Codex 0.152.0

你在 Codex 的輸入框切到 Vim 模式（改用鍵盤符號下指令與搜尋的編輯方式，名字來自一款老牌文字編輯器）改一段長草稿，想找回前面提到的某個檔名，以前得整段慢慢往回捲——現在按 / 往下找、? 往上找，符合的字會被反白標出來，找到了按 n 跳下一個、N 跳回上一個，

。另外，每次開新草稿，游標現在直接能打字：不管是剛送出一則訊息，還是打完斜線指令（/ 開頭觸發的內建指令）後接著開新的一段，都不用再手動切換才能輸入。

被限流（rate limit，用量到頂被暫時擋下）擋下來的時候，以前只會看到一條乾巴巴的橫幅告訴你現在不能用，接下來要自己想辦法。現在這條橫幅上直接放了幾個按鈕：查目前用量、管理額度、重設限制、管理訂閱方案，點哪個就辦哪件事，不必先關掉提示再去別的地方翻設定。

終端機介面跟 codex exec 這個非互動指令，現在會即時顯示憑證（credential，證明你是合法使用者的金鑰）更新到哪一步了，連透過 Amazon Bedrock 重新認證這種比較繞的流程也一樣有進度可看。以前遇到需要重新拿授權，畫面往往就是卡住不動，你根本不知道它是在處理還是當機了。

MCP（Model Context Protocol，讓 AI 連接外部工具的協定）伺服器的名稱以前只能用簡單字元。現在可以放 :、@、/、. 這些符號，等於能照套件命名慣例取名，像 @scope/name 這種寫法，指令列跟認證流程全部認得。同時，個別 MCP 工具現在能設定 output_token_limit，限制單次輸出的 token（AI 處理與計費的文字單位）上限，而且不管你怎麼把工作階段接回來繼續用，這個截斷規則都會維持一致。

以前 Codex 會內建一個規劃工具，現在這個功能改成預設不啟用。如果你要用，得自己在設定檔裡加上這一段：

{"tools": {"update_plan": {"enabled": true}}}

也就是把 tools.update_plan.enabled 設成 true。這不是新推出的功能，而是把原本一律開著的東西收緊成選用項目。

自動批准的審查機制這次補強了記性。以前訊息一長、對話一多，審查能保留的內容就有限，現在可以留住更長的訊息，也能看見更大範圍的對話轉錄（transcript，對話的逐字紀錄）。更關鍵的是，就算對話因為太長被系統壓縮（compaction，把舊對話濃縮成摘要以節省空間），使用者當初下的指示、給過的答案，以及已經核可的有效授權，都不會被壓縮掉，審查時仍然找得到依據。

這批修復把幾個容易在背景出包的地方一次補齊。app-server 的客戶端現在能自己設定 thread/shellCommand 的逾時時間，就算要等超過一小時也設得了；把之前的工作階段接回來繼續用時，沒特別指定路徑就會自動還原成上次的工作目錄，客戶端更新描述資訊也不會動到原本的檔案系統權限。MCP 工具在快取（cache，把查過的結果暫存起來直接用）更新、遠端外掛換版本時仍然能正常呼叫。認證要重試時，也會換上輔助程式剛給的新標頭（headers，請求裡帶的驗證資訊）。開模型選單重新整理清單時，原本反白選著的那個也不會跳掉。外掛推薦則是改成開機時就先載入，不用等到你打第一句才臨時去抓。

When editing long drafts in Vim mode within the Codex composer, finding an earlier filename no longer requires manual scrolling. You can now press / to search forward or ? to search backward with highlighted matches, and navigate with n and N to jump between occurrences.

Furthermore, fresh drafts now start directly ready for typing: whether you just submitted a message or ran a slash command, you can start typing immediately without manually switching modes.

When hit by a rate limit, users previously saw only a plain banner stating that requests were temporarily blocked, leaving them to figure out the next step on their own. The banner now includes direct action buttons to check current usage, manage credits, reset limits, and manage subscription plans, allowing you to handle each task immediately without dismissing the alert to search through settings.

The terminal UI and the non-interactive codex exec command now display live progress when refreshing credentials, including multi-step flows like Amazon Bedrock reauthentication. Previously, reauthentication would leave the screen hanging with no feedback, making it impossible to tell whether the process was working or frozen.

MCP server names were previously limited to simple characters. They can now include :, @, /, and ., enabling standard package-style naming such as @scope/name across all CLI commands and authentication flows. In addition, individual MCP tools now support an output_token_limit setting to cap output tokens per invocation, maintaining consistent truncation rules across resumed sessions.

Codex previously enabled its built-in planning tool out of the box, but the feature is now disabled by default. To use it, you must add the following snippet to your configuration file:

{"tools": {"update_plan": {"enabled": true}}}

This sets tools.update_plan.enabled to true. Rather than introducing a new feature, this update makes an existing capability opt-in.

Automatic approval reviews now retain richer conversation context. While earlier reviews were limited when handling lengthy exchanges, the system can now hold longer messages and larger conversation transcripts. Crucially, when conversations undergo history compaction to condense past turns, original user instructions, provided answers, and granted authorizations are preserved rather than compressed away, ensuring reviews retain full context.

This release fixes multiple background edge cases. Clients connecting to app-server can now configure thread/shellCommand timeouts, including deadlines longer than one hour. Resumed threads automatically restore their previous working directory when none is specified, and client metadata updates preserve existing filesystem permissions. In addition, MCP tools remain callable during cache refreshes and remote plugin updates, while authentication retries adopt newly refreshed helper headers. Opening the model picker refreshes the model list without losing your currently selected item, and plugin recommendations begin loading at startup to minimize delays before the first prompt.