feat: enhance openai-compatible provider support and patch responses API instructions handling

- Add compatibility patch for langchain-openai responses API to ensure system messages are extracted as top-level instructions, addressing Codex endpoint requirements.
- Update provider list: add Alibaba, Volcengine, and Tencent TokenHub; adjust SiliconFlow and MiniMax endpoints; refine provider ordering and model list strategies.
- Extend models.dev-only listing logic for providers lacking stable models.list endpoints.
- Increase models.dev cache TTL for improved efficiency.
- Add tests for openai responses API and streaming compatibility patches.

app/agent/llm/helper.py

@@ -182,6 +182,77 @@ def _patch_deepseek_reasoning_content_support():
     logger.debug("已修补 langchain-deepseek thinking tool-call 的 reasoning_content 回传兼容性")
 
 
+def _patch_openai_responses_instructions_support():
+    """
+    修补 langchain-openai 在使用 use_responses_api=True 时，
+    提取 system 消息为顶层 instructions 字段。
+    由于 Codex 等模型 (Responses API) 强依赖 instructions 字段，
+    如果没有该字段会报 400 "Instructions are required"。
+    """
+    try:
+        from langchain_openai import ChatOpenAI
+    except Exception as err:
+        logger.debug(f"跳过 langchain-openai instructions 修补：{err}")
+        return
+
+    if getattr(ChatOpenAI, "_moviepilot_responses_instructions_patched", False):
+        return
+
+    original_get_request_payload = getattr(ChatOpenAI, "_get_request_payload", None)
+    if not callable(original_get_request_payload):
+        logger.warning("langchain-openai 缺少 _get_request_payload，无法修补 instructions")
+        return
+
+    @wraps(original_get_request_payload)
+    def _patched_get_request_payload(self, input_, *, stop=None, **kwargs):
+        payload = original_get_request_payload(self, input_, stop=stop, **kwargs)
+
+        base_url = str(getattr(self, "openai_api_base", "") or "").lower()
+
+        # 处理 GitHub Copilot 端点兼容性
+        if "githubcopilot.com" in base_url:
+            payload.pop("stream_options", None)
+            payload.pop("metadata", None)
+
+        # 处理 ChatGPT 官方 Responses API (Codex) 端点兼容性
+        is_codex = "chatgpt.com/backend-api/codex" in base_url
+        
+        if is_codex and (getattr(self, "use_responses_api", False) or "input" in payload):
+            instructions = payload.get("instructions", "")
+            inputs = payload.get("input", [])
+            new_inputs = []
+
+            for msg in inputs:
+                if isinstance(msg, dict) and msg.get("role") == "system":
+                    content = msg.get("content")
+                    if isinstance(content, str) and content.strip():
+                        if instructions:
+                            instructions += "\n\n" + content
+                        else:
+                            instructions = content
+                else:
+                    new_inputs.append(msg)
+
+            payload["input"] = new_inputs
+            payload["instructions"] = instructions or "You are a helpful assistant."
+            payload["store"] = False
+            
+            # Codex 端点不支持的部分常见补全参数，统一清理避免 400 报错
+            unsupported_keys = [
+                "presence_penalty", "frequency_penalty", "top_p", "n", "user", 
+                "stop", "metadata", "logit_bias", "logprobs", "top_logprobs",
+                "stream_options", "temperature"
+            ]
+            for key in unsupported_keys:
+                payload.pop(key, None)
+
+        return payload
+
+    ChatOpenAI._get_request_payload = _patched_get_request_payload
+    ChatOpenAI._moviepilot_responses_instructions_patched = True
+    logger.debug("已修补 langchain-openai responses API 的 instructions 兼容性")
+
+
 class LLMHelper:
     """LLM模型相关辅助功能"""
 
@@ -559,6 +630,12 @@ class LLMHelper:
         else:
             from langchain_openai import ChatOpenAI
 
+            _patch_openai_responses_instructions_support()
+            
+            # ChatGPT Codex 端点强制要求 stream: True
+            if runtime.get("use_responses_api") and "chatgpt.com/backend-api/codex" in str(runtime.get("base_url") or ""):
+                streaming = True
+
             model = ChatOpenAI(
                 model=model_name,
                 api_key=runtime["api_key"],

app/agent/llm/provider.py

@@ -89,7 +89,7 @@ class LLMProviderManager(metaclass=Singleton):
     """统一维护 provider 目录、models.dev 缓存和 OAuth 状态。"""
 
     _MODELS_DEV_URL = "https://models.dev/api.json"
-    _MODELS_DEV_CACHE_TTL = 5 * 60
+    _MODELS_DEV_CACHE_TTL = 12 * 60 * 60
     _CHATGPT_CLIENT_ID = "app_EMoamEEZ73f0CkXaXp7hrann"
     _CHATGPT_ISSUER = "https://auth.openai.com"
     _CHATGPT_CODEX_BASE_URL = "https://chatgpt.com/backend-api/codex"
@@ -201,26 +201,46 @@ class LLMProviderManager(metaclass=Singleton):
                 description="通过 GitHub Copilot 订阅接入。",
                 sort_order=50,
             ),
-            ProviderSpec(
-                id="nvidia",
-                name="Nvidia",
-                runtime="openai_compatible",
-                models_dev_provider_id="nvidia",
-                default_base_url="https://integrate.api.nvidia.com/v1",
-                api_key_hint="填写 Nvidia API Key。",
-                description="Nvidia 集成推理平台。",
-                sort_order=60,
-            ),
             ProviderSpec(
                 id="siliconflow",
                 name="硅基流动",
                 runtime="openai_compatible",
                 models_dev_provider_id="siliconflow",
-                default_base_url="https://api.siliconflow.com/v1",
+                default_base_url="https://api.siliconflow.cn/v1",
                 api_key_hint="填写硅基流动 API Key。",
                 description="SiliconFlow 官方兼容端点。",
+                sort_order=60,
+            ),
+            ProviderSpec(
+                id="alibaba",
+                name="阿里云百炼",
+                runtime="openai_compatible",
+                models_dev_provider_id="alibaba",
+                default_base_url="https://dashscope.aliyuncs.com/compatible-mode/v1",
+                api_key_hint="填写 DashScope / Alibaba API Key。",
+                description="阿里云百炼兼容端点。",
                 sort_order=70,
             ),
+            ProviderSpec(
+                id="volcengine",
+                name="火山方舟",
+                runtime="openai_compatible",
+                default_base_url="https://ark.cn-beijing.volces.com/api/v3",
+                api_key_hint="填写火山方舟 API Key。",
+                description="字节跳动火山引擎兼容端点。",
+                sort_order=80,
+            ),
+            ProviderSpec(
+                id="tencent",
+                name="Tencent",
+                runtime="openai_compatible",
+                models_dev_provider_id="tencent-tokenhub",
+                default_base_url="https://tokenhub.tencentmaas.com/v1",
+                api_key_hint="填写 Tencent API Key。",
+                model_list_strategy="models_dev_only",
+                description="腾讯兼容端点。",
+                sort_order=90,
+            ),
             ProviderSpec(
                 id="ollama-cloud",
                 name="Ollama Cloud",
@@ -229,28 +249,28 @@ class LLMProviderManager(metaclass=Singleton):
                 default_base_url="https://ollama.com/v1",
                 api_key_hint="填写 Ollama Cloud API Key。",
                 description="Ollama Cloud 云端模型服务。",
-                sort_order=80,
+                sort_order=100,
             ),
             ProviderSpec(
-                id="alibaba",
+                id="nvidia",
-                name="Alibaba",
+                name="Nvidia",
                 runtime="openai_compatible",
-                models_dev_provider_id="alibaba",
+                models_dev_provider_id="nvidia",
-                default_base_url="https://dashscope-intl.aliyuncs.com/compatible-mode/v1",
+                default_base_url="https://integrate.api.nvidia.com/v1",
-                api_key_hint="填写 DashScope / Alibaba API Key。",
+                api_key_hint="填写 Nvidia API Key。",
-                description="阿里云百炼兼容端点。",
+                description="Nvidia 集成推理平台。",
-                sort_order=90,
+                sort_order=110,
             ),
             ProviderSpec(
                 id="minimax",
                 name="MiniMax",
                 runtime="anthropic_compatible",
                 models_dev_provider_id="minimax",
-                default_base_url="https://api.minimax.io/anthropic/v1",
+                default_base_url="https://api.minimaxi.com/anthropic/v1",
                 api_key_hint="填写 MiniMax API Key。",
                 model_list_strategy="anthropic_compatible",
                 description="MiniMax Anthropic-compatible 端点。",
-                sort_order=100,
+                sort_order=120,
             ),
             ProviderSpec(
                 id="xiaomi",
@@ -260,17 +280,7 @@ class LLMProviderManager(metaclass=Singleton):
                 default_base_url="https://api.xiaomimimo.com/v1",
                 api_key_hint="填写 Xiaomi API Key。",
                 description="小米 Mimo 兼容端点。",
-                sort_order=110,
+                sort_order=130,
-            ),
-            ProviderSpec(
-                id="tencent",
-                name="Tencent",
-                runtime="openai_compatible",
-                models_dev_provider_id="tencent",
-                default_base_url="https://api.lkeap.cloud.tencent.com/coding/v3",
-                api_key_hint="填写 Tencent API Key。",
-                description="腾讯兼容端点。",
-                sort_order=120,
             ),
             ProviderSpec(
                 id="openai",
@@ -640,12 +650,13 @@ class LLMProviderManager(metaclass=Singleton):
             )
         return sorted(results, key=lambda item: item["name"].lower())
 
-    async def _list_models_from_anthropic_compatible(
+    async def _list_models_from_models_dev_only(
             self,
             provider_id: str,
+            transport: str = "openai",
     ) -> list[dict[str, Any]]:
         """
-        Anthropic-compatible 生态没有像 OpenAI 那样统一稳定的 models.list 行为，
+        某些 provider 没有统一稳定的 models.list 行为，
         因此优先读取 models.dev 目录；若未来 provider 暴露标准 models 接口，
         再平滑补充实时刷新即可。
         """
@@ -660,7 +671,7 @@ class LLMProviderManager(metaclass=Singleton):
                     model_id=model_id,
                     display_name=metadata.get("name") or model_id,
                     metadata=metadata,
-                    transport="anthropic",
+                    transport=transport,
                     source="models.dev",
                 )
             )
@@ -832,8 +843,15 @@ class LLMProviderManager(metaclass=Singleton):
             )
 
         if spec.model_list_strategy == "anthropic_compatible":
-            return await self._list_models_from_anthropic_compatible(
+            return await self._list_models_from_models_dev_only(
                 provider_id=provider_id,
+                transport="anthropic",
+            )
+            
+        if spec.model_list_strategy == "models_dev_only":
+            return await self._list_models_from_models_dev_only(
+                provider_id=provider_id,
+                transport="openai",
             )
 
         # openai-compatible / deepseek 默认走官方 models 端点。
@@ -1348,7 +1366,8 @@ class LLMProviderManager(metaclass=Singleton):
             auth = None
             try:
                 auth = await self._resolve_chatgpt_oauth()
-            except Exception:
+            except Exception as err:
+                print(err)
                 pass
 
             if auth:

test_openai_copilot_patch.py

@@ -0,0 +1,9 @@
+from app.agent.llm.helper import _patch_openai_responses_instructions_support
+from langchain_openai import ChatOpenAI
+from langchain_core.messages import SystemMessage, HumanMessage
+import json
+
+_patch_openai_responses_instructions_support()
+model = ChatOpenAI(model="gpt-4o", openai_api_key="sk-123", base_url="https://api.githubcopilot.com", stream_usage=True)
+payload = model._get_request_payload([SystemMessage(content="Hello system"), HumanMessage(content="Hello user")])
+print(json.dumps(payload, indent=2))

test_openai_responses_patch3.py

@@ -0,0 +1,9 @@
+from app.agent.llm.helper import _patch_openai_responses_instructions_support
+from langchain_openai import ChatOpenAI
+from langchain_core.messages import SystemMessage, HumanMessage
+import json
+
+_patch_openai_responses_instructions_support()
+model = ChatOpenAI(model="gpt-4o", openai_api_key="sk-123", use_responses_api=True, temperature=0.7)
+payload = model._get_request_payload([SystemMessage(content="Hello system"), HumanMessage(content="Hello user")])
+print(json.dumps(payload, indent=2))

test_openai_stream_patch.py

@@ -0,0 +1,14 @@
+import asyncio
+from app.agent.llm.helper import LLMHelper
+from app.core.config import settings
+import json
+
+async def run():
+    llm = await LLMHelper.get_llm(
+        streaming=False,
+        provider="chatgpt",
+        model="gpt-5.1-codex",
+    )
+    print("streaming:", llm.streaming)
+    
+asyncio.run(run())