Codex 0.149.0

同時交辦多個 AI 任務時，現在不用在終端機裡盲目切換。

在命令列輸入 codex agents 即可開啟互動式 dashboard（集中查看與操作所有任務的控制面板）。你能在這個畫面中搜尋、啟動、開啟、重新命名以及停止各項任務。如果習慣特定的鍵盤操作，介面也支援自訂快速鍵，讓你用最順手的方式管理所有背景工作。

在終端機與 Codex 對話時，如果想換到另一個資料夾處理程式碼，現在不必關掉重開。在 TUI（以文字排版的終端機互動介面）的工作階段中，你可以直接輸入 /cd 切換工作目錄，或者用 /pwd 與 /cwd 印出當前所在的路徑。這讓你在同一次交談裡隨時移動工作位置，省去退出對話再重新連線的麻煩。

另開終端機想傳話給正在跑的任務時，codex queue 提供了直接插話的管道。

這個新指令可將訊息送進本機或遠端既有的工作階段（保存對話與狀態的執行單元）。排入佇列的訊息現在能確實喚醒閒置中的工作階段；碰上重複的工作階段名稱時，系統能更聰明地辨識目標；同時完整保留貼上文字或延遲執行指令的原有語意，確保工作不失真。

習慣在終端機用 Vim 模式打字的人，編輯提示詞的手感現在更加完整。系統擴充了字元替換功能，也加入多個常用的變更動作。例如你可以用 cw 換掉當前單字，用 c$ 改寫游標到行尾的文字，或者按 cc 清空整行重新輸入。這些快捷動作讓修改輸入內容時更加順手，不必依賴退格鍵逐字倒退。

遇到連線異常或環境卡住時，診斷工具能幫你更快找出病灶。

執行 codex doctor 指令時，檢查範圍現在涵蓋四大關鍵項目：端點防護軟體的阻擋、網路與 proxy（代理伺服器，替你轉發網路請求的中介主機）連線故障、桌面應用程式的運作狀態，以及版本更新伺服器的連線能力。當工具無法正常工作時，你能第一時間釐清是公司安全軟體攔截還是網路設定出錯。

透過 SDK（軟體開發套件，讓程式碼直接叫用工具功能的套件庫）呼叫 Codex 的開發者，現在擁有更細緻的控制權。你可以在程式碼中直接傳入精確的 CLI 設定覆寫值，依照當前任務需求微調各項參數。此外，針對需要深度思考的複雜問題，推理力度的選項新加入了 max 與 ultra 兩檔設定，讓模型投入最充足的推理資源來解析難題。

透過即時通道接收輸出時，短暫的網路波動不再造成內容遺失。

負責即時傳輸的 WebRTC（讓瀏覽器與應用程式直接進行影音或資料串流的通訊協定）側邊頻道連線，在遭遇非預期的傳輸中斷時，現在會自動嘗試重新連線。重新連上之後，中斷期間積壓、尚未送出的輸出資料會完整補送，不會被無故丟棄，確保傳輸過程不漏接任何字句。

在 Windows Terminal（微軟推出的現代化命令列終端機）中使用 Codex 的人，回頭翻閱過去的討論紀錄變得更加方便。以前在行內 TUI 模式下互動時，對話內容往往無法留在終端機自帶的 scrollback（向上捲動查看歷史輸出的緩衝區）裡；現在這些行內歷程會完整保留。你可以隨時用滑鼠滾輪往上滑，查看先前的程式碼與回應。

背景掛著多個任務時，系統不再因為保留太多舊輸出而吃掉記憶體。

針對目前未處於活躍狀態的 TUI 對話串，其 replay buffer（用來重播串流過程的快取緩衝區）現在被設下了明確的儲存上限。這項限制能防止背景對話過度保留落落長的文字串流，大幅減輕記憶體負擔，讓長時間開著多個工作階段的操作更加輕盈穩定。

有意為 Codex 開源專案貢獻程式碼的開發者，現在有了更清晰的協作指引。官方在貢獻規範中明確指出，外部開發者的提案應先透過 GitHub issue 提出並進行架構設計討論，取得共識後再著手實作，而非直接送出 pull request（請求專案管理者合併程式碼的申請）。在動手寫程式前先對齊設計，可避免白費力氣。

在安全容器內執行未受信任的程式碼時，現在有更詳盡的防護風險指引。官方文件深入記錄了 secure devcontainer（用來隔離環境的開發容器）的信任限制。內容特別點出了 DNS 資料外洩的潛在威脅，說明惡意程式可能利用網域名稱查詢將機密資料外洩。這份說明提醒團隊在隔離環境中仍需嚴格審視網路外連政策。

When juggling multiple AI tasks, you no longer have to switch blindly across terminal windows.

Run codex agents in your terminal to open an interactive dashboard. Inside this view, you can search, start, open, rename, and stop tasks. The interface also provides configurable shortcuts so you can navigate and manage all your background jobs using your preferred keystrokes.

When talking with Codex in the terminal, you no longer need to exit and restart just to switch folders. Inside a TUI (text-based user interface) session, you can now enter /cd to change your working directory, or run /pwd and /cwd to print your current location. This allows you to navigate across different project folders within the same conversation without losing your ongoing thread.

When you need to send instructions from another window to an existing task, codex queue offers a direct channel.

This new command delivers messages to existing local or remote sessions. Queued messages now reliably wake idle sessions, resolve duplicate session names more intelligently, and preserve the original semantics of pasted or deferred commands without corruption.

For developers accustomed to Vim bindings in the terminal, editing prompts is now much smoother. Codex has expanded Vim editing with single-character replacement and additional change motions. You can now use cw to change a word, c$ to change to the end of the line, or cc to rewrite the entire current line. These key combinations let you reshape text in place without backspacing character by character.

When troubleshooting connection hiccups or frozen environments, the built-in diagnostic tool can pinpoint issues faster.

Running codex doctor now inspects four critical areas: endpoint protection software interference, network and proxy failures, desktop application state, and update server connectivity. When something breaks, you can immediately tell whether corporate security software or network routing is at fault.

Developers calling Codex through an SDK (Software Development Kit) now enjoy much tighter control over execution. You can directly pass exact CLI configuration overrides from your code to fine-tune settings for specific tasks. In addition, when tackling complex logic that requires deep deliberation, you can set the reasoning effort to max or ultra to dedicate maximum reasoning capacity to the problem.

When receiving output over real-time channels, brief network hiccups will no longer cause missing data.

Realtime WebRTC sideband connections now automatically reconnect after an unexpected transport loss. Once reconnected, pending output generated during the interruption is preserved and delivered rather than dropped, ensuring you never miss streamed responses.

For developers running Codex inside Windows Terminal, reviewing past discussions is now completely effortless. Previously in inline TUI mode, conversation output was not properly retained in the terminal's native scrollback buffer. Inline TUI history now remains available in Windows Terminal scrollback, allowing you to scroll up anytime to inspect previous code snippets and explanations.

When keeping multiple background tasks open, Codex will no longer consume excessive memory by hoarding old streamed text.

Replay buffers for inactive TUI threads are now strictly bounded. This restriction prevents inactive sessions from retaining massive amounts of streamed output in memory, reducing system resource usage and keeping long-running sessions lightweight and stable.

For developers looking to contribute to Codex, the contribution guidelines are now clearer. The documentation clarifies that external contributions should begin with an issue and architectural design discussion rather than jumping straight to a pull request. Aligning on the design before writing code helps avoid wasted effort.

When running untrusted code inside secure containers, developers now have clearer guidance on potential security boundaries. The documentation now details the trust limitations of secure devcontainers (isolated containerized development environments). It specifically highlights DNS exfiltration risks, explaining how malicious code could leak sensitive data via domain lookup requests. This guidance reminds teams to audit outbound network policies even within containerized setups.