feat(litellm): 添加 reasoning_content 支持并强制禁用思考模式

- 在 LiteLLMProvider 中添加 "reasoning_content" 到允许的消息键集合中
- 修改 _apply_thinking_mode 方法以强制禁用思考模式，不再基于模型名称判断
- 总是设置 enable_thinking 为 False 并添加 thinking.type: disabled 配置
- 更新相关测试用例验证新的思考模式行为

fix(web): 修复非运行状态下的直接处理逻辑

- 创建 _run_web_direct 辅助函数来处理代理服务的直接提交/处理逻辑
- 当代理服务未运行时使用 process_direct 而不是 submit_direct
- 更新 REST 和 WebSocket 接口以使用新的处理逻辑
- 添加相应的单元测试验证非运行状态下使用直接处理

test(config): 添加代理配置重载功能的测试

- 添加 test_reload_agent_config_updates_booted_loop_config 测试函数
- 验证配置文件更新后代理循环能够正确加载新配置
- 测试模型、API 基础地址和 API 密钥的更新

chore(frontend): 默认禁用前端思考模式偏好

- 将前端思考模式存储的默认值从 true 改为 false
- 确保窗口未定义时返回 false 而不是 true
- 更新本地存储缺失时的默认行为为禁用思考模式

app-instance/backend/tests/unit/test_litellm_thinking_mode.py

@@ -45,10 +45,13 @@ def test_qwen_thinking_mode_is_sent_as_chat_template_kwargs(monkeypatch: pytest.
     )
 
     assert response.content == "可以"
-    assert captured["extra_body"] == {"chat_template_kwargs": {"enable_thinking": False}}
+    assert captured["extra_body"] == {
+        "chat_template_kwargs": {"enable_thinking": False},
+        "thinking": {"type": "disabled"},
+    }
 
 
-def test_non_qwen_thinking_mode_is_not_sent(monkeypatch: pytest.MonkeyPatch) -> None:
+def test_thinking_mode_disabled_is_sent_without_model_name_matching(monkeypatch: pytest.MonkeyPatch) -> None:
     captured: dict = {}
 
     class Message:
@@ -85,7 +88,72 @@ def test_non_qwen_thinking_mode_is_not_sent(monkeypatch: pytest.MonkeyPatch) ->
         )
     )
 
-    assert "extra_body" not in captured
+    assert captured["extra_body"] == {
+        "chat_template_kwargs": {"enable_thinking": False},
+        "thinking": {"type": "disabled"},
+    }
+
+
+def test_litellm_provider_preserves_reasoning_content_for_tool_round_trip() -> None:
+    messages = [
+        {
+            "role": "assistant",
+            "content": "",
+            "reasoning_content": "must be passed back",
+            "tool_calls": [
+                {
+                    "id": "call-1",
+                    "type": "function",
+                    "function": {"name": "lookup", "arguments": "{}"},
+                }
+            ],
+        }
+    ]
+
+    assert LiteLLMProvider._sanitize_messages(messages)[0]["reasoning_content"] == "must be passed back"
+
+
+def test_thinking_mode_is_forced_disabled_even_when_requested_enabled(monkeypatch: pytest.MonkeyPatch) -> None:
+    captured: dict = {}
+
+    class Message:
+        content = "ok"
+        reasoning_content = None
+        tool_calls = []
+
+    class Choice:
+        message = Message()
+        finish_reason = "stop"
+
+    class Response:
+        choices = [Choice()]
+        usage = None
+
+    async def fake_acompletion(**kwargs):
+        captured.update(kwargs)
+        return Response()
+
+    monkeypatch.setattr("beaver.engine.providers.litellm.acompletion", fake_acompletion)
+    monkeypatch.setattr("beaver.engine.providers.litellm.litellm", SimpleNamespace())
+
+    provider = LiteLLMProvider(
+        api_key="sk-test",
+        api_base="https://oai.example.com/v1",
+        default_model="gpt-4.1-mini",
+        provider_name="openai",
+    )
+    asyncio.run(
+        provider.chat(
+            [{"role": "user", "content": "reply ok"}],
+            model="gpt-4.1-mini",
+            thinking_enabled=True,
+        )
+    )
+
+    assert captured["extra_body"] == {
+        "chat_template_kwargs": {"enable_thinking": False},
+        "thinking": {"type": "disabled"},
+    }
 
 
 def test_litellm_provider_sanitizes_tool_call_arguments(monkeypatch: pytest.MonkeyPatch) -> None: