Grok Build 1.0.48

當你請 AI 寫程式寫到一半想推翻重來、想跳回先前的步驟時，終端機裡的歷史清單現在不會再被雜訊淹沒。

以前打 /rewind（倒轉對話的指令）準備挑選要退回的時間點，清單裡常會混雜系統在背景自動產生的過渡紀錄。現在這項指令只會列出你親手敲進去的提示詞（prompt，向 AI 發出的指令），自動產生的執行輪次通通直接跳過。

如果你在專案設定檔裡切換或掛載自己慣用的模型，現在終端機依然能完整看見官方的重要提示。在 config.toml（管理設定選項的檔案）中設定自訂模型表格時，有時需要指定不同的模型來源。以前只要填入這些自訂內容，終端機原本會顯示的公告橫幅（notice banner，說明模型條款或使用須知的提示區塊）就會莫名失蹤。現在無論你怎麼自訂模型表格，原始公告都會安然保留在畫面上。

請 AI 幫忙翻找或修改程式碼時，最怕遇到專案裡夾雜非純文字檔案導致工具直接卡死。現在 read_file、search_replace 與 grep 遇到特殊檔案（例如系統管道或裝置捷徑）時更為穩定。面對超大檔案或二進位內容（binary content，非一般純文字的圖檔或編譯檔）時，也不容易突然當掉。就算真的無法讀取，工具也會直接給出清楚具體的錯誤訊息，不再丟出令人困惑的故障代碼。

在終端機輸入問題卻因為尚未登入而被擋下時，完成驗證後的下一步現在完全交由你決定。以前只要碰上需要登入的限制，被拒絕的提示詞（prompt，向 AI 發出的指令）在登入完成後常會自動重試。現在登入成功後它不會再擅自重新發送，只會把被擋下的登入區塊留著，需要你手動再次送出。

When you want to scrap code changes halfway through and jump back to a previous step while pairing with AI, the history list in your terminal is no longer flooded with noise.

Previously, running /rewind to select a rollback point often displayed a list mixed with intermediate records that the system generated automatically in the background. Now, this command only lists prompts you typed yourself, skipping automatically generated execution turns entirely.

If you switch to or mount preferred models in your project configuration file, your terminal still displays official notices in full. When configuring a custom model table in config.toml, you sometimes need to specify different model sources. Previously, filling in these custom details caused the terminal's notice banner to mysteriously disappear. Now, no matter how you customize the model table, the original announcement remains safely on screen.

When asking AI to search or modify code, the biggest fear is running into non-plain-text files in a project that freeze tools completely. Now, read_file, search_replace, and grep are significantly more stable when encountering special files such as system pipes or device links. When facing oversized files or binary content, they are also far less likely to crash unexpectedly. Even if a file cannot be read, the tools now provide clear, specific error messages instead of throwing confusing failure codes.

When a question entered in the terminal is blocked because you are not logged in, the next step after completing authentication is now entirely up to you. Previously, whenever you hit a login requirement, rejected prompts often retried automatically once authentication finished. Now, it no longer resends anything automatically after a successful login; instead, it retains the blocked login prompt block, requiring you to submit it again manually.