Codex 0.159.0

當你看著模型正洋洋灑灑回答、或是執行耗時的程式呼叫時，只要打字就能立刻把它的思緒拉回新方向。啟用 instant_interrupt 設定後，你不必苦等當前輸出跑完或強行中斷重來，送出新提示詞就能在運算當下直接引導 Codex 調整下一步。

如果你發現剛才的提問有誤或想臨時補充細節，這項機制讓你隨時插話，省去反覆等待與重啟對話的時間。

在打了一長串提示詞卻臨時想切去處理別的工作時，留在輸入框的草稿現在不會無故消失。就算當前是一個尚未展開對話的全新空白工作階段（session，開啟工具後與 AI 互動的獨立工作階段），切換任務時輸入內容依然會妥善保留。同時，你還沒送出第一則訊息的對話串，現在也能直接列在清單中或執行歸檔。

這讓你在整理多個進行中的任務時，不必擔心手滑切換而弄丟寫到一半的想法。

當你開啟新的工作階段時，終端機現在會以更精簡的歡迎畫面與統一標頭迎接你，並在對話進行中或結束後適時提供操作小提示。如果在執行時跳出警告檢視器，關閉檢視器時會自動清除已看過的警告項目，若有想稍後再確認的重點，只要按下鍵盤上的 k 鍵就能特意保留。此外，當 Codex 提出執行計畫並詢問是否採用時，你不再被卡在固定的確認畫面，而是能自由向上捲動對話記錄，仔細核對先前的上下文再做決定。這些調整讓你在終端機裡查閱與操作時更加直覺流暢。

當你請 AI 繪製架構圖或整理資料表格時，畫面呈現與複製貼上的格式現在更加精準。內建的 Mermaid（用純文字繪製圖表的語法）渲染器現在支援更多流程圖連線樣式、文字標籤與節點群組，排版不再輕易破版。同時，在終端機選取並複製對話記錄時，會完整保留 Markdown 表格結構、文字樣式與縮排空白，且支援更多終端機的反白即複製功能。

你不再需要手動補齊歪掉的欄位或重畫群組架構。

剛安裝好工具準備登入帳號時，驗證步驟變得更加順暢可靠。透過本機登入 ChatGPT 帳號時，系統現在能穩定自動開啟預設瀏覽器完成授權；如果瀏覽器未能如期跳出，初次引導畫面也額外提供了快捷鍵，可以直接複製登入網址並手動貼到瀏覽器開啟。

這避免了過去登入視窗偶爾卡住、使用者找不到授權網址而卡在開場說明的窘境。

在 Windows 系統上啟動工具時，桌面不再被突然彈出的黑色視窗打擾。執行 MCP（讓 AI 連接外部工具的協定）伺服器、程式模式主機或管道指令時，能避免跳出多餘的主控台視窗，且在權限受限的啟動環境下會自動切換為內嵌備援模式。同時，應用伺服器（app-server，負責處理外部用戶端請求的後端服務）現在支援分頁讀取功能，用戶端能指定從特定的對話項目開始拉取歷史記錄。

這讓你在 Windows 上的多工體驗更加乾淨整潔。

如果你習慣自己在終端機掌握對話節奏，畫面現在不會再主動彈出干擾思考的後續問題建議。這次更新正式移除了自動追問提示建議功能，相關的 tui.prompt_suggestions 設定項也一併清理下架。同時，系統不再預先打包隨附的 plugin-creator（外掛製作工具）技能，讓內建環境回歸輕量。

以前畫面上可能會出現用不到的建議或預載工具，現在介面與設定檔都變得更加清爽，把主導權完全交還給你的指令。

When you watch the model outputting a lengthy response or running a time-consuming program call, you can steer its train of thought in a new direction simply by typing. After enabling the instant_interrupt setting, you do not have to wait for the current output to finish or forcefully interrupt and start over; sending a new prompt directly guides Codex to adjust its next step mid-computation.

If you spot a mistake in your previous prompt or want to add details spontaneously, this mechanism lets you chime in at any time, saving you the wait of restarting conversations.

When you type out a lengthy prompt but suddenly need to switch to another task, the draft in the input box no longer disappears. Even if the current task is a brand-new blank session without any messages yet, your entered input is properly preserved when switching tasks. In addition, conversation threads where you have not sent a first message can now be listed directly or archived.

This ensures you will not lose half-written ideas when switching between multiple active tasks.

When you start a new session, the terminal greets you with a cleaner welcome screen and unified header, offering helpful tips during or after conversations. If the warning viewer appears during execution, closing it automatically clears reviewed warnings, though you can press the k key to keep specific items for later review. Furthermore, when Codex presents an execution plan and asks for confirmation, you are no longer stuck on a rigid confirmation screen; instead, you can freely scroll up through conversation history to check prior context before deciding. These adjustments make reviewing and interacting in the terminal more intuitive and fluid.

When asking AI to generate architecture diagrams or data tables, rendering and copy-pasting formats are now significantly more accurate. The built-in Mermaid renderer supports more flowchart link styles, text labels, and node subgraphs, making layouts less prone to breaking. Meanwhile, selecting and copying conversation logs in the terminal fully preserves Markdown table structures, text styles, and indentation spacing, with broader support for copy-on-select terminal behaviors.

You no longer need to manually realign misaligned columns or redraw node groups.

When logging into your account after installing the tool, the authentication workflow is much smoother and more dependable. When logging into a ChatGPT account locally, the system reliably opens your default browser to complete authorization; if the browser fails to pop up, the onboarding screen provides a keyboard shortcut to directly copy the login URL and paste it manually.

This avoids past issues where login windows hung or users got stuck without an authorization URL during onboarding.

When launching tools on Windows, your desktop is no longer disrupted by sudden black terminal pop-ups. Spawning MCP servers, programmatic mode hosts, or piped commands avoids extraneous console windows, and the system automatically falls back to an embedded mode in permission-restricted launch environments. Meanwhile, the app-server now supports pagination, allowing clients to fetch history starting from a specific conversation item.

This makes multitasking on Windows cleaner and more organized.

If you prefer managing your conversation flow in the terminal yourself, the screen will no longer pop up follow-up prompt suggestions that interrupt your train of thought. This update officially removes automated follow-up prompt suggestions and retires the corresponding tui.prompt_suggestions setting. In addition, the system no longer bundles the built-in plugin-creator skill out of the box, keeping the default environment lightweight.

Where unused suggestions or preloaded tools once cluttered the view, the interface and configuration are now cleaner, handing full control back to your commands.