Codex 0.155.0

想直接用語音向 AI 說明需求而不想一直敲鍵盤？Codex 在支援的建置版本中推出了實驗性的 /voice 語音對話功能。

你只需先在對話中輸入 /experimental 指令開啟實驗性功能清單，就能啟用這項模式。進入語音對話後，終端機畫面會同步顯示即時對話轉錄（transcript，將語音同步轉成文字的逐字紀錄），同時提供專門的麥克風控制項供你切換收音狀態，讓口語溝通與文字紀錄同步並行。

等待 AI 思考程式碼架構時，你不再需要對著空白畫面發呆。現在文字終端機介面的狀態列（status row，終端機底部的即時狀態提示區）會同步顯示即時推論摘要，清楚交代模型當前的思考步驟。當該回合順利完成後，畫面還會主動標註完成時間戳記。

有了這項改進，你隨時都能掌握目前的運算進度與確切完成時刻。

管理多個背景代理人與平行分支時，專案總覽介面變得更井然有序。代理人總覽頁面現在支援任務的隱藏、封存與刪除功能，方便你隨時整理已完成或暫時不需追蹤的項目。此外，系統也加入了工作樹（worktree，Git 讓同一專案平行開多份副本的功能）的歸屬詳細資訊，並支援在二次確認後安全刪除乾淨無殘留修改的託管工作樹。

這讓維護龐大專案時的平行分支變得更加清爽安全。

在 Mac 上使用命令列介面串接外部工具時，安全驗證變得更俐落省時。針對支援的 Mac 機種，本機文字終端機的工作階段現在支援以 Touch ID 指紋辨識來驗證 MCP（Model Context Protocol，讓 AI 連接外部工具的協定）的操作請求。同時，當外部 MCP 伺服器的 OAuth 授權憑證過期時，系統能精確回報錯誤狀態。一旦授權權杖（token，用來驗證身分的授權令牌）重新整理失敗，介面還會主動提供清楚的重新連線指引，引導你完成登入驗證。

常駐背景執行的伺服器現在具備更靈活的更新彈性與還原韌性。你可以自訂守護行程（daemon，在背景持續運作的常駐服務程式）的自動更新排程，或隨時執行 codex app-server daemon update 指令手動檢查並套用更新。更棒的是，即使守護行程在背景重新啟動，已儲存的對話串與當前進行中的目標也能自動恢復，不會因服務重啟而遺失工作脈絡。

透過 Amazon Bedrock 調用模型時，雲端憑證的管理變得更加自動化。系統現在可以直接執行你所設定的自訂指令來取得 AWS 登入憑證，並利用快取（caching，把查過的資料暫存起來直接用）機制將憑證暫存備用。當憑證即將過期時，背景會依據有效期限自動重新整理，並在遭遇認證異常時主動執行復原程序。

這免去了頻繁手動更新金鑰的繁瑣步驟。

仰賴系統自動審核高權限操作時，判斷流程現在更具可信度與容錯力。自動核准審查機制在評估時能更可靠地保留完整操作步驟與授權佐證，提供完整的稽核依據。面對網路或伺服器引發的暫時性故障，審查流程會自動重試；同時也能明確區分審查流程本身的執行失敗與真正偵測到危險操作的警示結果，避免將系統異常誤判為安全威脅。

在同一台電腦切換不同使用者帳號時，前一個帳號的敏感連線與資源將被徹底清空。一旦完成身分切換，系統會立即讓屬於前一身分的遠端控制工作階段全部失效，並清除快取（cache，把連線狀態暫存起來的機制）的 WebSocket 連線資料與模型目錄，確保不同帳號間的權限與資料完全隔離。

在受限環境中執行腳本或指令時，系統安全防護進一步升級。Codex 成功阻擋了在受限的 WSL 沙盒（sandbox，隔離程式執行權限的保護環境）中逃逸並啟動 Windows 程序的漏洞。同時，系統也強化了中介代理 Shell 快照的保護機制，防止機密的登入憑證在環境快照中遭到外洩。

這讓你在隔離環境測試未知程式碼時更加安心。

使用 Python 開發套件呼叫 Codex 服務時，套件版本與核心工具將保持緊密同步。Python SDK 與執行階段套件的發布流程現在全面對齊穩定版 CLI 發布，雙方採用完全一致的版本號碼。在 SDK 正式上架發布前，系統還會嚴格驗證執行階段所需的各項資源檔案，確保套件環境完備無虞。

這能有效避免套件版本錯置或缺檔的困擾。

Want to explain your requirements to the AI directly with your voice instead of constantly typing on a keyboard? Codex has introduced experimental /voice conversations on supported builds.

You only need to type the /experimental command in a conversation to open the experimental features list and enable this mode. After entering a voice conversation, the terminal screen simultaneously displays live transcripts alongside dedicated microphone controls to toggle audio capture states, keeping spoken communication and written records in sync.

You no longer need to stare at a blank screen while waiting for the AI to reason about code architecture. The terminal UI status row now displays live reasoning summaries in real time, clearly showing the model's current thinking steps. Once a turn completes successfully, the screen also proactively displays a completion timestamp.

With this improvement, you can stay informed about ongoing computation progress and exact completion times.

Managing multiple background agents and parallel branches is now much more organized. The agents overview page now supports hiding, archiving, and deleting tasks, making it easy to organize completed items or tasks you no longer need to track. In addition, the system includes worktree ownership details and supports confirmed deletion of clean managed worktrees without lingering modifications.

This keeps parallel branches cleaner and safer when maintaining large projects.

Security verification is now faster and more streamlined when connecting external tools via the command line on macOS. On supported Mac models, local terminal UI sessions now support Touch ID verification for MCP (Model Context Protocol) operation requests. Meanwhile, when OAuth credentials on external MCP servers expire, the system accurately reports the error state. If token refresh fails, the interface proactively provides clear reconnection guidance to walk you through login verification.

Background servers now offer more flexible updates and recovery resilience. You can configure automatic update schedules for the daemon or run codex app-server daemon update at any time to manually check and apply updates. Furthermore, even if the daemon restarts in the background, saved threads and active goals recover automatically without losing workflow context.

Cloud credential management is now more automated when invoking models via Amazon Bedrock. The system can now execute configured custom commands directly to obtain AWS credentials and cache them for subsequent use. When credentials near expiration, background processes automatically refresh them based on their lifespan and initiate recovery procedures if authentication anomalies occur.

This eliminates the hassle of manually updating keys frequently.

Automated review of elevated permissions is now more reliable and fault-tolerant. Automatic approval reviews preserve complete action steps and authorization evidence more reliably during evaluation, providing a solid audit trail. The review process automatically retries upon transient network or server failures, while clearly distinguishing review execution failures from unsafe-action findings to avoid misclassifying system anomalies as security threats.

Switching between user accounts on the same machine now completely clears sensitive connections and resources from the previous identity. Once account switching completes, the system immediately invalidates all remote-control sessions belonging to the prior identity and clears cached WebSocket connection state and model catalogs, ensuring complete isolation of permissions and data between accounts.

Security safeguards are further enhanced when running scripts or commands in restricted environments. Codex successfully blocks escapes from restricted WSL sandboxes that attempt to launch Windows processes. Meanwhile, the system hardens brokered shell snapshots to prevent confidential credentials from leaking through environment snapshots.

This gives you greater peace of mind when testing untrusted code in isolated environments.

Package versions and core tools now stay closely synchronized when invoking Codex services via Python. Publishing workflows for the Python SDK and runtime packages are now fully aligned with stable CLI releases, using matching version numbers across both. Before SDK publication, the system rigorously verifies required runtime assets to ensure a complete and ready environment.

This effectively prevents mismatched versions or missing asset issues.