Merge pull request #741 from SuanmoSuanyangTechnology/feature/agent-tool_xjn

fix(app)

api/app/controllers/ontology_controller.py

@@ -163,6 +163,7 @@ def _get_ontology_service(
             api_key=api_key_config.api_key,
             base_url=api_key_config.api_base,
             is_omni=api_key_config.is_omni,
+            support_thinking="thinking" in (api_key_config.capability or []),
             max_retries=3,
             timeout=60.0
         )

api/app/controllers/public_share_controller.py

@@ -453,6 +453,9 @@ async def chat(
         # 流式返回
         agent_config = agent_config_4_app_release(release)
 
+        if not (agent_config.model_parameters.get("deep_thinking", False) and payload.thinking):
+            agent_config.model_parameters["deep_thinking"] = False
+
         if payload.stream:
             async def event_generator():
                 async for event in app_chat_service.agnet_chat_stream(
@@ -634,7 +637,8 @@ async def config_query(
             "app_type": release.app.type,
             "variables": release.config.get("variables"),
             "memory": release.config.get("memory", {}).get("enabled"),
-            "features": release.config.get("features")
+            "features": release.config.get("features"),
+            "model_parameters": release.config.get("model_parameters")
         }
     elif release.app.type == AppType.MULTI_AGENT:
         content = {

api/app/controllers/service/app_api_controller.py

@@ -144,6 +144,11 @@ async def chat(
         # print(app.current_release.default_model_config_id)
         agent_config = agent_config_4_app_release(app.current_release)
         # print(agent_config.default_model_config_id)
+
+        # thinking 开关：仅当 agent 配置了 deep_thinking 且请求 thinking=True 时才启用
+        if not (agent_config.model_parameters.get("deep_thinking", False) and payload.thinking):
+            agent_config.model_parameters["deep_thinking"] = False
+
         # 流式返回
         if payload.stream:
             async def event_generator():

api/app/controllers/service/end_user_api_controller.py

@@ -42,7 +42,7 @@ async def create_end_user(
     payload = CreateEndUserRequest(**body)
     workspace_id = api_key_auth.workspace_id
 
-    logger.info(f"Create end user request - other_id: {payload.other_id}, workspace_id: {workspace_id}")
+    logger.info("Create end user request - other_id: %s, workspace_id: %s", payload.other_id, workspace_id)
 
     # Resolve memory_config_id: explicit > workspace default
     memory_config_id = None

api/app/core/agent/langchain_agent.py

@@ -37,7 +37,10 @@ class LangChainAgent:
             tools: Optional[Sequence[BaseTool]] = None,
             streaming: bool = False,
             max_iterations: Optional[int] = None,  # 最大迭代次数（None 表示自动计算）
-            max_tool_consecutive_calls: int = 3  # 单个工具最大连续调用次数
+            max_tool_consecutive_calls: int = 3,  # 单个工具最大连续调用次数
+            deep_thinking: bool = False,  # 是否启用深度思考模式
+            thinking_budget_tokens: Optional[int] = None,  # 深度思考 token 预算
+            capability: Optional[List[str]] = None  # 模型能力列表，用于校验是否支持深度思考
     ):
         """初始化 LangChain Agent
 
@@ -60,6 +63,7 @@ class LangChainAgent:
         self.streaming = streaming
         self.is_omni = is_omni
         self.max_tool_consecutive_calls = max_tool_consecutive_calls
+        self.deep_thinking = deep_thinking and ("thinking" in (capability or []))
 
         # 工具调用计数器：记录每个工具的连续调用次数
         self.tool_call_counter: Dict[str, int] = {}
@@ -82,6 +86,13 @@ class LangChainAgent:
             f"auto_calculated={max_iterations is None}"
         )
 
+        # 根据 capability 校验是否真正支持深度思考
+        actual_deep_thinking = self.deep_thinking
+        if deep_thinking and not actual_deep_thinking:
+            logger.warning(
+                f"模型 {model_name} 不支持深度思考（capability 中无 'thinking'），已自动关闭 deep_thinking"
+            )
+
         # 创建 RedBearLLM（支持多提供商）
         model_config = RedBearModelConfig(
             model_name=model_name,
@@ -89,10 +100,13 @@ class LangChainAgent:
             api_key=api_key,
             base_url=api_base,
             is_omni=is_omni,
+            deep_thinking=actual_deep_thinking,
+            thinking_budget_tokens=thinking_budget_tokens if actual_deep_thinking else None,
+            support_thinking="thinking" in (capability or []),
             extra_params={
                 "temperature": temperature,
                 "max_tokens": max_tokens,
-                "streaming": streaming  # 使用参数控制流式
+                "streaming": streaming
             }
         )
 
@@ -249,6 +263,33 @@ class LangChainAgent:
 
         return messages
 
+    @staticmethod
+    def _extract_tokens_from_message(msg) -> int:
+        """从 AIMessage 或类似对象中提取 total_tokens，兼容多种 provider 格式
+
+        支持的格式：
+        - response_metadata.token_usage.total_tokens (OpenAI/ChatOpenAI)
+        - response_metadata.usage.total_tokens (部分 provider)
+        - usage_metadata.total_tokens (LangChain 新版)
+        """
+        total = 0
+        # 1. response_metadata
+        response_meta = getattr(msg, "response_metadata", None)
+        if response_meta and isinstance(response_meta, dict):
+            # 尝试 token_usage 路径
+            token_usage = response_meta.get("token_usage") or response_meta.get("usage", {})
+            if isinstance(token_usage, dict):
+                total = token_usage.get("total_tokens", 0)
+        # 2. usage_metadata（LangChain 新版 AIMessage 属性）
+        if not total:
+            usage_meta = getattr(msg, "usage_metadata", None)
+            if usage_meta:
+                if isinstance(usage_meta, dict):
+                    total = usage_meta.get("total_tokens", 0)
+                else:
+                    total = getattr(usage_meta, "total_tokens", 0)
+        return total or 0
+
     def _build_multimodal_content(self, text: str, files: List[Dict[str, Any]]) -> List[Dict[str, Any]]:
         """
         构建多模态消息内容
@@ -283,6 +324,17 @@ class LangChainAgent:
 
         return content_parts
 
+    @staticmethod
+    def _extract_reasoning_content(msg) -> str:
+        """从 AIMessage 中提取深度思考内容（reasoning_content）
+
+        所有 provider 统一通过 additional_kwargs.reasoning_content 传递：
+        - DeepSeek-R1 / QwQ: 原生字段
+        - Volcano (Doubao-thinking): 由 VolcanoChatOpenAI 从 delta.reasoning_content 注入
+        """
+        additional = getattr(msg, "additional_kwargs", None) or {}
+        return additional.get("reasoning_content") or additional.get("reasoning", "")
+
     async def chat(
             self,
             message: str,
@@ -348,6 +400,7 @@ class LangChainAgent:
 
             logger.debug(f"输出消息数量: {len(output_messages)}")
             total_tokens = 0
+            reasoning_content = ""
             for msg in reversed(output_messages):
                 if isinstance(msg, AIMessage):
                     logger.debug(f"找到 AI 消息，content 类型: {type(msg.content)}")
@@ -382,8 +435,8 @@ class LangChainAgent:
                     else:
                         content = str(msg.content)
                         logger.debug(f"转换为字符串: {content[:100]}...")
-                    response_meta = msg.response_metadata if hasattr(msg, 'response_metadata') else None
+                    total_tokens = self._extract_tokens_from_message(msg)
-                    total_tokens = response_meta.get("token_usage", {}).get("total_tokens", 0) if response_meta else 0
+                    reasoning_content = self._extract_reasoning_content(msg) if self.deep_thinking else ""
                     break
 
             logger.info(f"最终提取的内容长度: {len(content)}")
@@ -399,6 +452,8 @@ class LangChainAgent:
                     "total_tokens": total_tokens
                 }
             }
+            if reasoning_content:
+                response["reasoning_content"] = reasoning_content
 
             logger.debug(
                 "Agent 调用完成",
@@ -420,7 +475,7 @@ class LangChainAgent:
             history: Optional[List[Dict[str, str]]] = None,
             context: Optional[str] = None,
             files: Optional[List[Dict[str, Any]]] = None
-    ) -> AsyncGenerator[str, None]:
+    ) -> AsyncGenerator[str | int, None]:
         """执行流式对话
 
         Args:
@@ -431,6 +486,8 @@ class LangChainAgent:
 
         Yields:
             str: 消息内容块
+            int: token 统计
+            Dict: 深度思考内容 {"type": "reasoning", "content": "..."}
         """
         logger.info("=" * 80)
         logger.info(" chat_stream 方法开始执行")
@@ -451,6 +508,7 @@ class LangChainAgent:
             # 统一使用 agent 的 astream_events 实现流式输出
             logger.debug("使用 Agent astream_events 实现流式输出")
             full_content = ''
+            full_reasoning = ''
             try:
                 last_event = {}
                 async for event in self.agent.astream_events(
@@ -467,6 +525,13 @@ class LangChainAgent:
                         # LLM 流式输出
                         chunk = event.get("data", {}).get("chunk")
                         if chunk and hasattr(chunk, "content"):
+                            # 提取深度思考内容（仅在启用深度思考时）
+                            if self.deep_thinking:
+                                reasoning_chunk = self._extract_reasoning_content(chunk)
+                                if reasoning_chunk:
+                                    full_reasoning += reasoning_chunk
+                                    yield {"type": "reasoning", "content": reasoning_chunk}
+
                             # 处理多模态响应：content 可能是字符串或列表
                             chunk_content = chunk.content
                             if isinstance(chunk_content, str) and chunk_content:
@@ -497,6 +562,13 @@ class LangChainAgent:
                         chunk = event.get("data", {}).get("chunk")
                         if chunk:
                             if hasattr(chunk, "content"):
+                                # 提取深度思考内容（仅在启用深度思考时）
+                                if self.deep_thinking:
+                                    reasoning_chunk = self._extract_reasoning_content(chunk)
+                                    if reasoning_chunk:
+                                        full_reasoning += reasoning_chunk
+                                        yield {"type": "reasoning", "content": reasoning_chunk}
+
                                 chunk_content = chunk.content
                                 if isinstance(chunk_content, str) and chunk_content:
                                     full_content += chunk_content
@@ -535,12 +607,9 @@ class LangChainAgent:
                 output_messages = last_event.get("data", {}).get("output", {}).get("messages", [])
                 for msg in reversed(output_messages):
                     if isinstance(msg, AIMessage):
-                        response_meta = msg.response_metadata if hasattr(msg, 'response_metadata') else None
+                        stream_total_tokens = self._extract_tokens_from_message(msg)
-                        total_tokens = response_meta.get("token_usage", {}).get(
+                        logger.info(f"流式 token 统计: total_tokens={stream_total_tokens}")
-                            "total_tokens",
+                        yield stream_total_tokens
-                            0
-                        ) if response_meta else 0
-                        yield total_tokens
                         break
 
             except Exception as e:

api/app/core/memory/src/search.py

@@ -758,8 +758,7 @@ async def run_hybrid_search(
                     model_name=embedder_config_dict["model_name"],
                     provider=embedder_config_dict["provider"],
                     api_key=embedder_config_dict["api_key"],
-                    base_url=embedder_config_dict["base_url"],
+                    base_url=embedder_config_dict["base_url"]
-                    type="llm"
                 )
                 config_load_time = time.time() - config_load_start
                 logger.info(f"[PERF] Config loading took {config_load_time:.4f}s")

api/app/core/models/base.py

@@ -14,6 +14,7 @@ from pydantic import BaseModel, Field
 from app.core.error_codes import BizCode
 from app.core.exceptions import BusinessException
 from app.models.models_model import ModelProvider, ModelType
+from app.core.models.volcano_chat import VolcanoChatOpenAI
 
 T = TypeVar("T")
 
@@ -25,6 +26,9 @@ class RedBearModelConfig(BaseModel):
     api_key: str
     base_url: Optional[str] = None
     is_omni: bool = False  # 是否为 Omni 模型
+    deep_thinking: bool = False  # 是否启用深度思考模式
+    thinking_budget_tokens: Optional[int] = None  # 深度思考 token 预算
+    support_thinking: bool = False  # 模型是否支持 enable_thinking 参数（capability 含 thinking）
     # 请求超时时间（秒）- 默认120秒以支持复杂的LLM调用，可通过环境变量 LLM_TIMEOUT 配置
     timeout: float = Field(default_factory=lambda: float(os.getenv("LLM_TIMEOUT", "120.0")))
     # 最大重试次数 - 默认2次以避免过长等待，可通过环境变量 LLM_MAX_RETRIES 配置
@@ -44,7 +48,7 @@ class RedBearModelFactory:
         # 打印供应商信息用于调试
         from app.core.logging_config import get_business_logger
         logger = get_business_logger()
-        logger.debug(f"获取模型参数 - Provider: {provider}, Model: {config.model_name}, is_omni: {config.is_omni}")
+        logger.debug(f"获取模型参数 - Provider: {provider}, Model: {config.model_name}, is_omni: {config.is_omni}, deep_thinking: {config.deep_thinking}")
 
         # dashscope 的 omni 模型使用 OpenAI 兼容模式
         if provider == ModelProvider.DASHSCOPE and config.is_omni:
@@ -58,7 +62,7 @@ class RedBearModelFactory:
                 write=60.0,
                 pool=10.0,
             )
-            return {
+            params: Dict[str, Any] = {
                 "model": config.model_name,
                 "base_url": config.base_url,
                 "api_key": config.api_key,
@@ -66,6 +70,21 @@ class RedBearModelFactory:
                 "max_retries": config.max_retries,
                 **config.extra_params
             }
+            # 流式模式下启用 stream_usage 以获取 token 统计
+            is_streaming = bool(config.extra_params.get("streaming"))
+            if is_streaming:
+                params["stream_usage"] = True
+            # 只有支持 thinking 的模型才传 enable_thinking
+            if config.support_thinking:
+                model_kwargs: Dict[str, Any] = config.extra_params.get("model_kwargs", {})
+                if is_streaming:
+                    model_kwargs["enable_thinking"] = config.deep_thinking
+                    if config.deep_thinking and config.thinking_budget_tokens:
+                        model_kwargs["thinking_budget"] = config.thinking_budget_tokens
+                else:
+                    model_kwargs["enable_thinking"] = False
+                params["model_kwargs"] = model_kwargs
+            return params
 
         if provider in [ModelProvider.OPENAI, ModelProvider.XINFERENCE, ModelProvider.GPUSTACK, ModelProvider.OLLAMA, ModelProvider.VOLCANO]:
             # 使用 httpx.Timeout 对象来设置详细的超时配置
@@ -78,7 +97,7 @@ class RedBearModelFactory:
                 write=60.0,  # 写入超时：60秒
                 pool=10.0,  # 连接池超时：10秒
             )
-            return {
+            params: Dict[str, Any] = {
                 "model": config.model_name,
                 "base_url": config.base_url,
                 "api_key": config.api_key,
@@ -86,16 +105,47 @@ class RedBearModelFactory:
                 "max_retries": config.max_retries,
                 **config.extra_params
             }
+            # 流式模式下启用 stream_usage 以获取 token 统计
+            if config.extra_params.get("streaming"):
+                params["stream_usage"] = True
+            # 深度思考模式
+            is_streaming = bool(config.extra_params.get("streaming"))
+            if is_streaming:
+                if provider == ModelProvider.VOLCANO:
+                    # 火山引擎深度思考仅流式调用支持，非流式时不传 thinking 参数
+                    thinking_config: Dict[str, Any] = {
+                        "type": "enabled" if config.deep_thinking else "disabled"
+                    }
+                    if config.deep_thinking and config.thinking_budget_tokens:
+                        thinking_config["budget_tokens"] = config.thinking_budget_tokens
+                    params["extra_body"] = {"thinking": thinking_config}
+                else:
+                    # 始终显式传递 enable_thinking，不支持该参数的模型（如 DeepSeek-R1）会直接忽略
+                    model_kwargs: Dict[str, Any] = config.extra_params.get("model_kwargs", {})
+                    model_kwargs["enable_thinking"] = config.deep_thinking
+                    if config.deep_thinking and config.thinking_budget_tokens:
+                        model_kwargs["thinking_budget"] = config.thinking_budget_tokens
+                    params["model_kwargs"] = model_kwargs
+            return params
         elif provider == ModelProvider.DASHSCOPE:
-            # DashScope (通义千问) 使用自己的参数格式
+            params = {
-            # 注意: DashScopeEmbeddings 不支持 timeout 和 base_url 参数
-            # 只支持: model, dashscope_api_key, max_retries, client
-            return {
                 "model": config.model_name,
                 "dashscope_api_key": config.api_key,
                 "max_retries": config.max_retries,
                 **config.extra_params
             }
+            # 只有支持 thinking 的模型才传 enable_thinking
+            if config.support_thinking:
+                is_streaming = bool(config.extra_params.get("streaming"))
+                model_kwargs: Dict[str, Any] = config.extra_params.get("model_kwargs", {})
+                if is_streaming:
+                    model_kwargs["enable_thinking"] = config.deep_thinking
+                    if config.deep_thinking and config.thinking_budget_tokens:
+                        model_kwargs["thinking_budget"] = config.thinking_budget_tokens
+                else:
+                    model_kwargs["enable_thinking"] = False
+                params["model_kwargs"] = model_kwargs
+            return params
         elif provider == ModelProvider.BEDROCK:
             # Bedrock 使用 AWS 凭证
             # api_key 格式: "access_key_id:secret_access_key" 或只是 access_key_id
@@ -134,6 +184,13 @@ class RedBearModelFactory:
             elif "region_name" not in params:
                 params["region_name"] = "us-east-1"  # 默认区域
 
+            # 深度思考模式：Claude 3.7 Sonnet 等支持思考的模型
+            # 通过 additional_model_request_fields 传递 thinking 块，关闭时不传（Bedrock 无 disabled 选项）
+            if config.deep_thinking:
+                budget = config.thinking_budget_tokens or 10000
+                params["additional_model_request_fields"] = {
+                    "thinking": {"type": "enabled", "budget_tokens": budget}
+                }
             return params
         else:
             raise BusinessException(f"不支持的提供商: {provider}", code=BizCode.PROVIDER_NOT_SUPPORTED)
@@ -160,7 +217,9 @@ def get_provider_llm_class(config: RedBearModelConfig, type: ModelType = ModelTy
     # dashscope 的 omni 模型使用 OpenAI 兼容模式
     if provider == ModelProvider.DASHSCOPE and config.is_omni:
         return ChatOpenAI
-    if provider in [ModelProvider.OPENAI, ModelProvider.XINFERENCE, ModelProvider.GPUSTACK, ModelProvider.VOLCANO]:
+    if provider == ModelProvider.VOLCANO:
+        return VolcanoChatOpenAI
+    if provider in [ModelProvider.OPENAI, ModelProvider.XINFERENCE, ModelProvider.GPUSTACK]:
         if type == ModelType.LLM:
             return OpenAI
         elif type == ModelType.CHAT:

api/app/core/models/scripts/bedrock_models.yaml

@@ -11,6 +11,7 @@ models:
   tags:
   - 大语言模型
   logo: bedrock
+
 - name: amazon nova
   type: llm
   provider: bedrock
@@ -27,6 +28,7 @@ models:
   - stream-tool-call
   - vision
   logo: bedrock
+
 - name: anthropic claude
   type: llm
   provider: bedrock
@@ -35,6 +37,7 @@ models:
   is_official: true
   capability:
     - vision
+    - thinking
   is_omni: false
   tags:
   - 大语言模型
@@ -44,6 +47,7 @@ models:
   - stream-tool-call
   - document
   logo: bedrock
+
 - name: cohere
   type: llm
   provider: bedrock
@@ -58,6 +62,7 @@ models:
   - tool-call
   - stream-tool-call
   logo: bedrock
+
 - name: deepseek
   type: llm
   provider: bedrock
@@ -66,6 +71,7 @@ models:
   is_official: true
   capability:
     - vision
+    - thinking
   is_omni: false
   tags:
   - 大语言模型
@@ -74,6 +80,7 @@ models:
   - tool-call
   - stream-tool-call
   logo: bedrock
+
 - name: meta
   type: llm
   provider: bedrock
@@ -87,6 +94,7 @@ models:
   - agent-thought
   - tool-call
   logo: bedrock
+
 - name: mistral
   type: llm
   provider: bedrock
@@ -100,6 +108,7 @@ models:
   - agent-thought
   - tool-call
   logo: bedrock
+
 - name: openai
   type: llm
   provider: bedrock
@@ -114,6 +123,7 @@ models:
   - tool-call
   - stream-tool-call
   logo: bedrock
+
 - name: qwen
   type: llm
   provider: bedrock
@@ -128,6 +138,7 @@ models:
   - tool-call
   - stream-tool-call
   logo: bedrock
+
 - name: amazon.rerank-v1:0
   type: rerank
   provider: bedrock
@@ -139,6 +150,7 @@ models:
   tags:
   - 重排序模型
   logo: bedrock
+
 - name: cohere.rerank-v3-5:0
   type: rerank
   provider: bedrock
@@ -150,6 +162,7 @@ models:
   tags:
   - 重排序模型
   logo: bedrock
+
 - name: amazon.nova-2-multimodal-embeddings-v1:0
   type: embedding
   provider: bedrock
@@ -163,6 +176,7 @@ models:
   - 文本嵌入模型
   - vision
   logo: bedrock
+
 - name: amazon.titan-embed-text-v1
   type: embedding
   provider: bedrock
@@ -174,6 +188,7 @@ models:
   tags:
   - 文本嵌入模型
   logo: bedrock
+
 - name: amazon.titan-embed-text-v2:0
   type: embedding
   provider: bedrock
@@ -185,6 +200,7 @@ models:
   tags:
   - 文本嵌入模型
   logo: bedrock
+
 - name: cohere.embed-english-v3
   type: embedding
   provider: bedrock
@@ -196,6 +212,7 @@ models:
   tags:
   - 文本嵌入模型
   logo: bedrock
+
 - name: cohere.embed-multilingual-v3
   type: embedding
   provider: bedrock

api/app/core/models/scripts/dashscope_models.yaml

@@ -6,36 +6,42 @@ models:
   description: DeepSeek-R1-Distill-Qwen-14B大语言模型，支持智能体思考，32000上下文窗口，对话模式
   is_deprecated: false
   is_official: true
-  capability: []
+  capability:
+  - thinking
   is_omni: false
   tags:
   - 大语言模型
   - agent-thought
   logo: dashscope
+
 - name: deepseek-r1-distill-qwen-32b
   type: llm
   provider: dashscope
   description: DeepSeek-R1-Distill-Qwen-32B大语言模型，支持智能体思考，32000上下文窗口，对话模式
   is_deprecated: false
   is_official: true
-  capability: []
+  capability:
+  - thinking
   is_omni: false
   tags:
   - 大语言模型
   - agent-thought
   logo: dashscope
+
 - name: deepseek-r1
   type: llm
   provider: dashscope
   description: DeepSeek-R1大语言模型，支持智能体思考，131072超大上下文窗口，对话模式
   is_deprecated: false
   is_official: true
-  capability: []
+  capability:
+  - thinking
   is_omni: false
   tags:
   - 大语言模型
   - agent-thought
   logo: dashscope
+
 - name: deepseek-v3.1
   type: llm
   provider: dashscope
@@ -48,6 +54,7 @@ models:
   - 大语言模型
   - agent-thought
   logo: dashscope
+
 - name: deepseek-v3.2-exp
   type: llm
   provider: dashscope
@@ -60,6 +67,7 @@ models:
   - 大语言模型
   - agent-thought
   logo: dashscope
+
 - name: deepseek-v3.2
   type: llm
   provider: dashscope
@@ -72,6 +80,7 @@ models:
   - 大语言模型
   - agent-thought
   logo: dashscope
+
 - name: deepseek-v3
   type: llm
   provider: dashscope
@@ -84,6 +93,7 @@ models:
   - 大语言模型
   - agent-thought
   logo: dashscope
+
 - name: farui-plus
   type: llm
   provider: dashscope
@@ -98,6 +108,7 @@ models:
   - agent-thought
   - stream-tool-call
   logo: dashscope
+
 - name: glm-4.7
   type: llm
   provider: dashscope
@@ -112,6 +123,7 @@ models:
   - agent-thought
   - stream-tool-call
   logo: dashscope
+
 - name: qvq-max-latest
   type: llm
   provider: dashscope
@@ -119,7 +131,8 @@ models:
   is_deprecated: false
   is_official: true
   capability:
-    - vision
+  - vision
+  - thinking
   is_omni: false
   tags:
   - 大语言模型
@@ -127,6 +140,7 @@ models:
   - agent-thought
   - stream-tool-call
   logo: dashscope
+
 - name: qvq-max
   type: llm
   provider: dashscope
@@ -134,7 +148,8 @@ models:
   is_deprecated: false
   is_official: true
   capability:
-    - vision
+  - vision
+  - thinking
   is_omni: false
   tags:
   - 大语言模型
@@ -142,6 +157,7 @@ models:
   - agent-thought
   - stream-tool-call
   logo: dashscope
+
 - name: qwen-coder-turbo-0919
   type: llm
   provider: dashscope
@@ -155,13 +171,15 @@ models:
   - 代码模型
   - agent-thought
   logo: dashscope
+
 - name: qwen-max-latest
   type: llm
   provider: dashscope
   description: qwen-max-latest大语言模型，支持多工具调用、智能体思考、流式工具调用，131072上下文窗口，对话模式，支持联网搜索
   is_deprecated: false
   is_official: true
-  capability: []
+  capability:
+  - thinking
   is_omni: false
   tags:
   - 大语言模型
@@ -169,6 +187,7 @@ models:
   - agent-thought
   - stream-tool-call
   logo: dashscope
+
 - name: qwen-max-longcontext
   type: llm
   provider: dashscope
@@ -183,13 +202,15 @@ models:
   - agent-thought
   - stream-tool-call
   logo: dashscope
+
 - name: qwen-max
   type: llm
   provider: dashscope
   description: qwen-max大语言模型，支持多工具调用、智能体思考、流式工具调用，32768上下文窗口，对话模式，支持联网搜索
   is_deprecated: false
   is_official: true
-  capability: []
+  capability:
+  - thinking
   is_omni: false
   tags:
   - 大语言模型
@@ -197,6 +218,7 @@ models:
   - agent-thought
   - stream-tool-call
   logo: dashscope
+
 - name: qwen-mt-plus
   type: llm
   provider: dashscope
@@ -210,6 +232,7 @@ models:
   - 翻译模型
   - agent-thought
   logo: dashscope
+
 - name: qwen-mt-turbo
   type: llm
   provider: dashscope
@@ -223,6 +246,7 @@ models:
   - 翻译模型
   - agent-thought
   logo: dashscope
+
 - name: qwen-plus-0112
   type: llm
   provider: dashscope
@@ -237,6 +261,7 @@ models:
   - agent-thought
   - stream-tool-call
   logo: dashscope
+
 - name: qwen-plus-0125
   type: llm
   provider: dashscope
@@ -251,6 +276,7 @@ models:
   - agent-thought
   - stream-tool-call
   logo: dashscope
+
 - name: qwen-plus-0723
   type: llm
   provider: dashscope
@@ -265,6 +291,7 @@ models:
   - agent-thought
   - stream-tool-call
   logo: dashscope
+
 - name: qwen-plus-0806
   type: llm
   provider: dashscope
@@ -279,6 +306,7 @@ models:
   - agent-thought
   - stream-tool-call
   logo: dashscope
+
 - name: qwen-plus-0919
   type: llm
   provider: dashscope
@@ -293,6 +321,7 @@ models:
   - agent-thought
   - stream-tool-call
   logo: dashscope
+
 - name: qwen-plus-1125
   type: llm
   provider: dashscope
@@ -307,6 +336,7 @@ models:
   - agent-thought
   - stream-tool-call
   logo: dashscope
+
 - name: qwen-plus-1127
   type: llm
   provider: dashscope
@@ -321,6 +351,7 @@ models:
   - agent-thought
   - stream-tool-call
   logo: dashscope
+
 - name: qwen-plus-1220
   type: llm
   provider: dashscope
@@ -335,6 +366,7 @@ models:
   - agent-thought
   - stream-tool-call
   logo: dashscope
+
 - name: qwen-vl-max
   type: chat
   provider: dashscope
@@ -342,8 +374,8 @@ models:
   is_deprecated: false
   is_official: true
   capability:
-    - vision
+  - vision
-    - video
+  - video
   is_omni: false
   tags:
   - 大语言模型
@@ -352,6 +384,7 @@ models:
   - agent-thought
   - video
   logo: dashscope
+
 - name: qwen-vl-plus-0809
   type: chat
   provider: dashscope
@@ -359,8 +392,8 @@ models:
   is_deprecated: true
   is_official: true
   capability:
-    - vision
+  - vision
-    - video
+  - video
   is_omni: false
   tags:
   - 大语言模型
@@ -369,6 +402,7 @@ models:
   - agent-thought
   - video
   logo: dashscope
+
 - name: qwen-vl-plus-2025-01-02
   type: chat
   provider: dashscope
@@ -376,8 +410,8 @@ models:
   is_deprecated: false
   is_official: true
   capability:
-    - vision
+  - vision
-    - video
+  - video
   is_omni: false
   tags:
   - 大语言模型
@@ -386,6 +420,7 @@ models:
   - agent-thought
   - video
   logo: dashscope
+
 - name: qwen-vl-plus-2025-01-25
   type: chat
   provider: dashscope
@@ -393,8 +428,8 @@ models:
   is_deprecated: false
   is_official: true
   capability:
-    - vision
+  - vision
-    - video
+  - video
   is_omni: false
   tags:
   - 大语言模型
@@ -403,6 +438,7 @@ models:
   - agent-thought
   - video
   logo: dashscope
+
 - name: qwen-vl-plus-latest
   type: chat
   provider: dashscope
@@ -410,8 +446,8 @@ models:
   is_deprecated: false
   is_official: true
   capability:
-    - vision
+  - vision
-    - video
+  - video
   is_omni: false
   tags:
   - 大语言模型
@@ -420,6 +456,7 @@ models:
   - agent-thought
   - video
   logo: dashscope
+
 - name: qwen-vl-plus
   type: chat
   provider: dashscope
@@ -427,8 +464,8 @@ models:
   is_deprecated: false
   is_official: true
   capability:
-    - vision
+  - vision
-    - video
+  - video
   is_omni: false
   tags:
   - 大语言模型
@@ -437,6 +474,7 @@ models:
   - agent-thought
   - video
   logo: dashscope
+
 - name: qwen2.5-0.5b-instruct
   type: llm
   provider: dashscope
@@ -451,13 +489,15 @@ models:
   - agent-thought
   - stream-tool-call
   logo: dashscope
+
 - name: qwen3-14b
   type: llm
   provider: dashscope
   description: qwen3-14b大语言模型，支持多工具调用、智能体思考、流式工具调用，131072上下文窗口，对话模式
   is_deprecated: false
   is_official: true
-  capability: []
+  capability:
+  - thinking
   is_omni: false
   tags:
   - 大语言模型
@@ -465,13 +505,15 @@ models:
   - agent-thought
   - stream-tool-call
   logo: dashscope
+
 - name: qwen3-235b-a22b-instruct-2507
   type: llm
   provider: dashscope
   description: qwen3-235b-a22b-instruct-2507大语言模型，支持多工具调用、智能体思考、流式工具调用，131072上下文窗口，对话模式
   is_deprecated: false
   is_official: true
-  capability: []
+  capability:
+  - thinking
   is_omni: false
   tags:
   - 大语言模型
@@ -479,13 +521,15 @@ models:
   - agent-thought
   - stream-tool-call
   logo: dashscope
+
 - name: qwen3-235b-a22b-thinking-2507
   type: llm
   provider: dashscope
   description: qwen3-235b-a22b-thinking-2507大语言模型，支持多工具调用、智能体思考、流式工具调用，131072上下文窗口，对话模式
   is_deprecated: false
   is_official: true
-  capability: []
+  capability:
+  - thinking
   is_omni: false
   tags:
   - 大语言模型
@@ -493,13 +537,15 @@ models:
   - agent-thought
   - stream-tool-call
   logo: dashscope
+
 - name: qwen3-235b-a22b
   type: llm
   provider: dashscope
   description: qwen3-235b-a22b大语言模型，支持多工具调用、智能体思考、流式工具调用，131072上下文窗口，对话模式
   is_deprecated: false
   is_official: true
-  capability: []
+  capability:
+  - thinking
   is_omni: false
   tags:
   - 大语言模型
@@ -507,13 +553,15 @@ models:
   - agent-thought
   - stream-tool-call
   logo: dashscope
+
 - name: qwen3-30b-a3b-instruct-2507
   type: llm
   provider: dashscope
   description: qwen3-30b-a3b-instruct-2507大语言模型，支持多工具调用、智能体思考、流式工具调用，131072上下文窗口，对话模式
   is_deprecated: false
   is_official: true
-  capability: []
+  capability:
+  - thinking
   is_omni: false
   tags:
   - 大语言模型
@@ -521,13 +569,15 @@ models:
   - agent-thought
   - stream-tool-call
   logo: dashscope
+
 - name: qwen3-30b-a3b
   type: llm
   provider: dashscope
   description: qwen3-30b-a3b大语言模型，支持多工具调用、智能体思考、流式工具调用，131072上下文窗口，对话模式
   is_deprecated: false
   is_official: true
-  capability: []
+  capability:
+  - thinking
   is_omni: false
   tags:
   - 大语言模型
@@ -535,13 +585,15 @@ models:
   - agent-thought
   - stream-tool-call
   logo: dashscope
+
 - name: qwen3-32b
   type: llm
   provider: dashscope
   description: qwen3-32b大语言模型，支持多工具调用、智能体思考、流式工具调用，131072上下文窗口，对话模式
   is_deprecated: false
   is_official: true
-  capability: []
+  capability:
+  - thinking
   is_omni: false
   tags:
   - 大语言模型
@@ -549,13 +601,15 @@ models:
   - agent-thought
   - stream-tool-call
   logo: dashscope
+
 - name: qwen3-4b
   type: llm
   provider: dashscope
   description: qwen3-4b大语言模型，支持多工具调用、智能体思考、流式工具调用，131072上下文窗口，对话模式
   is_deprecated: false
   is_official: true
-  capability: []
+  capability:
+  - thinking
   is_omni: false
   tags:
   - 大语言模型
@@ -563,13 +617,15 @@ models:
   - agent-thought
   - stream-tool-call
   logo: dashscope
+
 - name: qwen3-8b
   type: llm
   provider: dashscope
   description: qwen3-8b大语言模型，支持多工具调用、智能体思考、流式工具调用，131072上下文窗口，对话模式
   is_deprecated: false
   is_official: true
-  capability: []
+  capability:
+  - thinking
   is_omni: false
   tags:
   - 大语言模型
@@ -577,65 +633,75 @@ models:
   - agent-thought
   - stream-tool-call
   logo: dashscope
+
 - name: qwen3-coder-30b-a3b-instruct
   type: llm
   provider: dashscope
   description: qwen3-coder-30b-a3b-instruct大语言模型，支持智能体思考，262144上下文窗口，对话模式
   is_deprecated: false
   is_official: true
-  capability: []
+  capability:
+  - thinking
   is_omni: false
   tags:
   - 大语言模型
   - 代码模型
   - agent-thought
   logo: dashscope
+
 - name: qwen3-coder-480b-a35b-instruct
   type: llm
   provider: dashscope
   description: qwen3-coder-480b-a35b-instruct大语言模型，支持智能体思考，262144上下文窗口，对话模式
   is_deprecated: false
   is_official: true
-  capability: []
+  capability:
+  - thinking
   is_omni: false
   tags:
   - 大语言模型
   - 代码模型
   - agent-thought
   logo: dashscope
+
 - name: qwen3-coder-plus-2025-09-23
   type: llm
   provider: dashscope
   description: qwen3-coder-plus-2025-09-23大语言模型，支持智能体思考，1000000上下文窗口，对话模式
   is_deprecated: false
   is_official: true
-  capability: []
+  capability:
+  - thinking
   is_omni: false
   tags:
   - 大语言模型
   - 代码模型
   - agent-thought
   logo: dashscope
+
 - name: qwen3-coder-plus
   type: llm
   provider: dashscope
   description: qwen3-coder-plus大语言模型，支持智能体思考，1000000上下文窗口，对话模式
   is_deprecated: false
   is_official: true
-  capability: []
+  capability:
+  - thinking
   is_omni: false
   tags:
   - 大语言模型
   - 代码模型
   - agent-thought
   logo: dashscope
+
 - name: qwen3-max-2025-09-23
   type: llm
   provider: dashscope
   description: qwen3-max-2025-09-23大语言模型，支持多工具调用、智能体思考、流式工具调用，262144上下文窗口，对话模式，支持联网搜索
   is_deprecated: false
   is_official: true
-  capability: []
+  capability:
+  - thinking
   is_omni: false
   tags:
   - 大语言模型
@@ -644,13 +710,15 @@ models:
   - stream-tool-call
   - 联网搜索
   logo: dashscope
+
 - name: qwen3-max-2026-01-23
   type: llm
   provider: dashscope
   description: qwen3-max-2026-01-23大语言模型，支持多工具调用、智能体思考、流式工具调用，262144上下文窗口，对话模式，支持联网搜索
   is_deprecated: false
   is_official: true
-  capability: []
+  capability:
+  - thinking
   is_omni: false
   tags:
   - 大语言模型
@@ -659,13 +727,15 @@ models:
   - stream-tool-call
   - 联网搜索
   logo: dashscope
+
 - name: qwen3-max-preview
   type: llm
   provider: dashscope
   description: qwen3-max-preview大语言模型，支持多工具调用、智能体思考、流式工具调用，262144上下文窗口，对话模式
   is_deprecated: false
   is_official: true
-  capability: []
+  capability:
+  - thinking
   is_omni: false
   tags:
   - 大语言模型
@@ -673,13 +743,15 @@ models:
   - agent-thought
   - stream-tool-call
   logo: dashscope
+
 - name: qwen3-max
   type: llm
   provider: dashscope
   description: qwen3-max大语言模型，支持多工具调用、智能体思考、流式工具调用，262144上下文窗口，对话模式，支持联网搜索
   is_deprecated: false
   is_official: true
-  capability: []
+  capability:
+  - thinking
   is_omni: false
   tags:
   - 大语言模型
@@ -688,13 +760,15 @@ models:
   - stream-tool-call
   - 联网搜索
   logo: dashscope
+
 - name: qwen3-next-80b-a3b-instruct
   type: llm
   provider: dashscope
   description: qwen3-next-80b-a3b-instruct大语言模型，支持多工具调用、智能体思考、流式工具调用，131072上下文窗口，对话模式
   is_deprecated: false
   is_official: true
-  capability: []
+  capability:
+  - thinking
   is_omni: false
   tags:
   - 大语言模型
@@ -702,13 +776,15 @@ models:
   - agent-thought
   - stream-tool-call
   logo: dashscope
+
 - name: qwen3-next-80b-a3b-thinking
   type: llm
   provider: dashscope
   description: qwen3-next-80b-a3b-thinking大语言模型，支持多工具调用、智能体思考、流式工具调用，131072上下文窗口，对话模式
   is_deprecated: false
   is_official: true
-  capability: []
+  capability:
+  - thinking
   is_omni: false
   tags:
   - 大语言模型
@@ -716,6 +792,7 @@ models:
   - agent-thought
   - stream-tool-call
   logo: dashscope
+
 - name: qwen3-omni-flash-2025-12-01
   type: llm
   provider: dashscope
@@ -723,9 +800,10 @@ models:
   is_deprecated: false
   is_official: true
   capability:
-    - vision
+  - vision
-    - video
+  - video
-    - audio
+  - audio
+  - thinking
   is_omni: true
   tags:
   - 大语言模型
@@ -735,6 +813,7 @@ models:
   - video
   - audio
   logo: dashscope
+
 - name: qwen3-vl-235b-a22b-instruct
   type: chat
   provider: dashscope
@@ -742,8 +821,9 @@ models:
   is_deprecated: false
   is_official: true
   capability:
-    - vision
+  - vision
-    - video
+  - video
+  - thinking
   is_omni: false
   tags:
   - 大语言模型
@@ -754,6 +834,7 @@ models:
   - vision
   - video
   logo: dashscope
+
 - name: qwen3-vl-235b-a22b-thinking
   type: chat
   provider: dashscope
@@ -761,8 +842,9 @@ models:
   is_deprecated: false
   is_official: true
   capability:
-    - vision
+  - vision
-    - video
+  - video
+  - thinking
   is_omni: false
   tags:
   - 大语言模型
@@ -773,6 +855,7 @@ models:
   - vision
   - video
   logo: dashscope
+
 - name: qwen3-vl-30b-a3b-instruct
   type: chat
   provider: dashscope
@@ -780,8 +863,9 @@ models:
   is_deprecated: false
   is_official: true
   capability:
-    - vision
+  - vision
-    - video
+  - video
+  - thinking
   is_omni: false
   tags:
   - 大语言模型
@@ -792,6 +876,7 @@ models:
   - vision
   - video
   logo: dashscope
+
 - name: qwen3-vl-30b-a3b-thinking
   type: chat
   provider: dashscope
@@ -799,8 +884,9 @@ models:
   is_deprecated: false
   is_official: true
   capability:
-    - vision
+  - vision
-    - video
+  - video
+  - thinking
   is_omni: false
   tags:
   - 大语言模型
@@ -811,6 +897,7 @@ models:
   - vision
   - video
   logo: dashscope
+
 - name: qwen3-vl-flash
   type: chat
   provider: dashscope
@@ -818,8 +905,9 @@ models:
   is_deprecated: false
   is_official: true
   capability:
-    - vision
+  - vision
-    - video
+  - video
+  - thinking
   is_omni: false
   tags:
   - 大语言模型
@@ -830,6 +918,7 @@ models:
   - vision
   - video
   logo: dashscope
+
 - name: qwen3-vl-plus-2025-09-23
   type: chat
   provider: dashscope
@@ -837,8 +926,9 @@ models:
   is_deprecated: false
   is_official: true
   capability:
-    - vision
+  - vision
-    - video
+  - video
+  - thinking
   is_omni: false
   tags:
   - 大语言模型
@@ -847,6 +937,7 @@ models:
   - agent-thought
   - video
   logo: dashscope
+
 - name: qwen3-vl-plus
   type: chat
   provider: dashscope
@@ -854,8 +945,9 @@ models:
   is_deprecated: false
   is_official: true
   capability:
-    - vision
+  - vision
-    - video
+  - video
+  - thinking
   is_omni: false
   tags:
   - 大语言模型
@@ -864,45 +956,52 @@ models:
   - agent-thought
   - video
   logo: dashscope
+
 - name: qwq-32b
   type: llm
   provider: dashscope
   description: qwq-32b大语言模型，支持智能体思考、流式工具调用，131072上下文窗口，对话模式
   is_deprecated: false
   is_official: true
-  capability: []
+  capability:
+  - thinking
   is_omni: false
   tags:
   - 大语言模型
   - agent-thought
   - stream-tool-call
   logo: dashscope
+
 - name: qwq-plus-0305
   type: llm
   provider: dashscope
   description: qwq-plus-0305大语言模型，支持智能体思考、流式工具调用，131072上下文窗口，对话模式
   is_deprecated: false
   is_official: true
-  capability: []
+  capability:
+  - thinking
   is_omni: false
   tags:
   - 大语言模型
   - agent-thought
   - stream-tool-call
   logo: dashscope
+
 - name: qwq-plus
   type: llm
   provider: dashscope
   description: qwq-plus大语言模型，支持智能体思考、流式工具调用，131072上下文窗口，对话模式
   is_deprecated: false
   is_official: true
-  capability: []
+  capability:
+  - thinking
   is_omni: false
   tags:
   - 大语言模型
   - agent-thought
   - stream-tool-call
   logo: dashscope
+
 - name: gte-rerank-v2
   type: rerank
   provider: dashscope
@@ -914,6 +1013,7 @@ models:
   tags:
   - 重排序模型
   logo: dashscope
+
 - name: gte-rerank
   type: rerank
   provider: dashscope
@@ -925,6 +1025,7 @@ models:
   tags:
   - 重排序模型
   logo: dashscope
+
 - name: multimodal-embedding-v1
   type: embedding
   provider: dashscope
@@ -932,13 +1033,14 @@ models:
   is_deprecated: false
   is_official: true
   capability:
-    - vision
+  - vision
   is_omni: false
   tags:
   - 嵌入模型
   - 多模态模型
   - vision
   logo: dashscope
+
 - name: text-embedding-v1
   type: embedding
   provider: dashscope
@@ -951,6 +1053,7 @@ models:
   - 嵌入模型
   - 文本嵌入
   logo: dashscope
+
 - name: text-embedding-v2
   type: embedding
   provider: dashscope
@@ -963,6 +1066,7 @@ models:
   - 嵌入模型
   - 文本嵌入
   logo: dashscope
+
 - name: text-embedding-v3
   type: embedding
   provider: dashscope
@@ -975,6 +1079,7 @@ models:
   - 嵌入模型
   - 文本嵌入
   logo: dashscope
+
 - name: text-embedding-v4
   type: embedding
   provider: dashscope

api/app/core/models/scripts/openai_models.yaml

@@ -20,6 +20,7 @@ models:
   - audio
   - video
   logo: openai
+
 - name: gpt-3.5-turbo-0125
   type: llm
   provider: openai
@@ -34,6 +35,7 @@ models:
   - agent-thought
   - stream-tool-call
   logo: openai
+
 - name: gpt-3.5-turbo-1106
   type: llm
   provider: openai
@@ -48,6 +50,7 @@ models:
   - agent-thought
   - stream-tool-call
   logo: openai
+
 - name: gpt-3.5-turbo-16k
   type: llm
   provider: openai
@@ -62,6 +65,7 @@ models:
   - agent-thought
   - stream-tool-call
   logo: openai
+
 - name: gpt-3.5-turbo-instruct
   type: llm
   provider: openai
@@ -73,6 +77,7 @@ models:
   tags:
   - 大语言模型
   logo: openai
+
 - name: gpt-3.5-turbo
   type: llm
   provider: openai
@@ -87,6 +92,7 @@ models:
   - agent-thought
   - stream-tool-call
   logo: openai
+
 - name: gpt-4-0125-preview
   type: llm
   provider: openai
@@ -101,6 +107,7 @@ models:
   - agent-thought
   - stream-tool-call
   logo: openai
+
 - name: gpt-4-1106-preview
   type: llm
   provider: openai
@@ -115,6 +122,7 @@ models:
   - agent-thought
   - stream-tool-call
   logo: openai
+
 - name: gpt-4-turbo-2024-04-09
   type: llm
   provider: openai
@@ -131,6 +139,7 @@ models:
   - stream-tool-call
   - vision
   logo: openai
+
 - name: gpt-4-turbo-preview
   type: llm
   provider: openai
@@ -145,6 +154,7 @@ models:
   - agent-thought
   - stream-tool-call
   logo: openai
+
 - name: gpt-4-turbo
   type: llm
   provider: openai
@@ -161,6 +171,7 @@ models:
   - stream-tool-call
   - vision
   logo: openai
+
 - name: o1-preview
   type: llm
   provider: openai
@@ -173,6 +184,7 @@ models:
   - 大语言模型
   - agent-thought
   logo: openai
+
 - name: o1
   type: llm
   provider: openai
@@ -181,6 +193,7 @@ models:
   is_official: true
   capability:
     - vision
+    - thinking
   is_omni: false
   tags:
   - 大语言模型
@@ -190,6 +203,7 @@ models:
   - vision
   - structured-output
   logo: openai
+
 - name: o3-2025-04-16
   type: llm
   provider: openai
@@ -198,6 +212,7 @@ models:
   is_official: true
   capability:
     - vision
+    - thinking
   is_omni: false
   tags:
   - 大语言模型
@@ -207,13 +222,15 @@ models:
   - stream-tool-call
   - structured-output
   logo: openai
+
 - name: o3-mini-2025-01-31
   type: llm
   provider: openai
   description: o3-mini-2025-01-31大语言模型，支持智能体思考、工具调用、流式工具调用、结构化输出，200000上下文窗口，对话模式
   is_deprecated: false
   is_official: true
-  capability: []
+  capability:
+    - thinking
   is_omni: false
   tags:
   - 大语言模型
@@ -222,13 +239,15 @@ models:
   - stream-tool-call
   - structured-output
   logo: openai
+
 - name: o3-mini
   type: llm
   provider: openai
   description: o3-mini大语言模型，支持智能体思考、工具调用、流式工具调用、结构化输出，200000上下文窗口，对话模式
   is_deprecated: false
   is_official: true
-  capability: []
+  capability:
+    - thinking
   is_omni: false
   tags:
   - 大语言模型
@@ -237,6 +256,7 @@ models:
   - stream-tool-call
   - structured-output
   logo: openai
+
 - name: o3-pro-2025-06-10
   type: llm
   provider: openai
@@ -245,6 +265,7 @@ models:
   is_official: true
   capability:
     - vision
+    - thinking
   is_omni: false
   tags:
   - 大语言模型
@@ -253,6 +274,7 @@ models:
   - vision
   - structured-output
   logo: openai
+
 - name: o3-pro
   type: llm
   provider: openai
@@ -261,6 +283,7 @@ models:
   is_official: true
   capability:
     - vision
+    - thinking
   is_omni: false
   tags:
   - 大语言模型
@@ -269,6 +292,7 @@ models:
   - vision
   - structured-output
   logo: openai
+
 - name: o3
   type: llm
   provider: openai
@@ -277,6 +301,7 @@ models:
   is_official: true
   capability:
     - vision
+    - thinking
   is_omni: false
   tags:
   - 大语言模型
@@ -286,6 +311,7 @@ models:
   - stream-tool-call
   - structured-output
   logo: openai
+
 - name: o4-mini-2025-04-16
   type: llm
   provider: openai
@@ -294,6 +320,7 @@ models:
   is_official: true
   capability:
     - vision
+    - thinking
   is_omni: false
   tags:
   - 大语言模型
@@ -303,6 +330,7 @@ models:
   - stream-tool-call
   - structured-output
   logo: openai
+
 - name: o4-mini
   type: llm
   provider: openai
@@ -311,6 +339,7 @@ models:
   is_official: true
   capability:
     - vision
+    - thinking
   is_omni: false
   tags:
   - 大语言模型
@@ -320,6 +349,7 @@ models:
   - stream-tool-call
   - structured-output
   logo: openai
+
 - name: text-embedding-3-large
   type: embedding
   provider: openai
@@ -331,6 +361,7 @@ models:
   tags:
   - 文本向量模型
   logo: openai
+
 - name: text-embedding-3-small
   type: embedding
   provider: openai
@@ -342,6 +373,7 @@ models:
   tags:
   - 文本向量模型
   logo: openai
+
 - name: text-embedding-ada-002
   type: embedding
   provider: openai

api/app/core/models/scripts/volcano_models.yaml

@@ -10,6 +10,7 @@ models:
   capability:
     - vision
     - video
+    - thinking
   is_omni: false
   tags:
   - 大语言模型
@@ -24,6 +25,7 @@ models:
   capability:
     - vision
     - video
+    - thinking
   is_omni: false
   tags:
   - 大语言模型
@@ -38,6 +40,7 @@ models:
   capability:
     - vision
     - video
+    - thinking
   is_omni: false
   tags:
   - 大语言模型
@@ -52,6 +55,7 @@ models:
   capability:
     - vision
     - video
+    - thinking
   is_omni: false
   tags:
   - 大语言模型
@@ -82,6 +86,7 @@ models:
   capability:
     - vision
     - video
+    - thinking
   is_omni: false
   tags:
   - 大语言模型
@@ -96,6 +101,7 @@ models:
   capability:
     - vision
     - video
+    - thinking
   is_omni: false
   tags:
   - 大语言模型
@@ -110,6 +116,7 @@ models:
   capability:
     - vision
     - video
+    - thinking
   is_omni: false
   tags:
   - 大语言模型
@@ -124,6 +131,7 @@ models:
   capability:
     - vision
     - video
+    - thinking
   is_omni: false
   tags:
   - 大语言模型
@@ -139,6 +147,7 @@ models:
   capability:
     - vision
     - video
+    - thinking
   is_omni: false
   tags:
   - 大语言模型

api/app/core/models/volcano_chat.py

@@ -0,0 +1,38 @@
+"""
+火山引擎 ChatOpenAI 扩展
+
+ChatOpenAI 在解析流式 SSE 时只取 delta.content，会丢弃 delta.reasoning_content。
+此类仅重写 _convert_chunk_to_generation_chunk，将 reasoning_content 补入 additional_kwargs。
+"""
+from __future__ import annotations
+
+from typing import Any, Optional
+
+from langchain_core.outputs import ChatGenerationChunk
+from langchain_openai import ChatOpenAI
+
+
+class VolcanoChatOpenAI(ChatOpenAI):
+    """火山引擎 Chat 模型，支持深度思考内容（reasoning_content）的流式透传。"""
+
+    def _convert_chunk_to_generation_chunk(
+        self,
+        chunk: dict,
+        default_chunk_class: type,
+        base_generation_info: Optional[dict],
+    ) -> Optional[ChatGenerationChunk]:
+        gen_chunk = super()._convert_chunk_to_generation_chunk(
+            chunk, default_chunk_class, base_generation_info
+        )
+        if gen_chunk is None:
+            return None
+
+        # 从原始 chunk 中提取 reasoning_content
+        choices = chunk.get("choices") or chunk.get("chunk", {}).get("choices", [])
+        if choices:
+            delta = choices[0].get("delta") or {}
+            reasoning: Any = delta.get("reasoning_content")
+            if reasoning:
+                gen_chunk.message.additional_kwargs["reasoning_content"] = reasoning
+
+        return gen_chunk

api/app/core/tools/mcp/client.py

@@ -99,7 +99,7 @@ class SimpleMCPClient:
             # 建立 SSE 连接
             response = await self._session.get(self.server_url)
             
-            if response.status not in (200, 202):
+            if not (200 <= response.status < 300):
                 error_text = await response.text()
                 raise MCPConnectionError(f"SSE 连接失败 {response.status}: {error_text}")
             
@@ -190,9 +190,7 @@ class SimpleMCPClient:
         
         try:
             async with self._session.post(self._endpoint_url, json=request) as response:
-                # MCP SSE 协议：POST 请求返回 200 或 202 均为正常
+                if not (200 <= response.status < 300):
-                # 202 Accepted 表示请求已接受，结果通过 SSE 流异步返回
-                if response.status not in (200, 202):
                     error_text = await response.text()
                     raise MCPConnectionError(f"请求失败 {response.status}: {error_text}")
             
@@ -207,7 +205,7 @@ class SimpleMCPClient:
             raise MCPConnectionError("endpoint URL 未初始化")
         
         async with self._session.post(self._endpoint_url, json=notification) as response:
-            if response.status not in (200, 202):
+            if not (200 <= response.status < 300):
                 logger.warning(f"通知发送失败: {response.status}")
     
     async def _initialize_modelscope_session(self):
@@ -225,7 +223,7 @@ class SimpleMCPClient:
         
         try:
             async with self._session.post(self.server_url, json=init_request) as response:
-                if response.status != 200:
+                if not (200 <= response.status < 300):
                     error_text = await response.text()
                     raise MCPConnectionError(f"初始化失败 {response.status}: {error_text}")
                 

api/app/core/workflow/nodes/llm/node.py

@@ -135,7 +135,8 @@ class LLMNode(BaseNode):
                 api_key=model_info.api_key,
                 base_url=model_info.api_base,
                 extra_params=extra_params,
-                is_omni=model_info.is_omni
+                is_omni=model_info.is_omni,
+                support_thinking="thinking" in (model_info.capability or []),
             ),
             type=model_info.model_type
         )

api/app/core/workflow/nodes/parameter_extractor/node.py

@@ -109,6 +109,7 @@ class ParameterExtractorNode(BaseNode):
             api_key = api_config.api_key
             api_base = api_config.api_base
             is_omni = api_config.is_omni
+            capability = api_config.capability
             model_type = config.type
 
         llm = RedBearLLM(
@@ -117,7 +118,8 @@ class ParameterExtractorNode(BaseNode):
                 provider=provider,
                 api_key=api_key,
                 base_url=api_base,
-                is_omni=is_omni
+                is_omni=is_omni,
+                support_thinking="thinking" in (capability or []),
             ),
             type=ModelType(model_type)
         )

api/app/core/workflow/nodes/question_classifier/node.py

@@ -62,6 +62,7 @@ class QuestionClassifierNode(BaseNode):
             api_key = api_config.api_key
             base_url = api_config.api_base
             is_omni = api_config.is_omni
+            capability = api_config.capability
             model_type = config.type
 
         return RedBearLLM(
@@ -70,7 +71,8 @@ class QuestionClassifierNode(BaseNode):
                 provider=provider,
                 api_key=api_key,
                 base_url=base_url,
-                is_omni=is_omni
+                is_omni=is_omni,
+                support_thinking="thinking" in (capability or []),
             ),
             type=ModelType(model_type)
         )

api/app/models/models_model.py

@@ -81,7 +81,7 @@ class ModelConfig(BaseModel):
     
     # 模型配置参数
     capability = Column(ARRAY(String), default=list, nullable=False, server_default=text("'{}'::varchar[]"),
-                        comment="模型能力列表（如['vision', 'audio', 'video']）")
+                        comment="模型能力列表（如['vision', 'audio', 'video', 'thinking']）")
     is_omni = Column(Boolean, default=False, nullable=False, server_default="false", comment="是否为Omni模型（使用特殊API调用）")
     config = Column(JSON, comment="模型配置参数")
     # - temperature : 控制生成文本的随机性。值越高，输出越随机、越有创造性；值越低，输出越确定、越保守。

api/app/schemas/app_schema.py

@@ -241,6 +241,8 @@ class ModelParameters(BaseModel):
     presence_penalty: float = Field(default=0.0, ge=-2.0, le=2.0, description="存在惩罚")
     n: int = Field(default=1, ge=1, le=10, description="生成的回复数量")
     stop: Optional[List[str]] = Field(default=None, description="停止序列")
+    deep_thinking: bool = Field(default=False, description="是否启用深度思考模式（需模型支持，如 DeepSeek-R1、QwQ 等）")
+    thinking_budget_tokens: Optional[int] = Field(default=None, ge=1024, le=131072, description="深度思考 token 预算（仅部分模型支持）")
 
 
 class VariableDefinition(BaseModel):
@@ -612,6 +614,7 @@ class AppChatRequest(BaseModel):
     user_id: Optional[str] = Field(default=None, description="用户ID（用于会话管理）")
     variables: Optional[Dict[str, Any]] = Field(default=None, description="自定义变量参数值")
     stream: bool = Field(default=False, description="是否流式返回")
+    thinking: bool = Field(default=False, description="是否启用深度思考（需Agent配置支持）")
     files: List[FileInput] = Field(default_factory=list, description="附件列表（支持多文件）")
 
 

api/app/schemas/conversation_schema.py

@@ -31,6 +31,7 @@ class ChatRequest(BaseModel):
     stream: bool = Field(default=False, description="是否流式返回")
     web_search: bool = Field(default=False, description="是否启用网络搜索")
     memory: bool = Field(default=True, description="是否启用记忆功能")
+    thinking: bool = Field(default=False, description="是否启用深度思考（需Agent配置支持）")
     files: Optional[List[FileInput]] = Field(default=None, description="附件列表（支持多文件）")
 
 

api/app/services/app_chat_service.py

@@ -117,7 +117,9 @@ class AppChatService:
             max_tokens=model_parameters.get("max_tokens", 2000),
             system_prompt=system_prompt,
             tools=tools,
-
+            deep_thinking=model_parameters.get("deep_thinking", False),
+            thinking_budget_tokens=model_parameters.get("thinking_budget_tokens"),
+            capability=api_key_obj.capability or [],
         )
 
         model_info = ModelInfo(
@@ -205,7 +207,8 @@ class AppChatService:
             "model": api_key_obj.model_name,
             "usage": result.get("usage", {"prompt_tokens": 0, "completion_tokens": 0, "total_tokens": 0}),
             "audio_url": None,
-            "citations": filtered_citations
+            "citations": filtered_citations,
+            "reasoning_content": result.get("reasoning_content")
         }
         if files:
             for f in files:
@@ -258,6 +261,7 @@ class AppChatService:
             "conversation_id": conversation_id,
             "message_id": str(message_id),
             "message": result["content"],
+            "reasoning_content": result.get("reasoning_content"),
             "usage": result.get("usage", {
                 "prompt_tokens": 0,
                 "completion_tokens": 0,
@@ -354,7 +358,10 @@ class AppChatService:
                 max_tokens=model_parameters.get("max_tokens", 2000),
                 system_prompt=system_prompt,
                 tools=tools,
-                streaming=True
+                streaming=True,
+                deep_thinking=model_parameters.get("deep_thinking", False),
+                thinking_budget_tokens=model_parameters.get("thinking_budget_tokens"),
+                capability=api_key_obj.capability or [],
             )
 
             model_info = ModelInfo(
@@ -403,6 +410,7 @@ class AppChatService:
 
             # 流式调用 Agent（支持多模态），同时并行启动 TTS
             full_content = ""
+            full_reasoning = ""
             total_tokens = 0
 
             text_queue: asyncio.Queue = asyncio.Queue()
@@ -426,6 +434,9 @@ class AppChatService:
             ):
                 if isinstance(chunk, int):
                     total_tokens = chunk
+                elif isinstance(chunk, dict) and chunk.get("type") == "reasoning":
+                    full_reasoning += chunk['content']
+                    yield f"event: reasoning\ndata: {json.dumps({'content': chunk['content']}, ensure_ascii=False)}\n\n"
                 else:
                     full_content += chunk
                     yield f"event: message\ndata: {json.dumps({'content': chunk}, ensure_ascii=False)}\n\n"
@@ -472,7 +483,8 @@ class AppChatService:
                 "model": api_key_obj.model_name,
                 "usage": {"prompt_tokens": 0, "completion_tokens": 0, "total_tokens": total_tokens},
                 "audio_url": None,
-                "citations": filtered_citations
+                "citations": filtered_citations,
+                "reasoning_content": full_reasoning or None
             }
 
             if files:
@@ -652,13 +664,13 @@ class AppChatService:
                     storage_type=storage_type,
                     user_rag_memory_id=user_rag_memory_id
             ):
-                if "sub_usage" in event:
+                # 拦截 sub_usage 事件，累加 token
+                if "event: sub_usage" in event:
                     if "data:" in event:
                         try:
                             data_line = event.split("data: ", 1)[1].strip()
                             data = json.loads(data_line)
-                            if "total_tokens" in data:
+                            total_tokens += data.get("total_tokens", 0)
-                                total_tokens += data["total_tokens"]
                         except:
                             pass
                 else:

api/app/services/conversation_service.py

@@ -534,6 +534,7 @@ class ConversationService:
         api_key = api_config.api_key
         api_base = api_config.api_base
         is_omni = api_config.is_omni
+        capability = api_config.capability
         model_type = config.type
 
         llm = RedBearLLM(
@@ -542,7 +543,8 @@ class ConversationService:
                 provider=provider,
                 api_key=api_key,
                 base_url=api_base,
-                is_omni=is_omni
+                is_omni=is_omni,
+                support_thinking="thinking" in (capability or []),
             ),
             type=ModelType(model_type)
         )

api/app/services/draft_run_service.py

@@ -595,6 +595,9 @@ class AgentRunService:
                 max_tokens=effective_params.get("max_tokens", 2000),
                 system_prompt=system_prompt,
                 tools=tools,
+                deep_thinking=effective_params.get("deep_thinking", False),
+                thinking_budget_tokens=effective_params.get("thinking_budget_tokens"),
+                capability=api_key_config.get("capability", []),
             )
 
             # 5. 处理会话ID（创建或验证），新会话时写入开场白
@@ -689,7 +692,8 @@ class AgentRunService:
                             "prompt_tokens": 0,
                             "completion_tokens": 0,
                             "total_tokens": 0
-                        })
+                        }),
+                        "reasoning_content": result.get("reasoning_content")
                     },
                     files=files,
                     processed_files=processed_files,
@@ -701,6 +705,7 @@ class AgentRunService:
 
             response = {
                 "message": result["content"],
+                "reasoning_content": result.get("reasoning_content"),
                 "conversation_id": conversation_id,
                 "usage": result.get("usage", {
                     "prompt_tokens": 0,
@@ -838,7 +843,10 @@ class AgentRunService:
                 max_tokens=effective_params.get("max_tokens", 2000),
                 system_prompt=system_prompt,
                 tools=tools,
-                streaming=True
+                streaming=True,
+                deep_thinking=effective_params.get("deep_thinking", False),
+                thinking_budget_tokens=effective_params.get("thinking_budget_tokens"),
+                capability=api_key_config.get("capability", []),
             )
 
             # 5. 处理会话ID（创建或验证），新会话时写入开场白
@@ -898,6 +906,7 @@ class AgentRunService:
 
             # 9. 流式调用 Agent（支持多模态），同时并行启动 TTS
             full_content = ""
+            full_reasoning = ""
             total_tokens = 0
 
             # 启动流式 TTS（文本边输出边合成）
@@ -916,6 +925,9 @@ class AgentRunService:
             ):
                 if isinstance(chunk, int):
                     total_tokens = chunk
+                elif isinstance(chunk, dict) and chunk.get("type") == "reasoning":
+                    full_reasoning += chunk["content"]
+                    yield self._format_sse_event("reasoning", {"content": chunk["content"]})
                 else:
                     full_content += chunk
                     yield self._format_sse_event("message", {"content": chunk})
@@ -944,7 +956,8 @@ class AgentRunService:
                     app_id=agent_config.app_id,
                     user_id=user_id,
                     meta_data={
-                        "usage": {"prompt_tokens": 0, "completion_tokens": 0, "total_tokens": total_tokens}
+                        "usage": {"prompt_tokens": 0, "completion_tokens": 0, "total_tokens": total_tokens},
+                        "reasoning_content": full_reasoning or None
                     },
                     files=files,
                     processed_files=processed_files,
@@ -1665,7 +1678,7 @@ class AgentRunService:
             """从 text_queue 取文本按句子切分后喂给 synthesizer"""
             import re
             buf = ""
-            sentence_end = re.compile(r'[\u3002\uff01\uff1f\.!?\n]')
+            sentence_end = re.compile(r'[\u3002\uff01\uff1f.!?\n]')
             while True:
                 chunk = await text_queue.get()
                 if chunk is None:
@@ -1894,6 +1907,7 @@ class AgentRunService:
                     "conversation_id": result['conversation_id'],
                     "parameters_used": model_info["parameters"],
                     "message": result.get("message"),
+                    "reasoning_content": result.get("reasoning_content"),
                     "usage": usage,
                     "elapsed_time": elapsed,
                     "tokens_per_second": (
@@ -2012,7 +2026,7 @@ class AgentRunService:
         # 需要从 ModelApiKey 获取实际的模型名称，或者在 ModelConfig 中添加 model 字段
         return None
 
-    def _with_parameters(self, agent_config: AgentConfig, parameters: Dict[str, Any]) -> AgentConfig:
+    def _with_parameters(self, agent_config: AgentConfig, parameters: Dict[str, Any]) -> tuple[AgentConfig, Any]:
         """创建一个带有覆盖参数的 agent_config（浅拷贝，只修改 model_parameters）
 
         Args:
@@ -2110,6 +2124,7 @@ class AgentRunService:
 
                 start_time = time.time()
                 full_content = ""
+                full_reasoning = ""
                 returned_conversation_id = model_conversation_id
                 audio_url = None
                 audio_status = None
@@ -2168,6 +2183,18 @@ class AgentRunService:
                                     "content": chunk
                                 }))
 
+                            # 转发深度思考事件（带模型标识）
+                            if event_type == "reasoning" and event_data:
+                                reasoning_chunk = event_data.get("content", "")
+                                full_reasoning += reasoning_chunk
+                                await event_queue.put(self._format_sse_event("model_reasoning", {
+                                    "model_index": idx,
+                                    "model_config_id": model_config_id,
+                                    "label": model_label,
+                                    "conversation_id": returned_conversation_id,
+                                    "content": event_data.get("content", "")
+                                }))
+
                             # 从 end 事件中提取 features 输出字段
                             if event_type == "end" and event_data:
                                 audio_url = event_data.get("audio_url")
@@ -2199,6 +2226,7 @@ class AgentRunService:
                     "conversation_id": returned_conversation_id,
                     "parameters_used": model_info["parameters"],
                     "message": full_content,
+                    "reasoning_content": full_reasoning or None,
                     "elapsed_time": elapsed,
                     "audio_url": audio_url,
                     "audio_status": audio_status,
@@ -2351,6 +2379,7 @@ class AgentRunService:
                 "label": r["label"],
                 "conversation_id": r.get("conversation_id"),
                 "message": r.get("message"),
+                "reasoning_content": r.get("reasoning_content"),
                 "elapsed_time": r.get("elapsed_time", 0),
                 "audio_url": r.get("audio_url"),
                 "audio_status": r.get("audio_status"),

api/app/services/llm_router.py

@@ -415,6 +415,7 @@ class LLMRouter:
                 api_key=api_key_config.api_key,
                 base_url=api_key_config.api_base,
                 is_omni=api_key_config.is_omni,
+                support_thinking="thinking" in (api_key_config.capability or []),
                 temperature=0.3,
                 max_tokens=500
             )

api/app/services/master_agent_router.py

@@ -393,6 +393,7 @@ class MasterAgentRouter:
                 api_key=api_key_config.api_key,
                 base_url=api_key_config.api_base,
                 is_omni=api_key_config.is_omni,
+                support_thinking="thinking" in (api_key_config.capability or []),
                 extra_params = extra_params
             )
 
@@ -403,6 +404,17 @@ class MasterAgentRouter:
             response = await llm.ainvoke(prompt)
             ModelApiKeyService.record_api_key_usage(self.db, api_key_config.id)
 
+            # 提取 token 消耗
+            self._last_routing_tokens = 0
+            if hasattr(response, 'usage_metadata') and response.usage_metadata:
+                um = response.usage_metadata
+                self._last_routing_tokens = um.get("total_tokens", 0) if isinstance(um, dict) else getattr(um, "total_tokens", 0)
+            elif hasattr(response, 'response_metadata') and response.response_metadata:
+                token_usage = response.response_metadata.get("token_usage") or response.response_metadata.get("usage", {})
+                if isinstance(token_usage, dict):
+                    self._last_routing_tokens = token_usage.get("total_tokens", 0)
+            logger.info(f"Master Agent 路由 token 消耗: {self._last_routing_tokens}")
+
             # 提取响应内容
             if hasattr(response, 'content'):
                 return response.content

api/app/services/memory_perceptual_service.py

@@ -232,7 +232,8 @@ class MemoryPerceptualService:
                     provider=model_config.provider,
                     api_key=model_config.api_key,
                     base_url=model_config.api_base,
-                    is_omni=model_config.is_omni
+                    is_omni=model_config.is_omni,
+                    support_thinking="thinking" in (model_config.capability or []),
                 )
             )
         return llm, model_config

api/app/services/model_parameter_merger.py

@@ -45,12 +45,20 @@ class ModelParameterMerger:
             "frequency_penalty": 0.0,
             "presence_penalty": 0.0,
             "n": 1,
-            "stop": None
+            "stop": None,
+            "deep_thinking": False,
+            "thinking_budget_tokens": None
         }
         
         # 合并参数：默认值 -> 模型配置 -> Agent 配置
         merged = default_params.copy()
         
+        # Pydantic 对象转为 dict
+        if model_config_params and hasattr(model_config_params, 'model_dump'):
+            model_config_params = model_config_params.model_dump()
+        if agent_config_params and hasattr(agent_config_params, 'model_dump'):
+            agent_config_params = agent_config_params.model_dump()
+        
         # 应用模型配置参数
         if model_config_params:
             for key in default_params:

api/app/services/model_service.py

@@ -85,15 +85,16 @@ class ModelConfigService:
 
     @staticmethod
     async def validate_model_config(
-            db: Session,
+        db: Session,
-            *,
+        *,
-            model_name: str,
+        model_name: str,
-            provider: str,
+        provider: str,
-            api_key: str,
+        api_key: str,
-            api_base: Optional[str] = None,
+        api_base: Optional[str] = None,
-            model_type: str = "llm",
+        model_type: str = "llm",
-            test_message: str = "Hello",
+        test_message: str = "Hello",
-            is_omni: bool = False
+        is_omni: bool = False,
+        capability: Optional[list] = None
     ) -> Dict[str, Any]:
         """验证模型配置是否有效
 
@@ -124,6 +125,7 @@ class ModelConfigService:
                 api_key=api_key,
                 base_url=api_base,
                 is_omni=is_omni,
+                support_thinking="thinking" in (capability or []),
                 temperature=0.7,
                 max_tokens=100
             )
@@ -320,7 +322,8 @@ class ModelConfigService:
                     api_base=api_key_data.api_base,
                     model_type=model_data.type,
                     test_message="Hello",
-                    is_omni=model_data.is_omni
+                    is_omni=model_data.is_omni,
+                    capability=model_data.capability
                 )
                 if not validation_result["valid"]:
                     raise BusinessException(
@@ -590,7 +593,8 @@ class ModelApiKeyService:
                 api_base=data.api_base,
                 model_type=model_config.type,
                 test_message="Hello",
-                is_omni=data.is_omni
+                is_omni=data.is_omni,
+                capability=model_config.capability
             )
             if not validation_result["valid"]:
                 # 记录验证失败的模型，但不抛出异常
@@ -675,7 +679,8 @@ class ModelApiKeyService:
                     api_base=api_key_data.api_base,
                     model_type=model_config.type,
                     test_message="Hello",
-                    is_omni=api_key_data.is_omni
+                    is_omni=api_key_data.is_omni,
+                    capability=model_config.capability
                 )
                 if not validation_result["valid"]:
                     raise BusinessException(
@@ -707,7 +712,8 @@ class ModelApiKeyService:
                 api_base=api_key_data.api_base or existing_api_key.api_base,
                 model_type=model_config.type,
                 test_message="Hello",
-                is_omni=model_config.is_omni
+                is_omni=model_config.is_omni,
+                capability=model_config.capability
             )
             if not validation_result["valid"]:
                 raise BusinessException(

api/app/services/multi_agent_orchestrator.py

@@ -287,6 +287,11 @@ class MultiAgentOrchestrator:
             sub_conversation_id = None
             total_tokens = 0
             
+            # 累加 Master Agent 路由决策消耗的 token
+            total_tokens += task_analysis.get("routing_tokens", 0)
+            # 累加 Master Agent 整合消耗的 token
+            total_tokens += getattr(self, '_last_merge_tokens', 0)
+
             if isinstance(results, dict):
                 sub_conversation_id = results.get("conversation_id") or results.get("result", {}).get("conversation_id")
                 # 提取 token 信息
@@ -358,12 +363,16 @@ class MultiAgentOrchestrator:
             variables=variables
         )
 
+        # 获取路由决策消耗的 token
+        routing_tokens = getattr(self.router, '_last_routing_tokens', 0)
+
         logger.info(
             "Master Agent 分析完成",
             extra={
                 "selected_agent": routing_decision.get("selected_agent_id"),
                 "confidence": routing_decision.get("confidence"),
-                "strategy": routing_decision.get("strategy")
+                "strategy": routing_decision.get("strategy"),
+                "routing_tokens": routing_tokens
             }
         )
 
@@ -372,7 +381,8 @@ class MultiAgentOrchestrator:
             "variables": variables or {},
             "sub_agents": self.config.sub_agents,
             "initial_context": variables or {},
-            "routing_decision": routing_decision
+            "routing_decision": routing_decision,
+            "routing_tokens": routing_tokens
         }
 
     async def _execute_sequential(
@@ -1032,6 +1042,11 @@ class MultiAgentOrchestrator:
 
         # 5. 流式执行子 Agent
         sub_conversation_id = None
+        # Master Agent 路由决策消耗的 token，通过 sub_usage 事件发送给上层
+        routing_tokens = task_analysis.get("routing_tokens", 0)
+        if routing_tokens > 0:
+            yield self._format_sse_event("sub_usage", {"total_tokens": routing_tokens})
+
         async for event in self._execute_sub_agent_stream(
             agent_data["config"],
             message,
@@ -1054,6 +1069,7 @@ class MultiAgentOrchestrator:
                 except:
                     pass
 
+            # 直接透传所有事件（包括 sub_usage），累加统一由上层处理
             yield event
 
         # 6. 如果有会话 ID，发送一个包含它的事件
@@ -2600,6 +2616,7 @@ class MultiAgentOrchestrator:
                 api_key=api_key_config.api_key,
                 base_url=api_key_config.api_base,
                 is_omni=api_key_config.is_omni,
+                support_thinking="thinking" in (api_key_config.capability or []),
                 temperature=0.7,  # 整合任务使用中等温度
                 max_tokens=2000
             )
@@ -2612,6 +2629,17 @@ class MultiAgentOrchestrator:
 
             ModelApiKeyService.record_api_key_usage(self.db, api_key_config.id)
 
+            # 提取整合消耗的 token
+            merge_tokens = 0
+            if hasattr(response, 'usage_metadata') and response.usage_metadata:
+                um = response.usage_metadata
+                merge_tokens = um.get("total_tokens", 0) if isinstance(um, dict) else getattr(um, "total_tokens", 0)
+            elif hasattr(response, 'response_metadata') and response.response_metadata:
+                token_usage = response.response_metadata.get("token_usage") or response.response_metadata.get("usage", {})
+                if isinstance(token_usage, dict):
+                    merge_tokens = token_usage.get("total_tokens", 0)
+            self._last_merge_tokens = merge_tokens
+
             # 提取响应内容
             if hasattr(response, 'content'):
                 merged_response = response.content
@@ -2621,7 +2649,8 @@ class MultiAgentOrchestrator:
             logger.info(
                 "Master Agent 整合完成",
                 extra={
-                    "merged_length": len(merged_response)
+                    "merged_length": len(merged_response),
+                    "merge_tokens": merge_tokens
                 }
             )
 
@@ -2766,6 +2795,7 @@ class MultiAgentOrchestrator:
                 api_key=api_key_config.api_key,
                 base_url=api_key_config.api_base,
                 is_omni=api_key_config.is_omni,
+                support_thinking="thinking" in (api_key_config.capability or []),
                 temperature=0.7,
                 max_tokens=2000,
                 extra_params={"streaming": True}  # 启用流式输出

api/app/services/prompt_optimizer_service.py

@@ -185,7 +185,8 @@ class PromptOptimizerService:
             provider=api_config.provider,
             api_key=api_config.api_key,
             base_url=api_config.api_base,
-            is_omni=api_config.is_omni
+            is_omni=api_config.is_omni,
+            support_thinking="thinking" in (api_config.capability or []),
         ), type=ModelType(model_config.type))
         try:
             prompt_path = os.path.join(os.path.dirname(os.path.abspath(__file__)), 'prompt')

api/app/services/shared_chat_service.py

@@ -248,7 +248,9 @@ class SharedChatService:
             max_tokens=model_parameters.get("max_tokens", 2000),
             system_prompt=system_prompt,
             tools=tools,
-
+            deep_thinking=model_parameters.get("deep_thinking", False),
+            thinking_budget_tokens=model_parameters.get("thinking_budget_tokens"),
+            capability=api_key_obj.capability or [],
         )
 
         # 加载历史消息
@@ -450,7 +452,10 @@ class SharedChatService:
                 max_tokens=model_parameters.get("max_tokens", 2000),
                 system_prompt=system_prompt,
                 tools=tools,
-                streaming=True
+                streaming=True,
+                deep_thinking=model_parameters.get("deep_thinking", False),
+                thinking_budget_tokens=model_parameters.get("thinking_budget_tokens"),
+                capability=api_key_obj.capability or [],
             )
 
             # 加载历史消息
@@ -479,6 +484,8 @@ class SharedChatService:
             ):
                 if isinstance(chunk, int):
                     total_tokens = chunk
+                elif isinstance(chunk, dict) and chunk.get("type") == "reasoning":
+                    yield f"event: reasoning\ndata: {json.dumps({'content': chunk['content']}, ensure_ascii=False)}\n\n"
                 else:
                     full_content += chunk
                     # 发送消息块事件