fix(llm): unify JSON output handling across providers and fix tool+json_output compatibility

- Remove redundant `response_format` injection for VOLCANO provider since it's unsupported; rely on system prompt injection instead
- Extend system prompt JSON injection logic to cover VOLCANO and tool-enabled cases universally
- Simplify model parameter construction by removing redundant `params["model_kwargs"] = model_kwargs` assignments
- Refactor `CompatibleChatOpenAI._get_request_payload` to strip `response_format` when tools are present, avoiding strict validation errors in langchain_openai
- Fix timestamp calculation order in `datetime_tool.py` to avoid integer truncation before multiplication

api/app/core/models/base.py

@@ -101,12 +101,10 @@ class RedBearModelFactory:
                         extra_body["enable_thinking"] = True
                     if config.thinking_budget_tokens:
                         extra_body["thinking_budget"] = config.thinking_budget_tokens
-                params["extra_body"] = extra_body
             # JSON 输出模式
             if config.json_output:
                 model_kwargs = params.setdefault("model_kwargs", {})
                 model_kwargs["response_format"] = {"type": "json_object"}
-                params["model_kwargs"] = model_kwargs
             return params
 
         if provider in [ModelProvider.OPENAI, ModelProvider.XINFERENCE, ModelProvider.GPUSTACK, ModelProvider.OLLAMA, ModelProvider.VOLCANO]:
@@ -148,11 +146,12 @@ class RedBearModelFactory:
                             extra_body["enable_thinking"] = True
                         if config.thinking_budget_tokens:
                             extra_body["thinking_budget"] = config.thinking_budget_tokens
-                    params["extra_body"] = extra_body
             # JSON 输出模式
             if config.json_output:
-                params.setdefault("model_kwargs", {})
-                params["model_kwargs"]["response_format"] = {"type": "json_object"}
+                model_kwargs = params.setdefault("model_kwargs", {})
+                # VOLCANO 模型不支持 response_format，JSON 输出由 system prompt 注入实现
+                if provider != ModelProvider.VOLCANO:
+                    model_kwargs["response_format"] = {"type": "json_object"}
             return params
         elif provider == ModelProvider.DASHSCOPE:
             params = {
@@ -172,11 +171,9 @@ class RedBearModelFactory:
                         model_kwargs["incremental_output"] = True
                     if config.thinking_budget_tokens:
                         model_kwargs["thinking_budget"] = config.thinking_budget_tokens
-                params["model_kwargs"] = model_kwargs
             if config.json_output:
                 model_kwargs = params.setdefault("model_kwargs", {})
                 model_kwargs["response_format"] = {"type": "json_object"}
-                params["model_kwargs"] = model_kwargs
             return params
         elif provider == ModelProvider.BEDROCK:
             # Bedrock 使用 AWS 凭证
@@ -225,8 +222,8 @@ class RedBearModelFactory:
                 }
             # JSON 输出模式
             if config.json_output:
-                params.setdefault("model_kwargs", {})
-                params["model_kwargs"]["response_format"] = {"type": "json_object"}
+                model_kwargs = params.setdefault("model_kwargs", {})
+                model_kwargs["response_format"] = {"type": "json_object"}
             return params
         else:
             raise BusinessException(f"不支持的提供商: {provider}", code=BizCode.PROVIDER_NOT_SUPPORTED)
@@ -261,12 +258,13 @@ def get_provider_llm_class(config: RedBearModelConfig, type: ModelType = ModelTy
     if provider == ModelProvider.VOLCANO:
         return CompatibleChatOpenAI
     if provider in [ModelProvider.OPENAI, ModelProvider.XINFERENCE, ModelProvider.GPUSTACK]:
-        if type == ModelType.LLM:
-            return OpenAI
-        elif type == ModelType.CHAT:
-            return ChatOpenAI
-        else:
-            raise BusinessException(f"不支持的模型提供商及类型: {provider}-{type}", code=BizCode.PROVIDER_NOT_SUPPORTED)
+        return CompatibleChatOpenAI
+        # if type == ModelType.LLM:
+        #     return OpenAI
+        # elif type == ModelType.CHAT:
+        #     return CompatibleChatOpenAI
+        # else:
+        #     raise BusinessException(f"不支持的模型提供商及类型: {provider}-{type}", code=BizCode.PROVIDER_NOT_SUPPORTED)
     elif provider == ModelProvider.DASHSCOPE:
         return ChatTongyi
     elif provider == ModelProvider.OLLAMA:

api/app/core/models/compatible_chat.py

@@ -8,12 +8,33 @@ from __future__ import annotations
 
 from typing import Any, Optional, Union
 
+from langchain_core.messages import BaseMessage
 from langchain_core.outputs import ChatGenerationChunk, ChatResult
 from langchain_openai import ChatOpenAI
 
 
 class CompatibleChatOpenAI(ChatOpenAI):
-    """火山和千问的omni兼容模型，支持深度思考内容（reasoning_content）的流式和非流式透传。"""
+    """火山和千问的omni兼容模型，支持深度思考内容（reasoning_content）的流式和非流式透传。
+
+    同时修复 json_output + tools 同时使用时 langchain_openai 强制走 .parse()/.stream()
+    导致 strict 校验报错的问题：有工具时从 payload 中移除 response_format，
+    让父类走普通 .create()/.astream() 路径，JSON 输出由 system prompt 指令保证。
+    """
+
+    def _get_request_payload(
+        self,
+        input_: list[BaseMessage],
+        *,
+        stop: list[str] | None = None,
+        **kwargs: Any,
+    ) -> dict:
+        payload = super()._get_request_payload(input_, stop=stop, **kwargs)
+        # 有工具时 langchain_openai 检测到 response_format 会切换到 .parse()/.stream()
+        # 接口，OpenAI SDK 要求此时所有工具必须 strict=True，动态生成的工具不满足。
+        # 移除 response_format，让父类走普通路径，JSON 输出由 system prompt 指令保证。
+        if payload.get("tools") and "response_format" in payload:
+            payload.pop("response_format")
+        return payload
 
     def _create_chat_result(self, response: Union[dict, Any], generation_info: Optional[dict] = None) -> ChatResult:
         result = super()._create_chat_result(response, generation_info)