fix(agent): refresh LLM runtime config on each call

Read the latest LLM connection settings when building runtime clients so Web updates take effect immediately instead of reusing module-import defaults.

Closes #5757

app/agent/llm/helper.py

@@ -512,9 +512,9 @@ class LLMHelper:
             provider: str | None = None,
             model: str | None = None,
             thinking_level: str | None = None,
-            api_key: str | None = settings.LLM_API_KEY,
+            api_key: str | None = None,
-            base_url: str | None = settings.LLM_BASE_URL,
+            base_url: str | None = None,
-            base_url_preset: str | None = settings.LLM_BASE_URL_PRESET,
+            base_url_preset: str | None = None,
     ):
         """
         获取LLM实例
@@ -525,12 +525,18 @@ class LLMHelper:
             是否启用思考模式）。支持的级别包括 "off"（关闭）、"auto"（自动）、"minimal"、"low"、"medium"、"high"、"max"/"xhigh"（最大）。
             不同模型对思考模式的支持和表现不同，具体映射关系请
             参考代码实现。对于不支持思考模式的模型，该参数将被忽略。
-        :param api_key: API Key，默认为配置项LLM_API_KEY。对于某些提供商（如 DeepSeek），可能需要同时提供 base_url。
+        :param api_key: API Key。未显式传入时使用当前配置项 LLM_API_KEY。对于某些提供商（如 DeepSeek），可能需要同时提供 base_url。
-        :param base_url: API Base URL，默认为配置项LLM_BASE_URL。
+        :param base_url: API Base URL。未显式传入时使用当前配置项 LLM_BASE_URL。
+        :param base_url_preset: Base URL 预设。未显式传入时使用当前配置项 LLM_BASE_URL_PRESET。
         :return: LLM实例
         """
         provider_name = str(provider if provider is not None else settings.LLM_PROVIDER).lower()
         model_name = model if model is not None else settings.LLM_MODEL
+        api_key_value = api_key if api_key is not None else settings.LLM_API_KEY
+        base_url_value = base_url if base_url is not None else settings.LLM_BASE_URL
+        base_url_preset_value = (
+            base_url_preset if base_url_preset is not None else settings.LLM_BASE_URL_PRESET
+        )
         normalized_thinking_level = cls._resolve_thinking_level(
             thinking_level=thinking_level,
         )
@@ -542,17 +548,17 @@ class LLMHelper:
             runtime = await LLMProviderManager().resolve_runtime(
                 provider_id=provider_name,
                 model=model_name,
-                api_key=api_key,
+                api_key=api_key_value,
-                base_url=base_url,
+                base_url=base_url_value,
-                base_url_preset_id=base_url_preset,
+                base_url_preset_id=base_url_preset_value,
             )
         except Exception as err:
             logger.debug(f"LLM provider 目录不可用，回退到旧运行时逻辑: {err}")
             runtime = cls._build_legacy_runtime(
                 provider_name=provider_name,
                 model_name=model_name,
-                api_key=api_key,
+                api_key=api_key_value,
-                base_url=base_url,
+                base_url=base_url_value,
             )
         model_name = runtime.get("model_id") or model_name
         thinking_kwargs = cls._build_thinking_kwargs(

tests/test_llm_helper_testcall.py

@@ -244,6 +244,70 @@ class LlmHelperTestCallTest(unittest.TestCase):
         self.assertEqual(len(calls), 1)
         self.assertEqual(calls[0].get("reasoning_effort"), "none")
 
+    def test_get_llm_reads_latest_settings_when_runtime_args_omitted(self):
+        resolve_calls = []
+        llm_calls = []
+
+        class _FakeProviderManager:
+            async def resolve_runtime(self, **kwargs):
+                resolve_calls.append(kwargs)
+                return {
+                    "provider_id": kwargs["provider_id"],
+                    "runtime": "openai_compatible",
+                    "model_id": kwargs["model"],
+                    "api_key": kwargs["api_key"],
+                    "base_url": kwargs["base_url"],
+                    "default_headers": None,
+                    "use_responses_api": None,
+                    "model_record": None,
+                    "model_metadata": None,
+                }
+
+        class _FakeChatOpenAI:
+            def __init__(self, **kwargs):
+                llm_calls.append(kwargs)
+                self.model = kwargs["model"]
+                self.profile = None
+
+        provider_module = ModuleType("app.agent.llm.provider")
+        provider_module.LLMProviderManager = _FakeProviderManager
+        openai_module = ModuleType("langchain_openai")
+        openai_module.ChatOpenAI = _FakeChatOpenAI
+
+        with patch.object(llm_module.settings, "LLM_PROVIDER", "deepseek"), patch.object(
+            llm_module.settings, "LLM_MODEL", "deepseek-chat"
+        ), patch.object(llm_module.settings, "LLM_API_KEY", "updated-key"), patch.object(
+            llm_module.settings, "LLM_BASE_URL", "https://updated.example.com/v1"
+        ), patch.object(
+            llm_module.settings, "LLM_BASE_URL_PRESET", "updated-preset"
+        ), patch.dict(
+            sys.modules,
+            {
+                "app.agent.llm.provider": provider_module,
+                "langchain_openai": openai_module,
+            },
+        ):
+            asyncio.run(llm_module.LLMHelper.get_llm())
+
+        self.assertEqual(len(resolve_calls), 1)
+        self.assertEqual(
+            resolve_calls[0],
+            {
+                "provider_id": "deepseek",
+                "model": "deepseek-chat",
+                "api_key": "updated-key",
+                "base_url": "https://updated.example.com/v1",
+                "base_url_preset_id": "updated-preset",
+            },
+        )
+        self.assertEqual(len(llm_calls), 1)
+        self.assertEqual(llm_calls[0].get("model"), "deepseek-chat")
+        self.assertEqual(llm_calls[0].get("api_key"), "updated-key")
+        self.assertEqual(
+            llm_calls[0].get("base_url"),
+            "https://updated.example.com/v1",
+        )
+
     def test_get_llm_maps_unified_max_to_openai_xhigh(self):
         calls = []