docs: add web search fix to v4.1.21 changelog

Author: lbjlaq

README.md

@@ -443,6 +443,13 @@ response = client.chat.completions.create(
         -   **[核心增强] Linux 版本检测与 HTTP 客户端鲁棒性**:
             -   **版本解析**: 增强了 Linux 平台的版本号提取逻辑 (`extract_semver`)，能从复杂的命令行输出中准确识别版本，提升了自动更新和环境检测的准确性。
             -   **客户端降级**: 为 HTTP 客户端构建过程增加了自动降级机制。当代理配置导致构建失败时，系统会自动回退到无代理模式或默认配置，防止因网络配置错误导致应用完全不可用。
+        -   **[核心修复] Cherry Studio 联网搜索空响应修复 (/v1/responses)**:
+            -   **SSE 事件补全**: 重写了 `create_codex_sse_stream`，补全了 OpenAI Responses API 规范要求的完整 SSE 事件生命周期（`response.output_item.added`、`content_part.added/done`、`output_item.done`、`response.completed`），解决了 Cherry Studio 因事件缺失导致无法组装响应内容的问题。
+            -   **联网搜索注入修复**: 过滤了 Cherry Studio 发送的 `builtin_web_search` 工具声明，防止其与 `inject_google_search_tool` 冲突，确保 Google Search 工具被正确注入。
+            -   **搜索引文回显**: 为 Codex 流式响应添加了 `groundingMetadata` 解析，支持在联网搜索结果中回显搜索查询和来源引文。
+        -   **[优化] Claude 协议联网与思考稳定性 (PR #2007)**:
+            -   **移除联网降级**: 移除了 Claude 协议中针对联网搜索的激进模型降级逻辑，避免不必要的模型回退。
+            -   **移除思考历史降级**: 移除了 `should_disable_thinking_due_to_history` 检查，不再因历史消息格式问题永久禁用思考模式，改为依赖 `thinking_recovery` 机制自动修复。
         -   **UI 优化 (Fix #2008)**: 改进了冷却时间的显示颜色 (使用蓝色)，提高了在小字体下的可读性。
     *   **v4.1.20 (2026-02-16)**:
         *   **[✨ 新春祝福] 祝大家马年一马当先，万事如意！Code 运昌隆，上线无 Bug！🧧**

README_EN.md

@@ -295,6 +295,13 @@ print(response.choices[0].message.content)
         -   **[Core Enhancement] Linux Version Detection & HTTP Client Robustness**:
             -   **Version Parsing**: Enhanced Linux version extraction logic (`extract_semver`) to accurately identify semantic versions from complex command outputs, improving auto-update and environment detection accuracy.
             -   **Client Fallback**: Added automatic fallback mechanisms for HTTP client construction. If proxy configuration fails, the system automatically reverts to no-proxy mode or default settings, preventing total application failure due to network misconfiguration.
+        -   **[Core Fix] Cherry Studio Web Search Empty Response (/v1/responses)**:
+            -   **SSE Event Completion**: Rewrote `create_codex_sse_stream` to emit the complete SSE event lifecycle required by the OpenAI Responses API specification (`response.output_item.added`, `content_part.added/done`, `output_item.done`, `response.completed`), resolving the issue where Cherry Studio failed to assemble response content due to missing events.
+            -   **Web Search Injection Fix**: Filtered out `builtin_web_search` tool declarations sent by Cherry Studio to prevent conflicts with `inject_google_search_tool`, ensuring the Google Search tool is correctly injected.
+            -   **Search Citation Echo**: Added `groundingMetadata` parsing to the Codex streaming response, enabling search query and source citation echo in web search results.
+        -   **[Optimization] Claude Protocol Web Search & Thinking Stability (PR #2007)**:
+            -   **Remove Web Search Downgrade**: Removed the aggressive model fallback logic for web search in the Claude protocol mapper, preventing unnecessary model downgrades.
+            -   **Remove Thinking History Downgrade**: Removed the `should_disable_thinking_due_to_history` check that could permanently disable thinking mode due to imperfect message history, now relying on `thinking_recovery` mechanism for automatic repair.
         -   **UI Improvement (Fix #2008)**: Enhanced the readability of cooldown times by changing the text color to blue.
     *   **v4.1.20 (2026-02-16)**:
         *   Fixed `400 INVALID_ARGUMENT` error in Claude Proxy during tool calls.