Vim 模式加了復原與重做,草稿手滑也不怕
你在 Vim 模式(改用鍵盤符號下指令的編輯方式,名字來自一款老牌文字編輯器)裡改草稿,手滑刪掉一整段附件說明,以前只能重打——現在按 u 就能復原上一步,按 Ctrl+R 則是反悔剛才的復原、重做回去,
。而且不管草稿裡貼了多少內容、掛了幾個附件,復原跟重做都會保留完整版本,不會漏東漏西,你不用擔心救回來的稿子少了一塊。
發布日期:2026-09-03
你在 Vim 模式(改用鍵盤符號下指令的編輯方式,名字來自一款老牌文字編輯器)裡改草稿,手滑刪掉一整段附件說明,以前只能重打——現在按 u 就能復原上一步,按 Ctrl+R 則是反悔剛才的復原、重做回去,
。而且不管草稿裡貼了多少內容、掛了幾個附件,復原跟重做都會保留完整版本,不會漏東漏西,你不用擔心救回來的稿子少了一塊。
想幫 Codex 加裝外掛(plugin,替它補上額外功能的附加元件)?現在外掛的指令列工具直接打通遠端市集:想看有什麼可裝,打指令就能列出清單;看中哪個,一行指令裝上;用膩了,一行指令移除,全程不用離開終端機。
想關掉自動回顧摘要,或把貼上大量文字的偵測設定搬到新位置,這裡有兩個新設定鍵。Codex 原本會自動生成回顧摘要,如果你只想在打 /recap 時自己叫出來,設定檔加一行:
{"tui": {"auto_recap": false}}也就是 tui.auto_recap = false,自動摘要關掉,手動 /recap 照樣能用。另一個是貼上偵測(paste burst,短時間內貼上大量文字的辨識機制)設定鍵搬家,改用巢狀寫法:
{"tui": {"disable_paste_burst": true}}也就是 tui.disable_paste_burst,而舊的頂層寫法目前仍當備援繼續有效,不用急著改。
回頭捲動 Codex 的終端機介面歷史紀錄時,現在看得到的東西更完整了:程式碼修改會顯示完整的補丁(patch,改動前後對照的差異紀錄),送進背景終端機視窗的輸入內容會留底,個別執行完的指令也各自獨立顯示,不會被吞掉或混成一團。
用 Plus 或 Team 方案的人,額度用得快不快,現在能更早知道。以前要等額度更接近用完才會收到提醒,常常是打到一半才發現要沒了。現在只要在大約五小時一輪的用量視窗裡,額度剩不到一半,就會提早跳出警告,你就有機會提前調整步調,不必臨時改變工作方式。
你透過外部的 app-server(讓其他程式跟 Codex 對話的伺服器介面)連線用 Codex,這條連線如果突然斷掉,
。具體來說,Codex 現在會自動重新連上這個外部 app-server,草稿跟對話轉錄(transcript,逐字紀錄)都完整保留。而斷線當下還不確定有沒有送出、或是還在排隊等處理的動作,會先暫停下來,交給你重新確認過再繼續,不會替你自作主張。
Codex 的 Guardian(把關風險動作的審查機制)這次分清該管與不管的界線。用 Full Access(完全授權模式)時,遇到只需按確認的動作,Guardian 現在直接跳過審查。換成每個動作都要核准的 User approval mode,Guardian 也不再背景幫提案打分數、預先熱機,但敏感動作檢查與需要你輸入答案的請求,處理方式不變。Guardian 過去審過的紀錄也扛得住考驗:對話被壓縮、程式重啟,或你自己開了分支,紀錄都留得住,仍遵守回捲(rollback,把狀態退回先前某刻)的邊界。子代理(subagent,替它跑某段子任務的分身)的歷史也被隔開,不會混進主線程。
Codex 新增了一個預設關閉的實驗設定 features.context_management.experimental_mode,如果你符合資格想試,在設定檔加上:
{"features": {"context_management": {"experimental_mode": true}}}也就是把這個巢狀設定鍵打開。開啟後,如果你是用 Codex 後端的 ChatGPT Plus、Pro 或 Pro Lite 訂閱、而且帳號被納入資格名單,就會啟用以 token(AI 處理與計費的文字單位)預算管理上下文的機制,外加歷史筆記功能跟 new_context 這個工具。用 API 金鑰(一組用來證明身分、直接呼叫底層服務的密碼字串)登入的工作階段、自訂供應商,以及暫時性的結構化執行緒,則不在開放範圍內,不會啟用這些新機制。
這批修復收攏了幾項後台細節。MCP(Model Context Protocol,讓 AI 連接外部工具的協定)工具的核准記憶改成綁定帳號,換帳號不會沿用另一個帳號核准過的設定;MCP 相對路徑執行檔,在 macOS 上啟動更可靠。對話紀錄壓縮現在併入共享歷史,codex exec resume 依目錄續接也認得壓縮過的紀錄。工作階段的根目錄就算是符號連結(symlink,指到別處的捷徑檔),開執行緒分支照樣正常。app-server 執行緒中繼資料多了可為空值的 model、reasoningEffort 欄位,模型目錄若開放,還能用 request_user_input_async 結構化、非同步發問,不必卡住等你即答。
When modifying drafts in Vim mode, accidentally deleting a section of attachment notes previously required retyping the whole piece. You can now press u to undo the last action and Ctrl+R to redo, reversing the undo operation.
Furthermore, regardless of how much content was pasted or how many attachments were included, undo and redo preserve the complete state without dropping parts of your draft.
Looking to install plugins to expand Codex with extra capabilities? The plugin CLI now connects directly to remote marketplaces: you can run a command to list available plugins, install a chosen plugin with a single command, or remove it just as easily when no longer needed, all without leaving your terminal.
This release introduces two new configuration keys for managing recap behavior and paste detection. Codex previously generated recap summaries automatically. If you prefer to trigger recaps manually with /recap, add the following line to your configuration file:
{"tui": {"auto_recap": false}}This sets tui.auto_recap = false, disabling automatic generation while keeping manual /recap invocations available. Additionally, the paste burst detection setting has moved to a nested structure:
{"tui": {"disable_paste_burst": true}}This introduces tui.disable_paste_burst, while the legacy top-level key continues to work as a fallback so immediate migration is not required.
Scrolling back through TUI history in Codex now provides a much clearer view of past actions. Code modifications display complete patch diffs, inputs sent to background terminal sessions are fully recorded, and individual completed commands appear distinctly without being omitted or merged together.
Subscribers on Plus and Team plans can now gauge their quota consumption earlier. While warnings were previously triggered only as allowances were nearly exhausted, alerts are now sent as soon as less than half of the allowance remains within an approximately five-hour usage window. This gives you time to pace your prompts without having to abruptly adjust your workflow.
When using Codex through an external app-server connection that suddenly drops,
. Codex now automatically reconnects to the external app-server, preserving drafts and conversation transcripts intact. Any submissions that were in-flight or queued when the connection dropped are held in a paused state for user review before proceeding.
Guardian reviews in Codex now feature clearer operational boundaries. In Full Access mode, Guardian skips reviews for confirmation-only actions. When running in User approval mode, it stops background proposal scoring and prewarming while preserving existing checks for sensitive actions and user input requests. Furthermore, Guardian review histories now survive compaction, application restarts, and user-created forks while strictly respecting rollback boundaries and isolating subagent history from the main thread.
Codex introduces the disabled-by-default features.context_management.experimental_mode configuration setting. To try it when eligible, add the following snippet to your configuration file:
{"features": {"context_management": {"experimental_mode": true}}}When enabled for eligible ChatGPT Plus, Pro, or Pro Lite sessions using the Codex backend, it activates token-budgeted context management, history notes, and the new_context tool. API-key sessions, custom providers, and temporary structured threads remain excluded from these new capabilities.
This release refines several backend implementation details. Remembered MCP tool approvals are now scoped to the active account so switching accounts will not reuse permissions granted by another user, and relative MCP executable paths launch more reliably on macOS. Rollout compression now incorporates shared histories, allowing codex exec resume to handle compressed rollouts when resuming by working directory, while thread forks operate cleanly with symlinked session roots. Additionally, app-server thread metadata includes nullable model and reasoningEffort fields, and request_user_input_async enables structured asynchronous prompts when supported by the model catalog.
Vim mode now supports undo with `u` and redo with `Ctrl+R`, preserving complete drafts including pasted content and attachments. ([#41941](https://github.com/openai/codex/pull/41941), [#42140](https://github.com/openai/codex/pull/42140))The plugin CLI can list, install, and remove plugins from remote marketplaces. ([#42150](https://github.com/openai/codex/pull/42150))Set `tui.auto_recap = false` to disable automatic recaps while keeping manual `/recap` available. ([#42101](https://github.com/openai/codex/pull/42101))
`tui.disable_paste_burst` replaces the top-level setting, which remains supported as a fallback. ([#41976](https://github.com/openai/codex/pull/41976))TUI history shows complete patches, input sent to background terminals, and individual completed commands. ([#41893](https://github.com/openai/codex/pull/41893), [#42107](https://github.com/openai/codex/pull/42107))Plus and Team users receive an earlier warning when less than half of their allowance remains in an approximately five-hour usage window. ([#42142](https://github.com/openai/codex/pull/42142))TUI sessions reconnect after an external app-server connection drops, preserving drafts and transcripts while keeping uncertain or queued submissions paused for review. ([#41911](https://github.com/openai/codex/pull/41911), [#41916](https://github.com/openai/codex/pull/41916), [#41918](https://github.com/openai/codex/pull/41918))Full Access skips Guardian reviews for confirmation-only actions. User approval mode skips background Guardian scoring and prewarming, while sensitive-action checks and requests for user input retain their existing handling. ([#42147](https://github.com/openai/codex/pull/42147), [#42256](https://github.com/openai/codex/pull/42256))
Guardian review history survives compaction, restarts, and user-created forks while respecting rollback boundaries and isolating subagent history. ([#41879](https://github.com/openai/codex/pull/41879), [#42065](https://github.com/openai/codex/pull/42065))Adds the disabled-by-default `features.context_management.experimental_mode` configuration. When enabled for eligible ChatGPT Plus, Pro, or Pro Lite sessions using the Codex backend, it activates token-budget context, history notes, and the `new_context` tool. API-key sessions, custom providers, and temporary structured threads remain excluded. ([#42385](https://github.com/openai/codex/pull/42385))Remembered MCP tool approvals are scoped to the selected app account, and relative MCP executable paths start more reliably on macOS. ([#42133](https://github.com/openai/codex/pull/42133), [#42117](https://github.com/openai/codex/pull/42117))
Rollout compression includes shared histories, `codex exec resume` handles compressed rollouts when selecting by working directory, and thread forks work with symlinked session roots. ([#42039](https://github.com/openai/codex/pull/42039), [#42135](https://github.com/openai/codex/pull/42135))
App-server thread metadata includes nullable `model` and `reasoningEffort` fields. Structured asynchronous questions are supported through `request_user_input_async` when enabled by the model catalog. ([#42151](https://github.com/openai/codex/pull/42151), [#42178](https://github.com/openai/codex/pull/42178))