0_app/android/app/src/main/AndroidManifest.xml
@@ -1,6 +1,7 @@
 <manifest xmlns:android="http://schemas.android.com/apk/res/android">
     <uses-permission android:name="android.permission.INTERNET"/>
     <uses-permission android:name="android.permission.RECORD_AUDIO"/>
+    <uses-permission android:name="android.permission.CAMERA"/>
     <uses-permission android:name="android.permission.POST_NOTIFICATIONS"/>
     <uses-permission android:name="android.permission.REQUEST_INSTALL_PACKAGES"/>
     <application
diff --git a/it0_app/android/app/src/main/java/io/flutter/plugins/GeneratedPluginRegistrant.java b/it0_app/android/app/src/main/java/io/flutter/plugins/GeneratedPluginRegistrant.java
index 57be377..7504b65 100644
--- a/it0_app/android/app/src/main/java/io/flutter/plugins/GeneratedPluginRegistrant.java
+++ b/it0_app/android/app/src/main/java/io/flutter/plugins/GeneratedPluginRegistrant.java
@@ -25,6 +25,11 @@ public final class GeneratedPluginRegistrant {
     } catch (Exception e) {
       Log.e(TAG, "Error registering plugin flutter_local_notifications, com.dexterous.flutterlocalnotifications.FlutterLocalNotificationsPlugin", e);
     }
+    try {
+      flutterEngine.getPlugins().add(new io.flutter.plugins.flutter_plugin_android_lifecycle.FlutterAndroidLifecyclePlugin());
+    } catch (Exception e) {
+      Log.e(TAG, "Error registering plugin flutter_plugin_android_lifecycle, io.flutter.plugins.flutter_plugin_android_lifecycle.FlutterAndroidLifecyclePlugin", e);
+    }
     try {
       flutterEngine.getPlugins().add(new com.it_nomads.fluttersecurestorage.FlutterSecureStoragePlugin());
     } catch (Exception e) {
@@ -40,6 +45,11 @@ public final class GeneratedPluginRegistrant {
     } catch (Exception e) {
       Log.e(TAG, "Error registering plugin flutter_tts, com.eyedeadevelopment.fluttertts.FlutterTtsPlugin", e);
     }
+    try {
+      flutterEngine.getPlugins().add(new io.flutter.plugins.imagepicker.ImagePickerPlugin());
+    } catch (Exception e) {
+      Log.e(TAG, "Error registering plugin image_picker_android, io.flutter.plugins.imagepicker.ImagePickerPlugin", e);
+    }
     try {
       flutterEngine.getPlugins().add(new dev.fluttercommunity.plus.packageinfo.PackageInfoPlugin());
     } catch (Exception e) {
diff --git a/it0_app/lib/features/chat/data/datasources/chat_remote_datasource.dart b/it0_app/lib/features/chat/data/datasources/chat_remote_datasource.dart
index 6e81963..e2ca16f 100644
--- a/it0_app/lib/features/chat/data/datasources/chat_remote_datasource.dart
+++ b/it0_app/lib/features/chat/data/datasources/chat_remote_datasource.dart
@@ -14,7 +14,7 @@ class ChatRemoteDatasource {
   Future<Map<String, dynamic>> createTask({
     required String sessionId,
     required String message,
-    List<String>? attachments,
+    List<Map<String, dynamic>>? attachments,
   }) async {
     final response = await _dio.post(
       ApiEndpoints.tasks,
diff --git a/it0_app/lib/features/chat/data/models/chat_message_model.dart b/it0_app/lib/features/chat/data/models/chat_message_model.dart
index 1db270f..644c69d 100644
--- a/it0_app/lib/features/chat/data/models/chat_message_model.dart
+++ b/it0_app/lib/features/chat/data/models/chat_message_model.dart
@@ -9,6 +9,7 @@ class ChatMessageModel {
   final Map<String, dynamic>? toolExecution;
   final Map<String, dynamic>? approvalRequest;
   final Map<String, dynamic>? metadata;
+  final List<Map<String, dynamic>>? attachments;
 
   const ChatMessageModel({
     required this.id,
@@ -19,6 +20,7 @@ class ChatMessageModel {
     this.toolExecution,
     this.approvalRequest,
     this.metadata,
+    this.attachments,
   });
 
   factory ChatMessageModel.fromJson(Map<String, dynamic> json) {
@@ -36,6 +38,9 @@ class ChatMessageModel {
       approvalRequest: json['approval_request'] as Map<String, dynamic>? ??
           json['approvalRequest'] as Map<String, dynamic>?,
       metadata: json['metadata'] as Map<String, dynamic>?,
+      attachments: (json['attachments'] as List<dynamic>?)
+          ?.map((a) => Map<String, dynamic>.from(a as Map))
+          .toList(),
     );
   }
 
@@ -49,6 +54,7 @@ class ChatMessageModel {
       if (toolExecution != null) 'toolExecution': toolExecution,
       if (approvalRequest != null) 'approvalRequest': approvalRequest,
       if (metadata != null) 'metadata': metadata,
+      if (attachments != null) 'attachments': attachments,
     };
   }
 
@@ -63,6 +69,11 @@ class ChatMessageModel {
       toolExecution: toolExecution != null ? _parseToolExecution(toolExecution!) : null,
       approvalRequest: approvalRequest != null ? _parseApprovalRequest(approvalRequest!) : null,
       metadata: metadata,
+      attachments: attachments?.map((a) => ChatAttachment(
+        base64Data: a['base64Data'] as String? ?? '',
+        mediaType: a['mediaType'] as String? ?? 'image/jpeg',
+        fileName: a['fileName'] as String?,
+      )).toList(),
     );
   }
 
@@ -94,6 +105,7 @@ class ChatMessageModel {
             }
           : null,
       metadata: entity.metadata,
+      attachments: entity.attachments?.map((a) => a.toJson()).toList(),
     );
   }
 
diff --git a/it0_app/lib/features/chat/data/repositories/chat_repository_impl.dart b/it0_app/lib/features/chat/data/repositories/chat_repository_impl.dart
index 673ad0f..3eff289 100644
--- a/it0_app/lib/features/chat/data/repositories/chat_repository_impl.dart
+++ b/it0_app/lib/features/chat/data/repositories/chat_repository_impl.dart
@@ -27,7 +27,7 @@ class ChatRepositoryImpl implements ChatRepository {
   Stream<StreamEvent> sendMessage({
     required String sessionId,
     required String message,
-    List<String>? attachments,
+    List<Map<String, dynamic>>? attachments,
   }) async* {
     // Create the task on the backend
     final response = await _remoteDatasource.createTask(
@@ -87,7 +87,7 @@ class ChatRepositoryImpl implements ChatRepository {
     final response = await _remoteDatasource.createTask(
       sessionId: sessionId,
       message: '[voice_input]',
-      attachments: [audioPath],
+      attachments: [{'filePath': audioPath, 'mediaType': 'audio/wav'}],
     );
 
     final returnedSessionId = response['sessionId'] as String? ??
diff --git a/it0_app/lib/features/chat/domain/entities/chat_message.dart b/it0_app/lib/features/chat/domain/entities/chat_message.dart
index 9dd98d3..9fbb00e 100644
--- a/it0_app/lib/features/chat/domain/entities/chat_message.dart
+++ b/it0_app/lib/features/chat/domain/entities/chat_message.dart
@@ -4,6 +4,24 @@ enum MessageType { text, toolUse, toolResult, approval, thinking, standingOrderD
 
 enum ToolStatus { executing, completed, error, blocked, awaitingApproval }
 
+class ChatAttachment {
+  final String base64Data;
+  final String mediaType;
+  final String? fileName;
+
+  const ChatAttachment({
+    required this.base64Data,
+    required this.mediaType,
+    this.fileName,
+  });
+
+  Map<String, dynamic> toJson() => {
+    'base64Data': base64Data,
+    'mediaType': mediaType,
+    if (fileName != null) 'fileName': fileName,
+  };
+}
+
 class ChatMessage {
   final String id;
   final MessageRole role;
@@ -14,6 +32,7 @@ class ChatMessage {
   final ApprovalRequest? approvalRequest;
   final bool isStreaming;
   final Map<String, dynamic>? metadata;
+  final List<ChatAttachment>? attachments;
 
   const ChatMessage({
     required this.id,
@@ -25,6 +44,7 @@ class ChatMessage {
     this.approvalRequest,
     this.isStreaming = false,
     this.metadata,
+    this.attachments,
   });
 
   ChatMessage copyWith({
@@ -37,6 +57,7 @@ class ChatMessage {
     ApprovalRequest? approvalRequest,
     bool? isStreaming,
     Map<String, dynamic>? metadata,
+    List<ChatAttachment>? attachments,
   }) {
     return ChatMessage(
       id: id ?? this.id,
@@ -48,6 +69,7 @@ class ChatMessage {
       approvalRequest: approvalRequest ?? this.approvalRequest,
       isStreaming: isStreaming ?? this.isStreaming,
       metadata: metadata ?? this.metadata,
+      attachments: attachments ?? this.attachments,
     );
   }
 }
diff --git a/it0_app/lib/features/chat/domain/repositories/chat_repository.dart b/it0_app/lib/features/chat/domain/repositories/chat_repository.dart
index e989831..eeb694c 100644
--- a/it0_app/lib/features/chat/domain/repositories/chat_repository.dart
+++ b/it0_app/lib/features/chat/domain/repositories/chat_repository.dart
@@ -6,7 +6,7 @@ abstract class ChatRepository {
   Stream<StreamEvent> sendMessage({
     required String sessionId,
     required String message,
-    List<String>? attachments,
+    List<Map<String, dynamic>>? attachments,
   });
 
   /// Sends a voice message (audio file path) and returns a stream of events.
diff --git a/it0_app/lib/features/chat/domain/usecases/send_message.dart b/it0_app/lib/features/chat/domain/usecases/send_message.dart
index cf8cdfb..1da967e 100644
--- a/it0_app/lib/features/chat/domain/usecases/send_message.dart
+++ b/it0_app/lib/features/chat/domain/usecases/send_message.dart
@@ -9,7 +9,7 @@ class SendMessage {
   Stream<StreamEvent> execute({
     required String sessionId,
     required String message,
-    List<String>? attachments,
+    List<Map<String, dynamic>>? attachments,
   }) {
     return _repository.sendMessage(
       sessionId: sessionId,
diff --git a/it0_app/lib/features/chat/presentation/pages/chat_page.dart b/it0_app/lib/features/chat/presentation/pages/chat_page.dart
index be9871b..11402e4 100644
--- a/it0_app/lib/features/chat/presentation/pages/chat_page.dart
+++ b/it0_app/lib/features/chat/presentation/pages/chat_page.dart
@@ -1,5 +1,7 @@
+import 'dart:convert';
 import 'package:flutter/material.dart';
 import 'package:flutter_riverpod/flutter_riverpod.dart';
+import 'package:image_picker/image_picker.dart';
 import '../../../../core/theme/app_colors.dart';
 import '../../domain/entities/chat_message.dart';
 import '../providers/chat_providers.dart';
@@ -23,14 +25,21 @@ class ChatPage extends ConsumerStatefulWidget {
 class _ChatPageState extends ConsumerState<ChatPage> {
   final _messageController = TextEditingController();
   final _scrollController = ScrollController();
+  final List<ChatAttachment> _pendingAttachments = [];
 
   // -- Send ------------------------------------------------------------------
 
   void _send() {
     final text = _messageController.text.trim();
-    if (text.isEmpty) return;
+    if (text.isEmpty && _pendingAttachments.isEmpty) return;
     _messageController.clear();
-    ref.read(chatProvider.notifier).sendMessage(text);
+    final attachments = _pendingAttachments.isNotEmpty
+        ? List<ChatAttachment>.from(_pendingAttachments)
+        : null;
+    if (_pendingAttachments.isNotEmpty) {
+      setState(() => _pendingAttachments.clear());
+    }
+    ref.read(chatProvider.notifier).sendMessage(text, attachments: attachments);
     _scrollToBottom();
   }
 
@@ -60,6 +69,109 @@ class _ChatPageState extends ConsumerState<ChatPage> {
     );
   }
 
+  // -- Attachments -----------------------------------------------------------
+
+  void _showAttachmentOptions() {
+    showModalBottomSheet(
+      context: context,
+      builder: (ctx) => SafeArea(
+        child: Column(
+          mainAxisSize: MainAxisSize.min,
+          children: [
+            ListTile(
+              leading: const Icon(Icons.photo_library),
+              title: const Text('从相册选择'),
+              onTap: () { Navigator.pop(ctx); _pickImage(ImageSource.gallery); },
+            ),
+            ListTile(
+              leading: const Icon(Icons.camera_alt),
+              title: const Text('拍照'),
+              onTap: () { Navigator.pop(ctx); _pickImage(ImageSource.camera); },
+            ),
+          ],
+        ),
+      ),
+    );
+  }
+
+  static const _maxAttachments = 5;
+
+  Future<void> _pickImage(ImageSource source) async {
+    if (_pendingAttachments.length >= _maxAttachments) {
+      if (mounted) {
+        ScaffoldMessenger.of(context).showSnackBar(
+          const SnackBar(content: Text('最多添加 $_maxAttachments 张图片')),
+        );
+      }
+      return;
+    }
+
+    final picker = ImagePicker();
+    final picked = await picker.pickImage(
+      source: source,
+      maxWidth: 1568,
+      maxHeight: 1568,
+      imageQuality: 85,
+    );
+    if (picked == null) return;
+
+    final bytes = await picked.readAsBytes();
+    final ext = picked.path.split('.').last.toLowerCase();
+    final mediaType = switch (ext) {
+      'png' => 'image/png',
+      'webp' => 'image/webp',
+      'gif' => 'image/gif',
+      _ => 'image/jpeg',
+    };
+
+    setState(() {
+      _pendingAttachments.add(ChatAttachment(
+        base64Data: base64Encode(bytes),
+        mediaType: mediaType,
+        fileName: picked.name,
+      ));
+    });
+  }
+
+  Widget _buildAttachmentPreview() {
+    return SizedBox(
+      height: 80,
+      child: ListView.builder(
+        scrollDirection: Axis.horizontal,
+        itemCount: _pendingAttachments.length,
+        itemBuilder: (ctx, i) {
+          final att = _pendingAttachments[i];
+          final bytes = base64Decode(att.base64Data);
+          return Stack(
+            children: [
+              Padding(
+                padding: const EdgeInsets.all(4),
+                child: ClipRRect(
+                  borderRadius: BorderRadius.circular(8),
+                  child: Image.memory(bytes, width: 72, height: 72, fit: BoxFit.cover, cacheWidth: 144, cacheHeight: 144),
+                ),
+              ),
+              Positioned(
+                top: 0,
+                right: 0,
+                child: GestureDetector(
+                  onTap: () => setState(() => _pendingAttachments.removeAt(i)),
+                  child: Container(
+                    decoration: const BoxDecoration(
+                      color: Colors.black54,
+                      shape: BoxShape.circle,
+                    ),
+                    child: const Icon(Icons.close, size: 16, color: Colors.white),
+                  ),
+                ),
+              ),
+            ],
+          );
+        },
+      ),
+    );
+  }
+
   /// Whether to show a virtual "working" node at the bottom of the timeline.
   /// True when the agent is streaming but no assistant message has appeared yet.
   bool _needsWorkingNode(ChatState chatState) {
@@ -88,6 +200,19 @@ class _ChatPageState extends ConsumerState<ChatPage> {
         isFirst: isFirst,
         isLast: isLast,
         icon: Icons.person_outline,
+        content: message.attachments != null && message.attachments!.isNotEmpty
+            ? Wrap(
+                spacing: 4,
+                runSpacing: 4,
+                children: message.attachments!.map((att) {
+                  final bytes = base64Decode(att.base64Data);
+                  return ClipRRect(
+                    borderRadius: BorderRadius.circular(8),
+                    child: Image.memory(bytes, width: 120, height: 120, fit: BoxFit.cover, cacheWidth: 240, cacheHeight: 240),
+                  );
+                }).toList(),
+              )
+            : null,
       );
     }
 
@@ -347,46 +472,57 @@ class _ChatPageState extends ConsumerState<ChatPage> {
         color: AppColors.surface,
         border: Border(top: BorderSide(color: AppColors.surfaceLight.withOpacity(0.5))),
       ),
-      child: Row(
+      child: Column(
+        mainAxisSize: MainAxisSize.min,
         children: [
-          Expanded(
-            child: TextField(
-              controller: _messageController,
-              decoration: InputDecoration(
-                hintText: isStreaming ? '追加指令...' : '输入指令...',
-                border: const OutlineInputBorder(
-                  borderRadius: BorderRadius.all(Radius.circular(24)),
+          if (_pendingAttachments.isNotEmpty) _buildAttachmentPreview(),
+          Row(
+            children: [
+              if (!isStreaming)
+                IconButton(
+                  icon: const Icon(Icons.add_circle_outline),
+                  tooltip: '添加图片',
+                  onPressed: isAwaitingApproval ? null : _showAttachmentOptions,
+                ),
+              Expanded(
+                child: TextField(
+                  controller: _messageController,
+                  decoration: InputDecoration(
+                    hintText: isStreaming ? '追加指令...' : '输入指令...',
+                    border: const OutlineInputBorder(
+                      borderRadius: BorderRadius.all(Radius.circular(24)),
+                    ),
+                    contentPadding: const EdgeInsets.symmetric(horizontal: 16, vertical: 10),
+                  ),
+                  textInputAction: TextInputAction.send,
+                  onSubmitted: (_) => isStreaming ? _inject() : _send(),
+                  enabled: !isAwaitingApproval,
                 ),
-                contentPadding: const EdgeInsets.symmetric(horizontal: 16, vertical: 10),
               ),
-              textInputAction: TextInputAction.send,
-              onSubmitted: (_) => isStreaming ? _inject() : _send(),
-              enabled: !isAwaitingApproval,
-            ),
+              const SizedBox(width: 8),
+              if (isStreaming)
+                Row(
+                  mainAxisSize: MainAxisSize.min,
+                  children: [
+                    IconButton(
+                      icon: const Icon(Icons.send, color: AppColors.info),
+                      tooltip: '追加指令',
+                      onPressed: _inject,
+                    ),
+                    IconButton(
+                      icon: const Icon(Icons.stop_circle_outlined, color: AppColors.error),
+                      tooltip: '停止',
+                      onPressed: () => ref.read(chatProvider.notifier).cancelCurrentTask(),
+                    ),
+                  ],
+                )
+              else
+                IconButton(
+                  icon: const Icon(Icons.send),
+                  onPressed: isAwaitingApproval ? null : _send,
+                ),
+            ],
           ),
-          const SizedBox(width: 8),
-          if (isStreaming)
-            // During streaming: show both inject-send and stop buttons
-            Row(
-              mainAxisSize: MainAxisSize.min,
-              children: [
-                IconButton(
-                  icon: const Icon(Icons.send, color: AppColors.info),
-                  tooltip: '追加指令',
-                  onPressed: _inject,
-                ),
-                IconButton(
-                  icon: const Icon(Icons.stop_circle_outlined, color: AppColors.error),
-                  tooltip: '停止',
-                  onPressed: () => ref.read(chatProvider.notifier).cancelCurrentTask(),
-                ),
-              ],
-            )
-          else
-            IconButton(
-              icon: const Icon(Icons.send),
-              onPressed: isAwaitingApproval ? null : _send,
-            ),
         ],
       ),
     );
diff --git a/it0_app/lib/features/chat/presentation/providers/chat_providers.dart b/it0_app/lib/features/chat/presentation/providers/chat_providers.dart
index 9b701d7..12a698b 100644
--- a/it0_app/lib/features/chat/presentation/providers/chat_providers.dart
+++ b/it0_app/lib/features/chat/presentation/providers/chat_providers.dart
@@ -125,19 +125,26 @@ class ChatNotifier extends StateNotifier<ChatState> {
   final Ref _ref;
   StreamSubscription<StreamEvent>? _eventSubscription;
 
+  // Token throttle: buffer text tokens and flush every 80ms to reduce rebuilds
+  final StringBuffer _textBuffer = StringBuffer();
+  final StringBuffer _thinkingBuffer = StringBuffer();
+  Timer? _flushTimer;
+  static const _flushInterval = Duration(milliseconds: 80);
+
   ChatNotifier(this._ref) : super(const ChatState());
 
   /// Sends a user message to the agent and processes the streamed response.
-  Future<void> sendMessage(String prompt) async {
-    if (prompt.trim().isEmpty) return;
+  Future<void> sendMessage(String prompt, {List<ChatAttachment>? attachments}) async {
+    if (prompt.trim().isEmpty && (attachments == null || attachments.isEmpty)) return;
 
     // Add the user message locally
     final userMsg = ChatMessage(
       id: DateTime.now().microsecondsSinceEpoch.toString(),
       role: MessageRole.user,
-      content: prompt,
+      content: prompt.isEmpty ? '[图片]' : prompt,
       timestamp: DateTime.now(),
       type: MessageType.text,
+      attachments: attachments,
     );
 
     state = state.copyWith(
@@ -153,7 +160,8 @@ class ChatNotifier extends StateNotifier<ChatState> {
 
       final stream = useCase.execute(
         sessionId: sessionId ?? 'new',
-        message: prompt,
+        message: prompt.isEmpty ? '[图片]' : prompt,
+        attachments: attachments?.map((a) => a.toJson()).toList(),
       );
 
       _eventSubscription?.cancel();
@@ -183,11 +191,15 @@ class ChatNotifier extends StateNotifier<ChatState> {
     switch (event) {
       case ThinkingEvent(:final content):
         _appendOrUpdateAssistantMessage(content, MessageType.thinking);
-        state = state.copyWith(agentStatus: AgentStatus.thinking);
+        if (state.agentStatus != AgentStatus.thinking) {
+          state = state.copyWith(agentStatus: AgentStatus.thinking);
+        }
 
       case TextEvent(:final content):
         _appendOrUpdateAssistantMessage(content, MessageType.text);
-        state = state.copyWith(agentStatus: AgentStatus.executing);
+        if (state.agentStatus != AgentStatus.executing) {
+          state = state.copyWith(agentStatus: AgentStatus.executing);
+        }
 
       case ToolUseEvent(:final toolName, :final input):
         final msg = ChatMessage(
@@ -259,11 +271,16 @@ class ChatNotifier extends StateNotifier<ChatState> {
         );
 
       case CompletedEvent(:final summary):
+        _flushBuffersSync();
         final hasAssistantText = state.messages.any(
           (m) => m.role == MessageRole.assistant && m.type == MessageType.text && m.content.isNotEmpty,
         );
         if (summary.isNotEmpty && !hasAssistantText) {
-          _appendOrUpdateAssistantMessage(summary, MessageType.text);
+          // Write summary directly to state (not via buffer) since we're about
+          // to set idle status — buffering would cause a brief missing-text gap.
+          state = state.copyWith(
+            messages: _applyBuffer(state.messages, summary, MessageType.text),
+          );
         }
         // Mark any remaining executing tools as completed
         final finalMessages = state.messages.map((m) {
@@ -284,6 +301,7 @@ class ChatNotifier extends StateNotifier<ChatState> {
         );
 
       case ErrorEvent(:final message):
+        _flushBuffersSync();
         state = state.copyWith(
           agentStatus: AgentStatus.error,
           error: message,
@@ -334,40 +352,83 @@ class ChatNotifier extends StateNotifier<ChatState> {
   }
 
   void _appendOrUpdateAssistantMessage(String content, MessageType type) {
-    if (state.messages.isNotEmpty) {
-      final last = state.messages.last;
+    // Buffer tokens and flush on a timer to reduce widget rebuilds
+    if (type == MessageType.thinking) {
+      _thinkingBuffer.write(content);
+    } else {
+      _textBuffer.write(content);
+    }
+    _flushTimer ??= Timer(_flushInterval, _flushBuffers);
+  }
+
+  /// Flush buffered tokens to state immediately (synchronous).
+  void _flushBuffersSync() {
+    _flushTimer?.cancel();
+    _flushTimer = null;
+    _flushBuffers();
+  }
+
+  /// Flush any buffered text/thinking tokens into state.messages.
+  void _flushBuffers() {
+    _flushTimer = null;
+    if (_textBuffer.isEmpty && _thinkingBuffer.isEmpty) return;
+
+    var messages = [...state.messages];
+    if (_textBuffer.isNotEmpty) {
+      messages = _applyBuffer(messages, _textBuffer.toString(), MessageType.text);
+      _textBuffer.clear();
+    }
+    if (_thinkingBuffer.isNotEmpty) {
+      messages = _applyBuffer(messages, _thinkingBuffer.toString(), MessageType.thinking);
+      _thinkingBuffer.clear();
+    }
+    state = state.copyWith(messages: messages);
+  }
+
+  List<ChatMessage> _applyBuffer(
+    List<ChatMessage> messages,
+    String content,
+    MessageType type,
+  ) {
+    if (messages.isNotEmpty) {
+      final last = messages.last;
       if (last.role == MessageRole.assistant && last.type == type) {
-        final updated = last.copyWith(content: last.content + content);
-        state = state.copyWith(
-          messages: [
-            ...state.messages.sublist(0, state.messages.length - 1),
-            updated,
-          ],
-        );
-        return;
+        return [
+          ...messages.sublist(0, messages.length - 1),
+          last.copyWith(content: last.content + content),
+        ];
       }
     }
-
-    final msg = ChatMessage(
-      id: DateTime.now().microsecondsSinceEpoch.toString(),
-      role: MessageRole.assistant,
-      content: content,
-      timestamp: DateTime.now(),
-      type: type,
-      isStreaming: true,
-    );
-    state = state.copyWith(messages: [...state.messages, msg]);
+    return [
+      ...messages,
+      ChatMessage(
+        id: DateTime.now().microsecondsSinceEpoch.toString(),
+        role: MessageRole.assistant,
+        content: content,
+        timestamp: DateTime.now(),
+        type: type,
+        isStreaming: true,
+      ),
+    ];
   }
 
   /// Starts a new chat — clears messages and resets sessionId.
   void startNewChat() {
     _eventSubscription?.cancel();
+    _flushTimer?.cancel();
+    _flushTimer = null;
+    _textBuffer.clear();
+    _thinkingBuffer.clear();
     state = const ChatState();
   }
 
   /// Switches to an existing session — loads its messages from the backend.
   Future<void> switchSession(String sessionId) async {
     _eventSubscription?.cancel();
+    _flushTimer?.cancel();
+    _flushTimer = null;
+    _textBuffer.clear();
+    _thinkingBuffer.clear();
     state = ChatState(sessionId: sessionId, agentStatus: AgentStatus.idle);
 
     try {
@@ -431,6 +492,9 @@ class ChatNotifier extends StateNotifier<ChatState> {
     final taskId = state.taskId;
     if (taskId == null && state.sessionId == null) return;
 
+    // Flush any buffered tokens before cancelling
+    _flushBuffersSync();
+
     // 1. IMMEDIATELY update UI — optimistic cancel
     _eventSubscription?.cancel();
     _eventSubscription = null;
@@ -483,9 +547,10 @@ class ChatNotifier extends StateNotifier<ChatState> {
       return sendMessage(message);
     }
 
-    // 1. Cancel current event subscription
+    // 1. Cancel current event subscription and flush buffered tokens
     _eventSubscription?.cancel();
     _eventSubscription = null;
+    _flushBuffersSync();
 
     // 2. Mark executing tools as completed
     final updatedMessages = state.messages.map((m) {
@@ -563,12 +628,17 @@ class ChatNotifier extends StateNotifier<ChatState> {
 
   void clearChat() {
     _eventSubscription?.cancel();
+    _flushTimer?.cancel();
+    _flushTimer = null;
+    _textBuffer.clear();
+    _thinkingBuffer.clear();
     state = const ChatState();
   }
 
   @override
   void dispose() {
     _eventSubscription?.cancel();
+    _flushTimer?.cancel();
     _ref.read(webSocketClientProvider).disconnect();
     super.dispose();
   }
diff --git a/it0_app/lib/features/chat/presentation/widgets/message_bubble.dart b/it0_app/lib/features/chat/presentation/widgets/message_bubble.dart
index fa5fb38..39d5047 100644
--- a/it0_app/lib/features/chat/presentation/widgets/message_bubble.dart
+++ b/it0_app/lib/features/chat/presentation/widgets/message_bubble.dart
@@ -1,5 +1,5 @@
 import 'package:flutter/material.dart';
-import 'package:flutter_markdown/flutter_markdown.dart';
+import 'package:gpt_markdown/gpt_markdown.dart';
 import '../../../../core/theme/app_colors.dart';
 import '../../domain/entities/chat_message.dart';
 
@@ -58,10 +58,11 @@ class MessageBubble extends StatelessWidget {
                 ),
               )
             else
-              MarkdownBody(
-                data: message.content,
-                selectable: true,
-                styleSheet: _markdownStyleSheet(context),
+              SelectionArea(
+                child: GptMarkdown(
+                  message.content,
+                  style: const TextStyle(color: AppColors.textPrimary, fontSize: 15),
+                ),
               ),
 
             // Timestamp
@@ -83,37 +84,6 @@ class MessageBubble extends StatelessWidget {
     );
   }
 
-  MarkdownStyleSheet _markdownStyleSheet(BuildContext context) {
-    return MarkdownStyleSheet(
-      p: const TextStyle(color: AppColors.textPrimary, fontSize: 15),
-      h1: const TextStyle(color: AppColors.textPrimary, fontSize: 22, fontWeight: FontWeight.bold),
-      h2: const TextStyle(color: AppColors.textPrimary, fontSize: 19, fontWeight: FontWeight.bold),
-      h3: const TextStyle(color: AppColors.textPrimary, fontSize: 17, fontWeight: FontWeight.w600),
-      strong: const TextStyle(color: AppColors.textPrimary, fontWeight: FontWeight.bold),
-      em: const TextStyle(color: AppColors.textSecondary, fontStyle: FontStyle.italic),
-      code: TextStyle(
-        color: AppColors.secondary,
-        backgroundColor: AppColors.background.withOpacity(0.5),
-        fontSize: 13,
-        fontFamily: 'monospace',
-      ),
-      codeblockDecoration: BoxDecoration(
-        color: AppColors.background.withOpacity(0.6),
-        borderRadius: BorderRadius.circular(8),
-      ),
-      codeblockPadding: const EdgeInsets.all(10),
-      blockquoteDecoration: BoxDecoration(
-        border: Border(left: BorderSide(color: AppColors.primary, width: 3)),
-      ),
-      blockquotePadding: const EdgeInsets.only(left: 12, top: 4, bottom: 4),
-      tableBorder: TableBorder.all(color: AppColors.surfaceLight, width: 0.5),
-      tableHead: const TextStyle(color: AppColors.textPrimary, fontWeight: FontWeight.bold, fontSize: 13),
-      tableBody: const TextStyle(color: AppColors.textSecondary, fontSize: 13),
-      tableCellsPadding: const EdgeInsets.symmetric(horizontal: 8, vertical: 4),
-      listBullet: const TextStyle(color: AppColors.textSecondary, fontSize: 15),
-    );
-  }
-
   String _formatTime(DateTime time) {
     final hour = time.hour.toString().padLeft(2, '0');
     final minute = time.minute.toString().padLeft(2, '0');
diff --git a/it0_app/lib/features/chat/presentation/widgets/stream_text_widget.dart b/it0_app/lib/features/chat/presentation/widgets/stream_text_widget.dart
index 24d148e..e27162c 100644
--- a/it0_app/lib/features/chat/presentation/widgets/stream_text_widget.dart
+++ b/it0_app/lib/features/chat/presentation/widgets/stream_text_widget.dart
@@ -1,11 +1,12 @@
 import 'package:flutter/material.dart';
-import 'package:flutter_markdown/flutter_markdown.dart';
+import 'package:gpt_markdown/gpt_markdown.dart';
 import '../../../../core/theme/app_colors.dart';
 
-/// Widget that renders streaming text with an animated cursor at the end,
-/// giving the appearance of real-time text generation.
-/// When streaming completes, renders Markdown.
-class StreamTextWidget extends StatefulWidget {
+/// Widget that renders streaming text as real-time Markdown.
+///
+/// During streaming: renders Markdown with a solid block cursor at the end.
+/// When streaming completes: renders Markdown with text selection enabled.
+class StreamTextWidget extends StatelessWidget {
   final String text;
   final bool isStreaming;
   final TextStyle? style;
@@ -17,89 +18,23 @@ class StreamTextWidget extends StatefulWidget {
     this.style,
   });
 
-  @override
-  State<StreamTextWidget> createState() => _StreamTextWidgetState();
-}
-
-class _StreamTextWidgetState extends State<StreamTextWidget>
-    with SingleTickerProviderStateMixin {
-  late AnimationController _cursorController;
-
-  @override
-  void initState() {
-    super.initState();
-    _cursorController = AnimationController(
-      vsync: this,
-      duration: const Duration(milliseconds: 600),
-    )..repeat(reverse: true);
-  }
-
-  @override
-  void dispose() {
-    _cursorController.dispose();
-    super.dispose();
-  }
-
   @override
   Widget build(BuildContext context) {
-    final defaultStyle = TextStyle(
-      color: AppColors.textPrimary,
-      fontSize: 15,
-    );
-    final effectiveStyle = widget.style ?? defaultStyle;
-
-    // Streaming complete — render full Markdown
-    if (!widget.isStreaming) {
-      return MarkdownBody(
-        data: widget.text,
-        selectable: true,
-        styleSheet: MarkdownStyleSheet(
-          p: effectiveStyle,
-          h1: effectiveStyle.copyWith(fontSize: 22, fontWeight: FontWeight.bold),
-          h2: effectiveStyle.copyWith(fontSize: 19, fontWeight: FontWeight.bold),
-          h3: effectiveStyle.copyWith(fontSize: 17, fontWeight: FontWeight.w600),
-          strong: effectiveStyle.copyWith(fontWeight: FontWeight.bold),
-          em: effectiveStyle.copyWith(fontStyle: FontStyle.italic, color: AppColors.textSecondary),
-          code: TextStyle(
-            color: AppColors.secondary,
-            backgroundColor: AppColors.background.withOpacity(0.5),
-            fontSize: 13,
-            fontFamily: 'monospace',
-          ),
-          codeblockDecoration: BoxDecoration(
-            color: AppColors.background.withOpacity(0.6),
-            borderRadius: BorderRadius.circular(8),
-          ),
-          codeblockPadding: const EdgeInsets.all(10),
-          tableBorder: TableBorder.all(color: AppColors.surfaceLight, width: 0.5),
-          tableHead: effectiveStyle.copyWith(fontWeight: FontWeight.bold, fontSize: 13),
-          tableBody: TextStyle(color: AppColors.textSecondary, fontSize: 13),
-          tableCellsPadding: const EdgeInsets.symmetric(horizontal: 8, vertical: 4),
-        ),
-      );
-    }
-
-    // Still streaming — show plain text with blinking cursor
-    return AnimatedBuilder(
-      animation: _cursorController,
-      builder: (context, _) {
-        return RichText(
-          text: TextSpan(
-            style: effectiveStyle,
-            children: [
-              TextSpan(text: widget.text),
-              TextSpan(
-                text: '\u2588', // Block cursor character
-                style: effectiveStyle.copyWith(
-                  color: effectiveStyle.color?.withOpacity(
-                    _cursorController.value,
-                  ),
-                ),
-              ),
-            ],
-          ),
+    final effectiveStyle = style ??
+        const TextStyle(
+          color: AppColors.textPrimary,
+          fontSize: 15,
         );
-      },
+
+    // Streaming: append solid block cursor; completed: plain markdown
+    final displayText = isStreaming ? '$text\u2588' : text;
+
+    return RepaintBoundary(
+      child: isStreaming
+          ? GptMarkdown(displayText, style: effectiveStyle)
+          : SelectionArea(
+              child: GptMarkdown(displayText, style: effectiveStyle),
+            ),
     );
   }
 }
diff --git a/it0_app/pubspec.lock b/it0_app/pubspec.lock
index f893719..48632e1 100644
--- a/it0_app/pubspec.lock
+++ b/it0_app/pubspec.lock
@@ -185,6 +185,14 @@ packages:
       url: "https://pub.dev"
     source: hosted
     version: "3.1.2"
+  cross_file:
+    dependency: transitive
+    description:
+      name: cross_file
+      sha256: "28bb3ae56f117b5aec029d702a90f57d285cd975c3c5c281eaca38dbc47c5937"
+      url: "https://pub.dev"
+    source: hosted
+    version: "0.3.5+2"
   crypto:
     dependency: "direct main"
     description:
@@ -281,6 +289,38 @@ packages:
       url: "https://pub.dev"
     source: hosted
     version: "7.0.1"
+  file_selector_linux:
+    dependency: transitive
+    description:
+      name: file_selector_linux
+      sha256: "2567f398e06ac72dcf2e98a0c95df2a9edd03c2c2e0cacd4780f20cdf56263a0"
+      url: "https://pub.dev"
+    source: hosted
+    version: "0.9.4"
+  file_selector_macos:
+    dependency: transitive
+    description:
+      name: file_selector_macos
+      sha256: "5e0bbe9c312416f1787a68259ea1505b52f258c587f12920422671807c4d618a"
+      url: "https://pub.dev"
+    source: hosted
+    version: "0.9.5"
+  file_selector_platform_interface:
+    dependency: transitive
+    description:
+      name: file_selector_platform_interface
+      sha256: "35e0bd61ebcdb91a3505813b055b09b79dfdc7d0aee9c09a7ba59ae4bb13dc85"
+      url: "https://pub.dev"
+    source: hosted
+    version: "2.7.0"
+  file_selector_windows:
+    dependency: transitive
+    description:
+      name: file_selector_windows
+      sha256: "62197474ae75893a62df75939c777763d39c2bc5f73ce5b88497208bc269abfd"
+      url: "https://pub.dev"
+    source: hosted
+    version: "0.9.3+5"
   fixnum:
     dependency: transitive
     description:
@@ -342,14 +382,22 @@ packages:
       url: "https://pub.dev"
     source: hosted
     version: "8.0.0"
-  flutter_markdown:
-    dependency: "direct main"
+  flutter_math_fork:
+    dependency: transitive
     description:
-      name: flutter_markdown
-      sha256: "08fb8315236099ff8e90cb87bb2b935e0a724a3af1623000a9cec930468e0f27"
+      name: flutter_math_fork
+      sha256: "6d5f2f1aa57ae539ffb0a04bb39d2da67af74601d685a161aff7ce5bda5fa407"
       url: "https://pub.dev"
     source: hosted
-    version: "0.7.7+1"
+    version: "0.7.4"
+  flutter_plugin_android_lifecycle:
+    dependency: transitive
+    description:
+      name: flutter_plugin_android_lifecycle
+      sha256: ee8068e0e1cd16c4a82714119918efdeed33b3ba7772c54b5d094ab53f9b7fd1
+      url: "https://pub.dev"
+    source: hosted
+    version: "2.0.33"
   flutter_riverpod:
     dependency: "direct main"
     description:
@@ -496,6 +544,14 @@ packages:
       url: "https://pub.dev"
     source: hosted
     version: "14.8.1"
+  gpt_markdown:
+    dependency: "direct main"
+    description:
+      name: gpt_markdown
+      sha256: "9b88dfaffea644070b648c204ca4a55745a49f4ad0b58ed0ab70913ad593c7a1"
+      url: "https://pub.dev"
+    source: hosted
+    version: "1.1.5"
   graphs:
     dependency: transitive
     description:
@@ -552,6 +608,70 @@ packages:
       url: "https://pub.dev"
     source: hosted
     version: "4.8.0"
+  image_picker:
+    dependency: "direct main"
+    description:
+      name: image_picker
+      sha256: "784210112be18ea55f69d7076e2c656a4e24949fa9e76429fe53af0c0f4fa320"
+      url: "https://pub.dev"
+    source: hosted
+    version: "1.2.1"
+  image_picker_android:
+    dependency: transitive
+    description:
+      name: image_picker_android
+      sha256: eda9b91b7e266d9041084a42d605a74937d996b87083395c5e47835916a86156
+      url: "https://pub.dev"
+    source: hosted
+    version: "0.8.13+14"
+  image_picker_for_web:
+    dependency: transitive
+    description:
+      name: image_picker_for_web
+      sha256: "66257a3191ab360d23a55c8241c91a6e329d31e94efa7be9cf7a212e65850214"
+      url: "https://pub.dev"
+    source: hosted
+    version: "3.1.1"
+  image_picker_ios:
+    dependency: transitive
+    description:
+      name: image_picker_ios
+      sha256: b9c4a438a9ff4f60808c9cf0039b93a42bb6c2211ef6ebb647394b2b3fa84588
+      url: "https://pub.dev"
+    source: hosted
+    version: "0.8.13+6"
+  image_picker_linux:
+    dependency: transitive
+    description:
+      name: image_picker_linux
+      sha256: "1f81c5f2046b9ab724f85523e4af65be1d47b038160a8c8deed909762c308ed4"
+      url: "https://pub.dev"
+    source: hosted
+    version: "0.2.2"
+  image_picker_macos:
+    dependency: transitive
+    description:
+      name: image_picker_macos
+      sha256: "86f0f15a309de7e1a552c12df9ce5b59fe927e71385329355aec4776c6a8ec91"
+      url: "https://pub.dev"
+    source: hosted
+    version: "0.2.2+1"
+  image_picker_platform_interface:
+    dependency: transitive
+    description:
+      name: image_picker_platform_interface
+      sha256: "567e056716333a1647c64bb6bd873cff7622233a5c3f694be28a583d4715690c"
+      url: "https://pub.dev"
+    source: hosted
+    version: "2.11.1"
+  image_picker_windows:
+    dependency: transitive
+    description:
+      name: image_picker_windows
+      sha256: d248c86554a72b5495a31c56f060cf73a41c7ff541689327b1a7dbccc33adfae
+      url: "https://pub.dev"
+    source: hosted
+    version: "0.2.2"
   intl:
     dependency: "direct main"
     description:
@@ -640,14 +760,6 @@ packages:
       url: "https://pub.dev"
     source: hosted
     version: "1.3.0"
-  markdown:
-    dependency: transitive
-    description:
-      name: markdown
-      sha256: "935e23e1ff3bc02d390bad4d4be001208ee92cc217cb5b5a6c19bc14aaa318c1"
-      url: "https://pub.dev"
-    source: hosted
-    version: "7.3.0"
   matcher:
     dependency: transitive
     description:
@@ -688,6 +800,14 @@ packages:
       url: "https://pub.dev"
     source: hosted
     version: "5.4.6"
+  nested:
+    dependency: transitive
+    description:
+      name: nested
+      sha256: "03bac4c528c64c95c722ec99280375a6f2fc708eec17c7b3f07253b626cd2a20"
+      url: "https://pub.dev"
+    source: hosted
+    version: "1.0.0"
   package_config:
     dependency: transitive
     description:
@@ -864,6 +984,14 @@ packages:
       url: "https://pub.dev"
     source: hosted
     version: "6.5.0"
+  provider:
+    dependency: transitive
+    description:
+      name: provider
+      sha256: "4e82183fa20e5ca25703ead7e05de9e4cceed1fbd1eadc1ac3cb6f565a09f272"
+      url: "https://pub.dev"
+    source: hosted
+    version: "6.1.5+1"
   pub_semver:
     dependency: transitive
     description:
@@ -1237,6 +1365,14 @@ packages:
       url: "https://pub.dev"
     source: hosted
     version: "1.0.2"
+  tuple:
+    dependency: transitive
+    description:
+      name: tuple
+      sha256: a97ce2013f240b2f3807bcbaf218765b6f301c3eff91092bcfa23a039e7dd151
+      url: "https://pub.dev"
+    source: hosted
+    version: "2.0.2"
   typed_data:
     dependency: transitive
     description:
diff --git a/it0_app/pubspec.yaml b/it0_app/pubspec.yaml
index da33668..bf0f4f2 100644
--- a/it0_app/pubspec.yaml
+++ b/it0_app/pubspec.yaml
@@ -34,8 +34,9 @@ dependencies:
 
   # UI
   fl_chart: ^0.67.0
-  flutter_markdown: ^0.7.0
+  gpt_markdown: ^1.1.5
   flutter_svg: ^2.0.10+1
+  image_picker: ^1.1.2
 
   # Voice
   record: ^6.0.0
diff --git a/packages/services/agent-service/src/application/dto/execute-task.dto.ts b/packages/services/agent-service/src/application/dto/execute-task.dto.ts
index a37697f..d507037 100644
--- a/packages/services/agent-service/src/application/dto/execute-task.dto.ts
+++ b/packages/services/agent-service/src/application/dto/execute-task.dto.ts
@@ -5,4 +5,9 @@ export class ExecuteTaskDto {
   maxTurns?: number;
   maxBudgetUsd?: number;
   skill?: { name: string; arguments: string };
+  attachments?: Array<{
+    base64Data: string;
+    mediaType: string;
+    fileName?: string;
+  }>;
 }
diff --git a/packages/services/agent-service/src/domain/entities/conversation-message.entity.ts b/packages/services/agent-service/src/domain/entities/conversation-message.entity.ts
index f319ce1..ffc3d5c 100644
--- a/packages/services/agent-service/src/domain/entities/conversation-message.entity.ts
+++ b/packages/services/agent-service/src/domain/entities/conversation-message.entity.ts
@@ -23,6 +23,13 @@ export class ConversationMessage {
   @Column({ type: 'jsonb', nullable: true })
   toolResults?: any[];
 
+  @Column({ type: 'jsonb', nullable: true })
+  attachments?: Array<{
+    mediaType: string;
+    base64Data: string;
+    fileName?: string;
+  }>;
+
   @Column({ type: 'int', nullable: true })
   tokenCount?: number;
 
diff --git a/packages/services/agent-service/src/domain/services/conversation-context.service.ts b/packages/services/agent-service/src/domain/services/conversation-context.service.ts
index b0558c5..620d998 100644
--- a/packages/services/agent-service/src/domain/services/conversation-context.service.ts
+++ b/packages/services/agent-service/src/domain/services/conversation-context.service.ts
@@ -18,7 +18,11 @@ export class ConversationContextService {
   /**
    * Save a user message to the conversation history.
    */
-  async saveUserMessage(sessionId: string, content: string): Promise<ConversationMessage> {
+  async saveUserMessage(
+    sessionId: string,
+    content: string,
+    attachments?: Array<{ base64Data: string; mediaType: string; fileName?: string }>,
+  ): Promise<ConversationMessage> {
     const tenantId = TenantContextService.getTenantId();
     const sequenceNumber = await this.messageRepository.getNextSequenceNumber(sessionId);
 
@@ -28,6 +32,7 @@ export class ConversationContextService {
     message.sessionId = sessionId;
     message.role = 'user';
     message.content = content;
+    message.attachments = attachments;
     message.tokenCount = this.estimateTokens(content);
     message.sequenceNumber = sequenceNumber;
     message.createdAt = new Date();
@@ -81,7 +86,26 @@ export class ConversationContextService {
 
     for (const msg of messages) {
       if (msg.role === 'user') {
-        history.push({ role: 'user', content: msg.content });
+        if (msg.attachments && msg.attachments.length > 0) {
+          // Build multimodal content blocks for messages with images
+          const contentBlocks: any[] = [];
+          for (const att of msg.attachments) {
+            contentBlocks.push({
+              type: 'image',
+              source: {
+                type: 'base64',
+                media_type: att.mediaType,
+                data: att.base64Data,
+              },
+            });
+          }
+          if (msg.content && msg.content !== '[图片]') {
+            contentBlocks.push({ type: 'text', text: msg.content });
+          }
+          history.push({ role: 'user', content: contentBlocks });
+        } else {
+          history.push({ role: 'user', content: msg.content });
+        }
       } else if (msg.role === 'assistant') {
         // If the assistant message has tool calls, build content blocks
         if (msg.toolCalls && msg.toolCalls.length > 0) {
diff --git a/packages/services/agent-service/src/infrastructure/engines/claude-api/claude-api-engine.ts b/packages/services/agent-service/src/infrastructure/engines/claude-api/claude-api-engine.ts
index f1a5344..ef492c8 100644
--- a/packages/services/agent-service/src/infrastructure/engines/claude-api/claude-api-engine.ts
+++ b/packages/services/agent-service/src/infrastructure/engines/claude-api/claude-api-engine.ts
@@ -59,10 +59,14 @@ export class ClaudeApiEngine implements AgentEnginePort {
     const tools = this.buildToolDefinitions(params.allowedTools);
 
     // Initialize conversation with history + user prompt
-    const messages: AnthropicMessage[] = [
-      ...(params.conversationHistory || []),
-      { role: 'user', content: params.prompt },
-    ];
+    // When history already ends with the current user message (e.g. multimodal with image blocks),
+    // don't add a duplicate plain-text user message.
+    const history = params.conversationHistory || [];
+    const lastHistoryMsg = history.length > 0 ? history[history.length - 1] : null;
+    const historyEndsWithUser = lastHistoryMsg?.role === 'user' && Array.isArray(lastHistoryMsg.content);
+    const messages: AnthropicMessage[] = historyEndsWithUser
+      ? [...history]
+      : [...history, { role: 'user', content: params.prompt }];
 
     let totalTokensUsed = 0;
     let turnCount = 0;
diff --git a/packages/services/agent-service/src/interfaces/rest/controllers/agent.controller.ts b/packages/services/agent-service/src/interfaces/rest/controllers/agent.controller.ts
index 75e7ace..80aead7 100644
--- a/packages/services/agent-service/src/interfaces/rest/controllers/agent.controller.ts
+++ b/packages/services/agent-service/src/interfaces/rest/controllers/agent.controller.ts
@@ -38,6 +38,7 @@ export class AgentController {
       allowedTools?: string[];
       engineType?: string;
       maxContextMessages?: number;
+      attachments?: Array<{ base64Data: string; mediaType: string; fileName?: string }>;
     },
   ) {
     // Allow callers to override the engine (e.g. voice uses claude_api for streaming)
@@ -85,15 +86,19 @@ export class AgentController {
     await this.taskRepository.save(task);
 
     // Save user message to conversation history
-    await this.contextService.saveUserMessage(session.id, body.prompt);
+    await this.contextService.saveUserMessage(session.id, body.prompt, body.attachments);
 
     // Load conversation history for context
     const maxCtx = body.maxContextMessages ?? 20;
     const conversationHistory = await this.contextService.loadContext(session.id, maxCtx);
     this.logger.log(`[Task ${task.id}] Loaded ${conversationHistory.length} history messages for session=${session.id}`);
 
-    // Pass conversation history (excluding the current user message, which is the last one)
-    const historyForEngine = conversationHistory.slice(0, -1);
+    // When the current message has attachments, keep it in history (it has image content blocks).
+    // Otherwise, strip it so the engine adds a plain-text user message.
+    const hasAttachments = body.attachments && body.attachments.length > 0;
+    const historyForEngine = hasAttachments
+      ? conversationHistory   // includes current user message with image blocks
+      : conversationHistory.slice(0, -1);
 
     // For SDK engine: load previous SDK session ID for native resume
     const isSdkEngine = engine.engineType === AgentEngineType.CLAUDE_AGENT_SDK;
diff --git a/packages/services/agent-service/src/main.ts b/packages/services/agent-service/src/main.ts
index 636c9a7..f7c1f32 100644
--- a/packages/services/agent-service/src/main.ts
+++ b/packages/services/agent-service/src/main.ts
@@ -2,6 +2,7 @@ import { NestFactory } from '@nestjs/core';
 import { Logger } from '@nestjs/common';
 import { ConfigService } from '@nestjs/config';
 import { WsAdapter } from '@nestjs/platform-ws';
+import * as express from 'express';
 import { AgentModule } from './agent.module';
 
 const logger = new Logger('AgentService');
@@ -16,6 +17,8 @@ process.on('uncaughtException', (error) => {
 
 async function bootstrap() {
   const app = await NestFactory.create(AgentModule);
+  // Increase body parser limit for base64 image attachments
+  app.use(express.json({ limit: '10mb' }));
   // Use raw WebSocket adapter instead of Socket.IO
   app.useWebSocketAdapter(new WsAdapter(app));
   const config = app.get(ConfigService);