Codex 0.148.0

想把終端機裡和 Codex 討論過的內容交給同事，現在不用再逐段反白複製。

在 TUI（文字介面的操作畫面）輸入 /export。完整對話會轉成 Markdown（用符號排出標題、清單等格式的純文字）。你可把結果放進剪貼簿，直接貼到別處；也可另存成新檔案。原本一段段散在畫面裡的提問、回覆和程式碼，會有可保存、可轉交的版本。

同一個任務做到一半，想試另一條路又不想毀掉原本脈絡時，可以用 codex exec fork 分出一個工作階段（一次連續工作的上下文紀錄）。

原本的工作階段會留下來，新分支可另行探索。暫時不需要的紀錄也能在 TUI 的 resume picker（挑選要接續哪段工作的清單）歸檔；之後可從同一處還原，再繼續接手。

問題已經想好時，不必把鍵盤空著等 TUI 初始化完成；現在可先寫下提示詞（交給 AI 的工作要求），把問題先整理成可送出的要求。

啟動畫面也會顯示接續工作階段與分岔工作的進度。於是你知道畫面仍在準備，還是正在替你恢復哪段工作或建立哪個分支，不會只面對毫無資訊的等待。

在符合資格的工作區，想知道目前這串對話大約用了多少 credits（可用額度）或費用，現在不必離開終端機找帳務頁面。

輸入 /status 可查看估算，終端機最底下的狀態列與終端機標題也會顯示同類資訊。這是預估值，不是承諾的最終帳單；可用與否則取決於你的工作區是否符合資格。

公司已用 AWS 的 Amazon Bedrock Runtime 跑 AI 時，Codex 現在把它列為內建 provider（提供模型連線的服務來源）。

它支援 AWS profile（AWS 的身分與連線設定）。也支援 region（服務所在區域）。GPT-5.6 的 routing 是把請求送往適合模型的規則，現在也支援。因此可依既有 AWS 身分、區域與路由來使用，而非另行接上一個外部提供者。

要在特定事件發生時自動接一段處理，hooks（掛鉤，特定事件發生時被觸發的小程式）現在不必卡著等自己的指令跑完。它能非同步執行命令，也能叫用 MCP（Model Context Protocol，讓 AI 連接外部工具的協定）工具。

這讓掛鉤可在事件後啟動命令或外部工具，同時不把等待綁死在同一條執行流程上。

隔一段時間再接回舊工作階段時，最怕從錯的資料夾開始，或發現先前的核准規則不見了。現在接續後會還原先前保存的 working directory（當時工作的資料夾）。approval policy（哪些動作需先徵得同意的規則）也會回來。逐字稿預覽也更準確。

因此在 TUI 選擇接續時，較能辨認將要接手的內容與它原來工作的脈絡。

工作進行時 provider（提供模型服務的連線來源）暫時出問題，不再等同這一輪工作必須就此中斷。turn 會嘗試穿過暫時的供應者中斷重新連線。MCP（Model Context Protocol，讓 AI 連接外部工具的協定）伺服器也會恢復。這發生在 OAuth（用既有帳號授權第三方服務的登入方式）重新驗證之後。

兩種恢復都不需重新啟動 Codex。

看清 TUI（文字介面的操作畫面）之前，終端機先前暫存的按鍵輸入不會再突然啟用畫面上的提示或選項。這可避免某個 prompt（等待你輸入或選擇的詢問）在你讀到內容前就被誤觸。這項修正針對啟動時殘留的終端機輸入。

你能先確認眼前的詢問，再決定要輸入或選哪個選項。

登入驗證尚未備妥時，Codex 不會只停在缺少 authentication（證明你身分的登入驗證）的狀態。畫面會顯示 onboarding（第一次使用時帶你完成必要步驟的引導）。這是認證缺少時顯示的初次使用引導。它把缺少驗證的情況接到初次使用所需步驟，讓你看得見接下來該完成什麼。

把別處複製的文字貼進 composer（撰寫提示詞的輸入區）時，CRLF 是某些文字系統用兩個符號表示換行的格式。自動折行產生的空白和很長的 URL（網站位址）也都可能讓畫面看起來怪異。現在編寫區與 transcript（對話紀錄顯示區）會正確處理這些內容。

貼上含特殊換行、折行空白或長網址的內容時，文字應維持可讀的呈現。

沙盒是一種把程式可做的事圈在安全範圍內的隔離環境。遇到不允許或無法讀取的路徑時，現在採取 fail closed（無法確認安全時先拒絕）的做法。

這項限制涵蓋 Linux 和 Windows。換句話說，路徑被拒絕或不可讀時，沙盒不會繼續放行，而是以拒絕來保住限制的邊界。

替 Codex 建立 skill（讓它學會特定工作方法的操作指引）時，內附的 skill-creator guide 現在更聚焦。這份隨工具附上的說明集中於建立這類指引的工作，內容重點就是建立這類操作指引。

準備撰寫 skill 時，可參考這份更聚焦的 guide。

準備讓 skill 通過驗證時，未完成的 TODO（提醒日後補做的佔位文字）不能再留在內容裡。validation（檢查內容是否符合要求的程序）現在會拒絕這類佔位文字。這項拒絕針對仍未完成的 TODO 佔位文字。

要通過驗證，skill 內容不能還留著這些未完成 TODO。

Need to hand a Codex conversation to a teammate? You no longer have to select and copy it piece by piece.

Type /export in the TUI, the text-based interface, to turn the complete conversation into Markdown. Put it on the clipboard for pasting elsewhere, or save it as a new file. Questions, replies, and code that were spread across the screen now have a shareable copy.

Halfway through a task, you may want to try another path without losing the original thread. Run codex exec fork to split off a session, the record of one continuous piece of work.

The original session remains available while you explore the new one. You can also archive an inactive session from the TUI resume picker and restore it there later.

With the question already in mind, you no longer need to leave the keyboard idle while the TUI initializes: start drafting the prompt, the instruction for AI.

Startup also shows progress for resuming and forking work. You can tell whether the screen is still preparing, restoring a session, or creating a branch instead of watching an unexplained wait.

In an eligible workspace, you can check an estimate of this thread's credits or cost without leaving the terminal for a billing page.

Run /status for the estimate; terminal status lines and titles can show it too. It remains an estimate rather than a final bill, and availability depends on workspace eligibility.

If your company already runs AI through Amazon Bedrock Runtime on AWS, Codex now includes it as a built-in provider, a source of model connections.

It supports AWS profiles and regions. It also supports routing for GPT-5.6. That lets it use your existing AWS identity, location, and routing choices rather than requiring another external provider connection.

When an event should trigger an automatic follow-up, hooks, small programs triggered by specific events, no longer have to wait for their commands to finish. They can run commands asynchronously and invoke MCP tools, where MCP is a protocol that connects AI to external tools.

A hook can start a command or external tool after an event without tying the same execution path to that wait.

When returning to an earlier session, starting in the wrong folder or losing its approval rules is disruptive. Resumed sessions now restore the persisted working directory, the folder used for that work, and approval policy, the rules for actions that need consent. Transcript previews are also more accurate.

The TUI resume choice better shows what you are about to pick up and its original working context.

A temporary provider outage, where the source of model service connection has trouble, no longer means the current turn must end there. Turns reconnect through temporary provider outages. MCP servers, with MCP being the protocol connecting AI to external tools, also recover after OAuth reauthentication, where an existing account authorizes a third-party service.

Neither recovery requires restarting Codex.

Before you can read the TUI, the text-based interface, buffered terminal keystrokes no longer unexpectedly activate prompts or choices on screen. That prevents a prompt, a question waiting for input or selection, from being triggered before you can read it.

You can inspect the question first, then decide what to enter or select.

If Codex lacks authentication, the login check that proves identity, it no longer leaves you only at that missing-authentication state. It shows onboarding, guidance through the steps needed for first use. This connects the missing check to the required first-use steps.

Pasting text into the composer, the area for writing prompts, can bring CRLF, a two-symbol line-break format used by some text systems. Wrapped whitespace and long URLs, or web addresses, can also make the display render incorrectly. The composer and transcript, the conversation record display, now handle them correctly.

Pasted special line breaks, wrapped spaces, and long URLs should now stay readable.

A sandbox is an isolated environment that limits what a program can do. When it meets a denied or unreadable path, it now fails closed: it refuses access when safety cannot be confirmed.

This restriction applies across Linux and Windows. A denied or unreadable path is not allowed through; the sandbox keeps the boundary by rejecting it.

When you create a skill, an instruction guide that teaches Codex a particular way of working, its bundled skill-creator guide is now more focused. That guide concentrates on creating this kind of instruction.

Use the more focused guide when writing a skill.

Before a skill can pass validation, unfinished TODO placeholders, text left to complete later, can no longer remain in its contents. Validation, the check that content meets requirements, now rejects those placeholders.

A skill that passes validation cannot keep unfinished TODO placeholders.