feat: update model support and multimodal

api_doc/openapi.yaml

@@ -117,7 +117,7 @@ components:
         default_model:
           type: string
           default: kimi
-          description: kimi/deepseek/qwen3-max/qwen3-vl-plus
+          description: kimi/deepseek/qwen3-vl-plus/minimax-m2.5
       additionalProperties: true
       example:
         enabled: true

api_doc/prompts_personalization.md

@@ -47,7 +47,7 @@
 | `default_run_mode` | "fast"/"thinking"/"deep"/null | null | 默认运行模式：非法值会变成 null |
 | `auto_generate_title` | bool | true | 是否自动生成对话标题 |
 | `tool_intent_enabled` | bool | true | 工具意图提示开关（属于配置结构的一部分） |
-| `default_model` | string | "kimi" | 默认模型：仅允许 `"kimi"|"deepseek"|"qwen3-max"|"qwen3-vl-plus"`，非法值回落到 `"kimi"` |
+| `default_model` | string | "kimi" | 默认模型：仅允许 `"kimi"|"deepseek"|"qwen3-vl-plus"|"minimax-m2.5"`，非法值回落到 `"kimi"` |
 
 ### 最小示例（启用 + 2 条注意事项）
 

config/model_profiles.py

@@ -11,6 +11,7 @@ CONTEXT_WINDOWS = {
     "kimi-k2.5": 256_000,
     "qwen3-max": 256_000,
     "qwen3-vl-plus": 256_000,
+    "minimax-m2.5": 204_800,
     "deepseek": 128_000,
 }
 
@@ -32,7 +33,12 @@ DEEPSEEK_THINK_MODEL = _env("MODEL_DEEPSEEK_THINK", "deepseek-reasoner")
 QWEN_BASE = _env("API_BASE_QWEN", "https://dashscope.aliyuncs.com/compatible-mode/v1")
 QWEN_KEY = _env("API_KEY_QWEN", _env("DASHSCOPE_API_KEY", ""))
 QWEN_MAX_MODEL = _env("MODEL_QWEN_MAX", "qwen3-max")
-QWEN_VL_MODEL = _env("MODEL_QWEN_VL", "qwen3-vl-plus")
+QWEN_VL_MODEL = _env("MODEL_QWEN_VL", "qwen3.5-plus")
+
+# MiniMax
+MINIMAX_BASE = _env("API_BASE_MINIMAX", "https://api.minimaxi.com/v1")
+MINIMAX_KEY = _env("API_KEY_MINIMAX", "")
+MINIMAX_MODEL = _env("MODEL_MINIMAX", "MiniMax-M2.5")
 
 
 MODEL_PROFILES = {
@@ -110,7 +116,8 @@ MODEL_PROFILES = {
         "thinking": None,  # 不支持思考
         "supports_thinking": False,
         "fast_only": True,
-        "name": "Qwen3-Max"
+        "name": "Qwen3-Max",
+        "hidden": True
     },
     "qwen3-vl-plus": {
         "context_window": CONTEXT_WINDOWS["qwen3-vl-plus"],
@@ -132,7 +139,30 @@ MODEL_PROFILES = {
         },
         "supports_thinking": True,
         "fast_only": False,
-        "name": "Qwen3-VL"
+        "name": "Qwen3.5"
+    },
+    "minimax-m2.5": {
+        "context_window": CONTEXT_WINDOWS["minimax-m2.5"],
+        "fast": {
+            "base_url": MINIMAX_BASE,
+            "api_key": MINIMAX_KEY,
+            "model_id": MINIMAX_MODEL,
+            "max_tokens": 65536,
+            "context_window": CONTEXT_WINDOWS["minimax-m2.5"],
+            "extra_params": {"reasoning_split": True}
+        },
+        "thinking": {
+            "base_url": MINIMAX_BASE,
+            "api_key": MINIMAX_KEY,
+            "model_id": MINIMAX_MODEL,
+            "max_tokens": 65536,
+            "context_window": CONTEXT_WINDOWS["minimax-m2.5"],
+            "extra_params": {"reasoning_split": True}
+        },
+        "supports_thinking": True,
+        "fast_only": False,
+        "deep_only": True,
+        "name": "MiniMax-M2.5"
     }
 }
 
@@ -158,9 +188,14 @@ MODEL_PROMPT_OVERRIDES = {
         "deep_thinking_line": "Qwen3-Max 不支持深度思考模式，将保持快速模式。"
     },
     "qwen3-vl-plus": {
-        "model_description": "你的基础模型是 Qwen3-VL-Plus，支持图文多模态理解，接口来自通义千问 DashScope。",
+        "model_description": "你的基础模型是 Qwen3.5，由通义千问提供，支持图文多模态理解。",
-        "thinking_model_line": "思考模式时，请求的模型仍为 Qwen3-VL-Plus（开启思考能力），后续请求会切回快速模型。",
+        "thinking_model_line": "思考模式时仍使用 Qwen3.5，并开启思考能力。",
-        "deep_thinking_line": "在深度思考模式中，请求的模型是 Qwen3-VL-Plus（思考版），以获得更强的分析能力。"
+        "deep_thinking_line": "深度思考模式下，所有请求都将启用思考能力，以获得更强的分析表现。"
+    },
+    "minimax-m2.5": {
+        "model_description": "你的基础模型是 MiniMax-M2.5，支持超长上下文，当前仅以深度思考模式运行。",
+        "thinking_model_line": "MiniMax-M2.5 为思考模型，快速模式不会使用。",
+        "deep_thinking_line": "深度思考模式下，所有请求持续输出思考过程并给出最终回答。"
     }
 }
 

core/main_terminal.py

@@ -1479,7 +1479,7 @@ class MainTerminal:
                 "type": "function",
                 "function": {
                     "name": "vlm_analyze",
-                    "description": "使用大参数视觉语言模型（Qwen-VL模型）理解图片：文字、物体、布局、表格等，仅支持本地路径。",
+                    "description": "使用大参数视觉语言模型（Qwen3.5）理解图片：文字、物体、布局、表格等，仅支持本地路径。",
                     "parameters": {
                         "type": "object",
                         "properties": self._inject_intent({
@@ -1872,7 +1872,7 @@ class MainTerminal:
                 }
             }
         ]
-        # 视觉模型（Qwen-VL / Kimi-k2.5）自带多模态能力，不再暴露 vlm_analyze，改为 view_image
+        # 视觉模型（Qwen3.5 / Kimi-k2.5）自带多模态能力，不再暴露 vlm_analyze，改为 view_image / view_video
         if getattr(self, "model_key", None) in {"qwen3-vl-plus", "kimi-k2.5"}:
             tools = [
                 tool for tool in tools
@@ -1895,24 +1895,23 @@ class MainTerminal:
                     }
                 }
             })
-            if getattr(self, "model_key", None) == "kimi-k2.5":
+            tools.append({
-                tools.append({
+                "type": "function",
-                    "type": "function",
+                "function": {
-                    "function": {
                     "name": "view_video",
                     "description": "将指定本地视频附加到工具结果中（tool 消息携带 video_url），便于模型查看视频内容。",
-                        "parameters": {
+                    "parameters": {
-                            "type": "object",
+                        "type": "object",
-                            "properties": self._inject_intent({
+                        "properties": self._inject_intent({
                             "path": {
                                 "type": "string",
                                 "description": "项目内的视频相对路径（不要以 /workspace 开头）；宿主机模式可用绝对路径。支持 mp4/mov/mkv/avi/webm。"
-                                }
+                            }
-                            }),
+                        }),
-                            "required": ["path"]
+                        "required": ["path"]
-                        }
                     }
-                })
+                }
+            })
         # 附加自定义工具（仅管理员可见）
         custom_tools = self._build_custom_tools()
         if custom_tools:
@@ -2045,7 +2044,11 @@ class MainTerminal:
                 if abs_path.stat().st_size > 50 * 1024 * 1024:
                     return json.dumps({"success": False, "error": "视频过大，需 <= 50MB"}, ensure_ascii=False)
                 self.pending_video_view = {"path": str(path)}
-                result = {"success": True, "message": "视频已附加到工具结果中，将随 tool 返回。", "path": path}
+                result = {
+                    "success": True,
+                    "message": "视频已附加到工具结果中，将随 tool 返回。",
+                    "path": path
+                }
             
             # 终端会话管理工具
             elif tool_name == "terminal_session":
@@ -2528,7 +2531,7 @@ class MainTerminal:
 
     def build_messages(self, context: Dict, user_input: str) -> List[Dict]:
         """构建消息列表（添加终端内容注入）"""
-        # 加载系统提示（Qwen-VL 使用专用提示）
+        # 加载系统提示（Qwen3.5 使用专用提示）
         prompt_name = "main_system_qwenvl" if getattr(self, "model_key", "kimi") in {"qwen3-vl-plus", "kimi-k2.5"} else "main_system"
         system_prompt = self.load_prompt(prompt_name)
         
@@ -2810,9 +2813,9 @@ class MainTerminal:
         normalized = mode.lower()
         if normalized not in allowed:
             raise ValueError(f"不支持的模式: {mode}")
-        # Qwen-VL 官方不支持深度思考模式
+        # 仅深度思考模型限制
-        if getattr(self, "model_key", None) == "qwen3-vl-plus" and normalized == "deep":
+        if getattr(self, "model_profile", {}).get("deep_only") and normalized != "deep":
-            raise ValueError("Qwen-VL 不支持深度思考模式")
+            raise ValueError("当前模型仅支持深度思考模式")
         # fast-only 模型限制
         if getattr(self, "model_profile", {}).get("fast_only") and normalized != "fast":
             raise ValueError("当前模型仅支持快速模式")
@@ -2840,9 +2843,9 @@ class MainTerminal:
     def set_model(self, model_key: str) -> str:
         profile = get_model_profile(model_key)
         if getattr(self.context_manager, "has_images", False) and model_key not in {"qwen3-vl-plus", "kimi-k2.5"}:
-            raise ValueError("当前对话包含图片，仅支持 Qwen-VL 或 Kimi-k2.5")
+            raise ValueError("当前对话包含图片，仅支持 Qwen3.5 或 Kimi-k2.5")
-        if getattr(self.context_manager, "has_videos", False) and model_key != "kimi-k2.5":
+        if getattr(self.context_manager, "has_videos", False) and model_key not in {"qwen3-vl-plus", "kimi-k2.5"}:
-            raise ValueError("当前对话包含视频，仅支持 Kimi-k2.5")
+            raise ValueError("当前对话包含视频，仅支持 Qwen3.5 或 Kimi-k2.5")
         self.model_key = model_key
         self.model_profile = profile
         # 将模型标识传递给底层 API 客户端，便于按模型做兼容处理
@@ -2852,9 +2855,9 @@ class MainTerminal:
         # fast-only 模型强制快速模式
         if profile.get("fast_only") and self.run_mode != "fast":
             self.set_run_mode("fast")
-        # Qwen-VL 不支持深度思考，自动回落到思考模式
+        # 仅深度思考模型强制 deep
-        if model_key == "qwen3-vl-plus" and self.run_mode == "deep":
+        if profile.get("deep_only") and self.run_mode != "deep":
-            self.set_run_mode("thinking")
+            self.set_run_mode("deep")
         # 如果模型支持思考，但当前 run_mode 为 thinking/deep，则保持；否则无需调整
         self.api_client.start_new_task(force_deep=self.deep_thinking_mode)
         return self.model_key

modules/admin_policy_manager.py

@@ -17,7 +17,7 @@ from config.paths import ADMIN_POLICY_FILE
 from modules.custom_tool_registry import CustomToolRegistry, build_default_tool_category
 
 # 可用的模型 key（与前端、model_profiles 保持一致）
-ALLOWED_MODELS = {"kimi", "deepseek", "qwen3-max", "qwen3-vl-plus"}
+ALLOWED_MODELS = {"kimi", "deepseek", "qwen3-vl-plus", "minimax-m2.5"}
 
 # UI 禁用项键名，前后端统一
 UI_BLOCK_KEYS = [

modules/ocr_client.py

@@ -13,7 +13,7 @@ from modules.file_manager import FileManager
 
 
 class OCRClient:
-    """封装 VLM（如 DeepSeek-OCR / Qwen-VL）调用逻辑。"""
+    """封装 VLM（如 DeepSeek-OCR / Qwen3.5）调用逻辑。"""
 
     def __init__(self, project_path: str, file_manager: FileManager):
         self.project_path = Path(project_path).resolve()

modules/personalization_manager.py

@@ -113,7 +113,7 @@ def sanitize_personalization_payload(
         base.update(fallback)
     data = payload or {}
     allowed_tool_categories = set(TOOL_CATEGORIES.keys())
-    allowed_models = {"kimi", "kimi-k2.5", "deepseek", "qwen3-max", "qwen3-vl-plus"}
+    allowed_models = {"kimi", "kimi-k2.5", "deepseek", "qwen3-vl-plus", "minimax-m2.5"}
     allowed_image_modes = {"original", "1080p", "720p", "540p"}
 
     def _resolve_short_field(key: str) -> str:

prompts/main_system.txt

@@ -57,7 +57,7 @@
 ### 3.3 视觉理解
 
 - **非视觉模型**：`vlm_analyze` 调用 VLM 分析图片
-- **视觉模型**（Qwen-VL / Kimi-k2.5）：`view_image` 直接查看图片，`view_video` 查看视频（Kimi-k2.5）
+- **视觉模型**（Qwen3.5 / Kimi-k2.5）：`view_image` 直接查看图片，`view_video` 查看视频
 
 ### 3.4 终端操作
 

prompts/main_system_qwenvl.txt

@@ -56,7 +56,7 @@
 
 ### 3.3 视觉理解（重点）
 
-你**自带多模态能力**，用户可以直接发送图片；如需主动查看本地图片/视频，可调用 `view_image`/`view_video` 指定路径，系统会在工具结果中附带媒体（tool 消息携带 image_url/video_url）供你查看。
+你**自带多模态能力**，用户可以直接发送图片/视频；如需主动查看本地图片/视频，可调用 `view_image`/`view_video` 指定路径，系统会在工具结果中附带媒体（tool 消息携带 image_url/video_url）供你查看。
 
 当用户提出"这是什么""识别文字/表格/票据""找瑕疵/细节""读屏/按钮含义"等图片分析任务时，优先采用下面的方法，保证细节充分、结论可验证：
 

server/_conversation_segment.py

@@ -1457,21 +1457,33 @@ async def handle_task_with_sender(terminal: WebTerminal, workspace: UserWorkspac
             if finish_reason:
                 last_finish_reason = finish_reason
             
-            # 处理思考内容
+            # 处理思考内容（兼容 reasoning_content / reasoning_details）
+            reasoning_content = ""
             if "reasoning_content" in delta:
-                reasoning_content = delta["reasoning_content"]
+                reasoning_content = delta.get("reasoning_content") or ""
-                if reasoning_content:
+            elif "reasoning_details" in delta:
-                    reasoning_chunks += 1
+                details = delta.get("reasoning_details")
-                    debug_log(f"  思考内容 #{reasoning_chunks}: {len(reasoning_content)} 字符")
+                if isinstance(details, list):
+                    parts = []
+                    for item in details:
+                        if isinstance(item, dict):
+                            text = item.get("text")
+                            if text:
+                                parts.append(text)
+                    if parts:
+                        reasoning_content = "".join(parts)
+            if reasoning_content:
+                reasoning_chunks += 1
+                debug_log(f"  思考内容 #{reasoning_chunks}: {len(reasoning_content)} 字符")
                 
-                    if not thinking_started:
+                if not thinking_started:
-                        in_thinking = True
+                    in_thinking = True
-                        thinking_started = True
+                    thinking_started = True
-                        sender('thinking_start', {})
+                    sender('thinking_start', {})
-                        await asyncio.sleep(0.05)
+                    await asyncio.sleep(0.05)
                 
-                    current_thinking += reasoning_content
+                current_thinking += reasoning_content
-                    sender('thinking_chunk', {'content': reasoning_content})
+                sender('thinking_chunk', {'content': reasoning_content})
             
             # 处理正常内容
             if "content" in delta:
@@ -2424,10 +2436,11 @@ async def handle_task_with_sender(terminal: WebTerminal, workspace: UserWorkspac
                     video_path = inj.get("path") if isinstance(inj, dict) else None
                     if video_path:
                         text_part = tool_result_content if isinstance(tool_result_content, str) else ""
+                        video_payload = [video_path]
                         tool_message_content = web_terminal.context_manager._build_content_with_images(
                             text_part,
                             [],
-                            [video_path]
+                            video_payload
                         )
                         tool_videos = [video_path]
                         if metadata_payload is None:

server/_socket_segment.py

@@ -196,7 +196,7 @@ def handle_message(data):
         emit('error', {'message': '消息不能为空'})
         return
     if images and getattr(terminal, "model_key", None) not in {"qwen3-vl-plus", "kimi-k2.5"}:
-        emit('error', {'message': '当前模型不支持图片，请切换到 Qwen-VL 或 Kimi-k2.5'})
+        emit('error', {'message': '当前模型不支持图片，请切换到 Qwen3.5 或 Kimi-k2.5'})
         return
     
     print(f"[WebSocket] 收到消息: {message}")

server/api_v1.py

@@ -677,11 +677,14 @@ def create_personalization_api():
 def list_models_api():
     items = []
     for key, profile in MODEL_PROFILES.items():
+        if profile.get("hidden"):
+            continue
         items.append({
             "model_key": key,
             "name": profile.get("name", key),
             "supports_thinking": profile.get("supports_thinking", False),
             "fast_only": profile.get("fast_only", False),
+            "deep_only": profile.get("deep_only", False),
         })
     return jsonify({"success": True, "items": items})
 

server/app_legacy.py

@@ -922,7 +922,7 @@ def get_user_resources(username: Optional[str] = None) -> Tuple[Optional[WebTerm
         terminal.admin_policy_version = policy.get("updated_at")
         # 若当前模型被禁用，则回退到第一个可用模型
         if terminal.model_key in disabled_models:
-            for candidate in ["kimi-k2.5", "kimi", "deepseek", "qwen3-vl-plus", "qwen3-max"]:
+            for candidate in ["kimi-k2.5", "kimi", "deepseek", "qwen3-vl-plus", "minimax-m2.5"]:
                 if candidate not in disabled_models:
                     try:
                         terminal.set_model(candidate)

server/chat_flow.py

@@ -1335,21 +1335,33 @@ async def handle_task_with_sender(terminal: WebTerminal, workspace: UserWorkspac
             if finish_reason:
                 last_finish_reason = finish_reason
             
-            # 处理思考内容
+            # 处理思考内容（兼容 reasoning_content / reasoning_details）
+            reasoning_content = ""
             if "reasoning_content" in delta:
-                reasoning_content = delta["reasoning_content"]
+                reasoning_content = delta.get("reasoning_content") or ""
-                if reasoning_content:
+            elif "reasoning_details" in delta:
-                    reasoning_chunks += 1
+                details = delta.get("reasoning_details")
-                    debug_log(f"  思考内容 #{reasoning_chunks}: {len(reasoning_content)} 字符")
+                if isinstance(details, list):
+                    parts = []
+                    for item in details:
+                        if isinstance(item, dict):
+                            text = item.get("text")
+                            if text:
+                                parts.append(text)
+                    if parts:
+                        reasoning_content = "".join(parts)
+            if reasoning_content:
+                reasoning_chunks += 1
+                debug_log(f"  思考内容 #{reasoning_chunks}: {len(reasoning_content)} 字符")
                 
-                    if not thinking_started:
+                if not thinking_started:
-                        in_thinking = True
+                    in_thinking = True
-                        thinking_started = True
+                    thinking_started = True
-                        sender('thinking_start', {})
+                    sender('thinking_start', {})
-                        await asyncio.sleep(0.05)
+                    await asyncio.sleep(0.05)
                 
-                    current_thinking += reasoning_content
+                current_thinking += reasoning_content
-                    sender('thinking_chunk', {'content': reasoning_content})
+                sender('thinking_chunk', {'content': reasoning_content})
             
             # 处理正常内容
             if "content" in delta:
@@ -2335,10 +2347,11 @@ async def handle_task_with_sender(terminal: WebTerminal, workspace: UserWorkspac
                     video_path = inj.get("path") if isinstance(inj, dict) else None
                     if video_path:
                         text_part = tool_result_content if isinstance(tool_result_content, str) else ""
+                        video_payload = [video_path]
                         tool_message_content = web_terminal.context_manager._build_content_with_images(
                             text_part,
                             [],
-                            [video_path]
+                            video_payload
                         )
                         tool_videos = [video_path]
                         if metadata_payload is None:

server/context.py

@@ -218,7 +218,7 @@ def get_user_resources(username: Optional[str] = None, workspace_id: Optional[st
             terminal.admin_policy_ui_blocks = policy.get("ui_blocks") or {}
             terminal.admin_policy_version = policy.get("updated_at")
             if terminal.model_key in disabled_models:
-                for candidate in ["kimi-k2.5", "kimi", "deepseek", "qwen3-vl-plus", "qwen3-max"]:
+                for candidate in ["kimi-k2.5", "kimi", "deepseek", "qwen3-vl-plus", "minimax-m2.5"]:
                     if candidate not in disabled_models:
                         try:
                             terminal.set_model(candidate)

server/socket_handlers.py

@@ -230,10 +230,10 @@ def handle_message(data):
         emit('error', {'message': '消息不能为空'})
         return
     if images and getattr(terminal, "model_key", None) not in {"qwen3-vl-plus", "kimi-k2.5"}:
-        emit('error', {'message': '当前模型不支持图片，请切换到 Qwen-VL 或 Kimi-k2.5'})
+        emit('error', {'message': '当前模型不支持图片，请切换到 Qwen3.5 或 Kimi-k2.5'})
         return
-    if videos and getattr(terminal, "model_key", None) != "kimi-k2.5":
+    if videos and getattr(terminal, "model_key", None) not in {"qwen3-vl-plus", "kimi-k2.5"}:
-        emit('error', {'message': '当前模型不支持视频，请切换到 Kimi-k2.5'})
+        emit('error', {'message': '当前模型不支持视频，请切换到 Qwen3.5 或 Kimi-k2.5'})
         return
     if images and videos:
         emit('error', {'message': '图片和视频请分开发送'})

static/src/app.ts

@@ -2612,16 +2612,16 @@ const appOptions = {
                 if (hasImages && !['qwen3-vl-plus', 'kimi-k2.5'].includes(this.currentModelKey)) {
                     this.uiPushToast({
                         title: '当前模型不支持图片',
-                        message: '请切换到 Qwen-VL 或 Kimi-k2.5 再发送图片',
+                        message: '请切换到 Qwen3.5 或 Kimi-k2.5 再发送图片',
                         type: 'error'
                     });
                     return;
                 }
 
-                if (hasVideos && this.currentModelKey !== 'kimi-k2.5') {
+                if (hasVideos && !['qwen3-vl-plus', 'kimi-k2.5'].includes(this.currentModelKey)) {
                     this.uiPushToast({
                         title: '当前模型不支持视频',
-                        message: '请切换到 Kimi-k2.5 后再发送视频',
+                        message: '请切换到 Qwen3.5 或 Kimi-k2.5 后再发送视频',
                         type: 'error'
                     });
                     return;
@@ -2876,7 +2876,7 @@ const appOptions = {
                 if (!['qwen3-vl-plus', 'kimi-k2.5'].includes(this.currentModelKey)) {
                     this.uiPushToast({
                         title: '当前模型不支持图片',
-                        message: '请选择 Qwen-VL 或 Kimi-k2.5 后再发送图片',
+                        message: '请选择 Qwen3.5 或 Kimi-k2.5 后再发送图片',
                         type: 'error'
                     });
                     return;
@@ -2891,10 +2891,10 @@ const appOptions = {
             },
 
             async openVideoPicker() {
-                if (this.currentModelKey !== 'kimi-k2.5') {
+                if (!['qwen3-vl-plus', 'kimi-k2.5'].includes(this.currentModelKey)) {
                     this.uiPushToast({
                         title: '当前模型不支持视频',
-                        message: '请切换到 Kimi-k2.5 后再发送视频',
+                        message: '请切换到 Qwen3.5 或 Kimi-k2.5 后再发送视频',
                         type: 'error'
                     });
                     return;
@@ -3163,7 +3163,7 @@ const appOptions = {
                 if (this.conversationHasImages && !['qwen3-vl-plus', 'kimi-k2.5'].includes(key)) {
                     this.uiPushToast({
                         title: '切换失败',
-                        message: '当前对话包含图片，仅支持 Qwen-VL 或 Kimi-k2.5',
+                        message: '当前对话包含图片，仅支持 Qwen3.5 或 Kimi-k2.5',
                         type: 'error'
                     });
                     return;
@@ -3187,16 +3187,11 @@ const appOptions = {
                         this.thinkingMode = data.thinking_mode ?? (data.run_mode !== 'fast');
                     } else {
                         // 前端兼容策略：根据模型特性自动调整运行模式
-                        if (key === 'qwen3-vl-plus') {
+                        const currentModel = modelStore.currentModel;
-                            // Qwen-VL 不支持深度思考，若当前为 deep 则回落到思考模式
+                        if (currentModel?.deepOnly) {
-                            if (this.runMode === 'deep') {
+                            this.runMode = 'deep';
-                                this.runMode = 'thinking';
+                            this.thinkingMode = true;
-                                this.thinkingMode = true;
+                        } else if (currentModel?.fastOnly) {
-                            } else {
-                                this.thinkingMode = this.runMode !== 'fast';
-                            }
-                        } else if (key === 'qwen3-max') {
-                            // Qwen-Max 仅快速模式
                             this.runMode = 'fast';
                             this.thinkingMode = false;
                         } else {
@@ -3244,12 +3239,12 @@ const appOptions = {
                 }
                 const modelStore = useModelStore();
                 const fastOnly = modelStore.currentModel?.fastOnly;
-                const currentModelKey = modelStore.currentModel?.key;
+                const deepOnly = modelStore.currentModel?.deepOnly;
                 if (fastOnly && mode !== 'fast') {
                     if (!options.suppressToast) {
                         this.uiPushToast({
                             title: '模式不可用',
-                            message: 'Qwen-Max只支持快速模式',
+                            message: '当前模型仅支持快速模式',
                             type: 'warning'
                         });
                     }
@@ -3257,12 +3252,11 @@ const appOptions = {
                     this.inputCloseMenus();
                     return;
                 }
-                // Qwen-VL 不支持深度思考模式
+                if (deepOnly && mode !== 'deep') {
-                if (currentModelKey === 'qwen3-vl-plus' && mode === 'deep') {
                     if (!options.suppressToast) {
                         this.uiPushToast({
                             title: '模式不可用',
-                            message: 'Qwen-VL 不支持深度思考模式，请使用快速或思考模式',
+                            message: '当前模型仅支持深度思考模式',
                             type: 'warning'
                         });
                     }

static/src/components/input/QuickMenu.vue

@@ -27,7 +27,7 @@
         发送图片
       </button>
       <button
-        v-if="currentModelKey === 'kimi-k2.5'"
+        v-if="['qwen3-vl-plus', 'kimi-k2.5'].includes(currentModelKey)"
         type="button"
         class="menu-entry"
         @click.stop="$emit('pick-video')"

static/src/components/personalization/PersonalizationDrawer.vue

@@ -252,7 +252,7 @@
                           </button>
                         </div>
                         <p class="behavior-hint">
-                          Qwen-Max 仅支持快速模式；Qwen-VL 不支持深度思考模式，选择时会给出提示。
+                          MiniMax-M2.5 仅支持深度思考模式，选择时会给出提示。
                         </p>
                       </div>
                       <div class="behavior-field">
@@ -720,8 +720,8 @@ const modelOptions = [
   { id: 'deepseek', label: 'DeepSeek', desc: '通用 + 思考强化', value: 'deepseek' },
   { id: 'kimi-k2.5', label: 'Kimi-k2.5', desc: '新版 Kimi，思考开关 + 图文多模态', value: 'kimi-k2.5', badge: '图文' },
   { id: 'kimi', label: 'Kimi-k2', desc: '旧版 Kimi-k2，兼顾通用对话', value: 'kimi' },
-  { id: 'qwen3-max', label: 'Qwen-Max', desc: '仅快速模式，不支持思考', value: 'qwen3-max', badge: '仅快速' },
+  { id: 'qwen3-vl-plus', label: 'Qwen3.5', desc: '图文多模态 + 深度思考', value: 'qwen3-vl-plus', badge: '图文' },
-  { id: 'qwen3-vl-plus', label: 'Qwen-VL', desc: '图文多模态，思考/快速均可', value: 'qwen3-vl-plus', badge: '图文' }
+  { id: 'minimax-m2.5', label: 'MiniMax-M2.5', desc: '仅深度思考，超长上下文', value: 'minimax-m2.5', badge: '深度思考' }
 ] as const;
 
 const filteredModelOptions = computed(() =>
@@ -815,11 +815,8 @@ const setDefaultModel = (value: string) => {
 
 const checkModeModelConflict = (mode: RunModeValue, model: string | null): boolean => {
   const warnings: string[] = [];
-  if (model === 'qwen3-max' && mode && mode !== 'fast') {
+  if (model === 'minimax-m2.5' && mode && mode !== 'deep') {
-    warnings.push('Qwen-Max 仅支持快速模式，已保持原设置。');
+    warnings.push('MiniMax-M2.5 仅支持深度思考模式，已保持原设置。');
-  }
-  if (model === 'qwen3-vl-plus' && mode === 'deep') {
-    warnings.push('Qwen-VL 不支持深度思考模式，已保持原设置。');
   }
   if (warnings.length) {
     uiStore.pushToast({

static/src/stores/model.ts

@@ -1,6 +1,6 @@
 import { defineStore } from 'pinia';
 
-export type ModelKey = 'kimi-k2.5' | 'kimi' | 'deepseek' | 'qwen3-max' | 'qwen3-vl-plus';
+export type ModelKey = 'kimi-k2.5' | 'kimi' | 'deepseek' | 'qwen3-vl-plus' | 'minimax-m2.5';
 
 export interface ModelOption {
   key: ModelKey;
@@ -8,6 +8,7 @@ export interface ModelOption {
   description: string;
   fastOnly: boolean;
   supportsThinking: boolean;
+  deepOnly?: boolean;
 }
 
 interface ModelState {
@@ -40,19 +41,20 @@ export const useModelStore = defineStore('model', {
         fastOnly: false,
         supportsThinking: true
       },
-      {
-        key: 'qwen3-max',
-        label: 'Qwen-Max',
-        description: '仅支持快速模式',
-        fastOnly: true,
-        supportsThinking: false
-      },
       {
         key: 'qwen3-vl-plus',
-        label: 'Qwen-VL',
+        label: 'Qwen3.5',
-        description: '支持图片输入',
+        description: '图文视频多模态 + 深度思考',
         fastOnly: false,
         supportsThinking: true
+      },
+      {
+        key: 'minimax-m2.5',
+        label: 'MiniMax-M2.5',
+        description: '仅深度思考，超长上下文',
+        fastOnly: false,
+        supportsThinking: true,
+        deepOnly: true
       }
     ]
   }),

static/src/stores/personalization.ts

@@ -397,7 +397,7 @@ export const usePersonalizationStore = defineStore('personalization', {
       this.clearFeedback();
     },
     setDefaultModel(model: string | null) {
-      const allowed = ['deepseek', 'kimi-k2.5', 'kimi', 'qwen3-max', 'qwen3-vl-plus'];
+      const allowed = ['deepseek', 'kimi-k2.5', 'kimi', 'qwen3-vl-plus', 'minimax-m2.5'];
       const target = typeof model === 'string' && allowed.includes(model) ? model : null;
       this.form = {
         ...this.form,

sub_agent/utils/api_client.py

@@ -4,7 +4,7 @@
 import httpx
 import json
 import asyncio
-from typing import List, Dict, Optional, AsyncGenerator
+from typing import List, Dict, Optional, AsyncGenerator, Any
 try:
     from config import API_BASE_URL, API_KEY, MODEL_ID, OUTPUT_FORMATS, DEFAULT_RESPONSE_MAX_TOKENS
 except ImportError:
@@ -98,6 +98,54 @@ class DeepSeekClient:
         
         return json.dumps(data, ensure_ascii=False)
 
+    def _extract_reasoning_delta(self, delta: Dict[str, Any]) -> str:
+        """统一提取思考内容，兼容 reasoning_content / reasoning_details。"""
+        if not isinstance(delta, dict):
+            return ""
+        if "reasoning_content" in delta:
+            return delta.get("reasoning_content") or ""
+        details = delta.get("reasoning_details")
+        if isinstance(details, list):
+            parts: List[str] = []
+            for item in details:
+                if isinstance(item, dict):
+                    text = item.get("text")
+                    if text:
+                        parts.append(text)
+            if parts:
+                return "".join(parts)
+        return ""
+
+    def _merge_system_messages(self, messages: List[Dict]) -> List[Dict]:
+        """
+        将多个 system 消息合并为一个。
+        """
+        if not messages:
+            return messages
+        merged_contents: List[str] = []
+        new_messages: List[Dict] = []
+        first_system_index: Optional[int] = None
+        for msg in messages:
+            if msg.get("role") == "system":
+                if first_system_index is None:
+                    first_system_index = len(new_messages)
+                content = msg.get("content", "")
+                if isinstance(content, str):
+                    merged_contents.append(content)
+                else:
+                    merged_contents.append(json.dumps(content, ensure_ascii=False))
+            else:
+                new_messages.append(msg)
+        if not merged_contents:
+            return messages
+        merged = {
+            "role": "system",
+            "content": "\n\n".join(c for c in merged_contents if c)
+        }
+        insert_at = first_system_index if first_system_index is not None else 0
+        new_messages.insert(insert_at, merged)
+        return new_messages
+    
     def start_new_task(self):
         """开始新任务（重置任务级别的状态）"""
         self.current_task_first_call = True
@@ -215,9 +263,11 @@ class DeepSeekClient:
         except (TypeError, ValueError):
             max_tokens = 4096
         
+        final_messages = self._merge_system_messages(messages)
+
         payload = {
             "model": self.model_id,
-            "messages": messages,
+            "messages": final_messages,
             "stream": stream,
             "thinking": {"type": "enabled" if current_thinking_mode else "disabled"},
             "max_tokens": max_tokens
@@ -332,9 +382,9 @@ class DeepSeekClient:
                 delta = chunk["choices"][0].get("delta", {})
                 
                 # 处理思考内容（只在思考模式开启时）
-                if "reasoning_content" in delta and should_show_thinking:
+                if should_show_thinking:
-                    reasoning_content = delta["reasoning_content"]
+                    reasoning_content = self._extract_reasoning_delta(delta)
-                    if reasoning_content:  # 只处理非空内容
+                    if reasoning_content:
                         if not in_thinking:
                             self._print("💭 [正在思考]\n", end="", flush=True)
                             in_thinking = True
@@ -568,9 +618,9 @@ class DeepSeekClient:
                 delta = chunk["choices"][0].get("delta", {})
                 
                 # 处理思考内容
-                if "reasoning_content" in delta and should_show_thinking:
+                if should_show_thinking:
-                    reasoning_content = delta["reasoning_content"]
+                    reasoning_content = self._extract_reasoning_delta(delta)
-                    if reasoning_content:  # 只处理非空内容
+                    if reasoning_content:
                         if not in_thinking:
                             self._print("💭 [正在思考]\n", end="", flush=True)
                             in_thinking = True

sub_agent/web_server.py

@@ -3354,22 +3354,34 @@ async def handle_task_with_sender(terminal: WebTerminal, message, sender, client
             if finish_reason:
                 last_finish_reason = finish_reason
             
-            # 处理思考内容
+            # 处理思考内容（兼容 reasoning_content / reasoning_details）
+            reasoning_content = ""
             if "reasoning_content" in delta:
-                reasoning_content = delta["reasoning_content"]
+                reasoning_content = delta.get("reasoning_content") or ""
-                if reasoning_content:
+            elif "reasoning_details" in delta:
-                    reasoning_chunks += 1
+                details = delta.get("reasoning_details")
-                    debug_log(f"  思考内容 #{reasoning_chunks}: {len(reasoning_content)} 字符")
+                if isinstance(details, list):
+                    parts = []
+                    for item in details:
+                        if isinstance(item, dict):
+                            text = item.get("text")
+                            if text:
+                                parts.append(text)
+                    if parts:
+                        reasoning_content = "".join(parts)
+            if reasoning_content:
+                reasoning_chunks += 1
+                debug_log(f"  思考内容 #{reasoning_chunks}: {len(reasoning_content)} 字符")
                 
-                    if should_show_thinking:
+                if should_show_thinking:
-                        if not thinking_started:
+                    if not thinking_started:
-                            in_thinking = True
+                        in_thinking = True
-                            thinking_started = True
+                        thinking_started = True
-                            sender('thinking_start', {})
+                        sender('thinking_start', {})
-                            await asyncio.sleep(0.05)
+                        await asyncio.sleep(0.05)
                     
-                        current_thinking += reasoning_content
+                    current_thinking += reasoning_content
-                        sender('thinking_chunk', {'content': reasoning_content})
+                    sender('thinking_chunk', {'content': reasoning_content})
             
             # 处理正常内容
             if "content" in delta:

utils/api_client.py

@@ -128,11 +128,12 @@ class DeepSeekClient:
                 match_texts.append("未找到匹配内容。")
             return "\n".join([header] + match_texts)
 
-    def _build_content_with_images(self, text: str, images: List[str], videos: Optional[List[str]] = None) -> Any:
+    def _build_content_with_images(self, text: str, images: List[str], videos: Optional[List[Any]] = None) -> Any:
         """将文本与图片/视频路径拼成多模态 content（用于 tool 消息）。"""
         videos = videos or []
         if not images and not videos:
             return text
+        qwen_video_fps = 2
         parts: List[Dict[str, Any]] = []
         if text:
             parts.append({"type": "text", "text": text})
@@ -150,8 +151,14 @@ class DeepSeekClient:
                 parts.append({"type": "image_url", "image_url": {"url": f"data:{mime};base64,{b64}"}})
             except Exception:
                 continue
-        for path in videos:
+        for item in videos:
             try:
+                if isinstance(item, dict):
+                    path = item.get("path") or ""
+                else:
+                    path = item
+                if not path:
+                    continue
                 abs_path = (base_path / path).resolve()
                 if not abs_path.exists() or not abs_path.is_file():
                     continue
@@ -160,7 +167,13 @@ class DeepSeekClient:
                     mime = "video/mp4"
                 data = abs_path.read_bytes()
                 b64 = base64.b64encode(data).decode("utf-8")
-                parts.append({"type": "video_url", "video_url": {"url": f"data:{mime};base64,{b64}"}})
+                payload: Dict[str, Any] = {
+                    "type": "video_url",
+                    "video_url": {"url": f"data:{mime};base64,{b64}"}
+                }
+                if self.model_key == "qwen3-vl-plus":
+                    payload["fps"] = qwen_video_fps
+                parts.append(payload)
             except Exception:
                 continue
         return parts if parts else text
@@ -184,6 +197,55 @@ class DeepSeekClient:
         
         return json.dumps(data, ensure_ascii=False)
 
+    def _extract_reasoning_delta(self, delta: Dict[str, Any]) -> str:
+        """统一提取思考内容，兼容 reasoning_content / reasoning_details。"""
+        if not isinstance(delta, dict):
+            return ""
+        if "reasoning_content" in delta:
+            return delta.get("reasoning_content") or ""
+        details = delta.get("reasoning_details")
+        if isinstance(details, list):
+            parts: List[str] = []
+            for item in details:
+                if isinstance(item, dict):
+                    text = item.get("text")
+                    if text:
+                        parts.append(text)
+            if parts:
+                return "".join(parts)
+        return ""
+
+    def _merge_system_messages(self, messages: List[Dict]) -> List[Dict]:
+        """
+        将多个 system 消息合并为一个（部分模型仅支持单条 system）。
+        保留原有顺序，把合并后的 system 放在第一条 system 的位置。
+        """
+        if not messages:
+            return messages
+        merged_contents: List[str] = []
+        new_messages: List[Dict] = []
+        first_system_index: Optional[int] = None
+        for msg in messages:
+            if msg.get("role") == "system":
+                if first_system_index is None:
+                    first_system_index = len(new_messages)
+                content = msg.get("content", "")
+                if isinstance(content, str):
+                    merged_contents.append(content)
+                else:
+                    merged_contents.append(json.dumps(content, ensure_ascii=False))
+            else:
+                new_messages.append(msg)
+        if not merged_contents:
+            return messages
+        merged = {
+            "role": "system",
+            "content": "\n\n".join(c for c in merged_contents if c)
+        }
+        insert_at = first_system_index if first_system_index is not None else 0
+        new_messages.insert(insert_at, merged)
+        return new_messages
+    
     def set_deep_thinking_mode(self, enabled: bool):
         """配置深度思考模式（持续使用思考模型）。"""
         self.deep_thinking_mode = bool(enabled)
@@ -445,16 +507,24 @@ class DeepSeekClient:
             else:
                 max_tokens = min(max_tokens, available)
 
+        lower_base_url = (api_config.get("base_url") or "").lower()
+        is_minimax = self.model_key == "minimax-m2.5" or "minimax" in lower_base_url
+
+        final_messages = self._merge_system_messages(messages)
+
         payload = {
             "model": api_config["model_id"],
-            "messages": messages,
+            "messages": final_messages,
             "stream": stream,
-            "max_tokens": max_tokens
         }
+        if is_minimax:
+            payload["max_completion_tokens"] = max_tokens
+        else:
+            payload["max_tokens"] = max_tokens
         # 部分平台（如 Qwen、DeepSeek）需要显式请求 usage 才会在流式尾包返回
         if stream:
             should_include_usage = False
-            if self.model_key in {"qwen3-max", "qwen3-vl-plus", "deepseek"}:
+            if self.model_key in {"qwen3-max", "qwen3-vl-plus", "deepseek", "minimax-m2.5"}:
                 should_include_usage = True
             # 兜底：根据 base_url 识别 openai 兼容的提供商
             if api_config["base_url"]:
@@ -462,14 +532,18 @@ class DeepSeekClient:
                 if any(keyword in lower_url for keyword in ["dashscope", "aliyuncs", "deepseek.com"]):
                     should_include_usage = True
             if should_include_usage:
-                payload.setdefault("stream_options", {})["include_usage"] = True
+                if is_minimax:
+                    payload["include_usage"] = True
+                else:
+                    payload.setdefault("stream_options", {})["include_usage"] = True
         # 注入模型额外参数（如 Qwen enable_thinking）
         extra_params = self.thinking_extra_params if current_thinking_mode else self.fast_extra_params
         if extra_params:
             payload.update(extra_params)
         if tools:
             payload["tools"] = tools
-            payload["tool_choice"] = "auto"
+            if not is_minimax:
+                payload["tool_choice"] = "auto"
 
         # 将本次请求落盘，便于出错时快速定位
         dump_path = self._dump_request_payload(payload, api_config, headers)
@@ -636,15 +710,14 @@ class DeepSeekClient:
                 delta = chunk["choices"][0].get("delta", {})
                 
                 # 处理思考内容
-                if "reasoning_content" in delta:
+                reasoning_content = self._extract_reasoning_delta(delta)
-                    reasoning_content = delta["reasoning_content"]
+                if reasoning_content:
-                    if reasoning_content:  # 只处理非空内容
+                    if not in_thinking:
-                        if not in_thinking:
+                        self._print("💭 [正在思考]\n", end="", flush=True)
-                            self._print("💭 [正在思考]\n", end="", flush=True)
+                        in_thinking = True
-                            in_thinking = True
+                        thinking_printed = True
-                            thinking_printed = True
+                    current_thinking += reasoning_content
-                        current_thinking += reasoning_content
+                    self._print(reasoning_content, end="", flush=True)
-                        self._print(reasoning_content, end="", flush=True)
                 
                 # 处理正常内容 - 独立的if，不是elif
                 if "content" in delta:
@@ -910,14 +983,13 @@ class DeepSeekClient:
                 delta = chunk["choices"][0].get("delta", {})
                 
                 # 处理思考内容
-                if "reasoning_content" in delta:
+                reasoning_content = self._extract_reasoning_delta(delta)
-                    reasoning_content = delta["reasoning_content"]
+                if reasoning_content:
-                    if reasoning_content:  # 只处理非空内容
+                    if not in_thinking:
-                        if not in_thinking:
+                        self._print("💭 [正在思考]\n", end="", flush=True)
-                            self._print("💭 [正在思考]\n", end="", flush=True)
+                        in_thinking = True
-                            in_thinking = True
+                    thinking_content += reasoning_content
-                        thinking_content += reasoning_content
+                    self._print(reasoning_content, end="", flush=True)
-                        self._print(reasoning_content, end="", flush=True)
                 
                 # 处理正常内容 - 独立的if而不是elif
                 if "content" in delta:

utils/context_manager.py

@@ -1414,12 +1414,14 @@ class ContextManager:
         except Exception:
             return None
 
-    def _build_content_with_images(self, text: str, images: List[str], videos: Optional[List[str]] = None) -> Any:
+    def _build_content_with_images(self, text: str, images: List[str], videos: Optional[List[Any]] = None) -> Any:
         """将文本与图片/视频路径组合成多模态content，图片转换为data URI，视频转换为 data URL。"""
         videos = videos or []
         if not images and not videos:
             return text
         parts: List[Dict[str, Any]] = []
+        supports_video_fps = getattr(getattr(self, "main_terminal", None), "model_key", None) == "qwen3-vl-plus"
+        qwen_video_fps = 2
         if text:
             parts.append({"type": "text", "text": text})
         for path in images:
@@ -1438,8 +1440,14 @@ class ContextManager:
                 parts.append({"type": "image_url", "image_url": {"url": data_url}})
             except Exception:
                 continue
-        for path in videos:
+        for item in videos:
             try:
+                if isinstance(item, dict):
+                    path = item.get("path") or ""
+                else:
+                    path = item
+                if not path:
+                    continue
                 abs_path = Path(self.project_path) / path
                 if not abs_path.exists() or not abs_path.is_file():
                     continue
@@ -1451,7 +1459,10 @@ class ContextManager:
                 data = abs_path.read_bytes()
                 b64 = base64.b64encode(data).decode("utf-8")
                 data_url = f"data:{mime};base64,{b64}"
-                parts.append({"type": "video_url", "video_url": {"url": data_url}})
+                payload: Dict[str, Any] = {"type": "video_url", "video_url": {"url": data_url}}
+                if supports_video_fps:
+                    payload["fps"] = qwen_video_fps
+                parts.append(payload)
             except Exception:
                 continue
         return parts if parts else text
@@ -1501,7 +1512,7 @@ class ContextManager:
 
     def build_messages(self, context: Dict, user_input: str) -> List[Dict]:
         """构建消息列表（添加终端内容注入）"""
-        # 加载系统提示（Qwen-VL 使用专用提示）
+        # 加载系统提示（Qwen3.5 使用专用提示）
         model_key = getattr(self.main_terminal, "model_key", "kimi") if hasattr(self, "main_terminal") else "kimi"
         prompt_name = "main_system_qwenvl" if model_key in {"qwen3-vl-plus", "kimi-k2.5"} else "main_system"
         system_prompt = self.load_prompt(prompt_name)