Codex 0.158.0

在全螢幕文字介面（TUI，以純文字排版的終端機操作畫面）瀏覽或整理對話紀錄時，操作方式變得更貼近日常習慣。

現在你可以在設定中開啟選取即複製與右鍵貼上功能。從對話紀錄選取複製的內容，會完整保留 Markdown（用簡單符號標註粗體與標題的純文字排版格式）語法，貼到其他筆記工具時不再失去格式。

當你想要串接需要特殊身分驗證的外部服務時，現在可以直接提供專屬金鑰完成連線。

Codex 現在支援需要預先註冊 OAuth 客戶端密鑰的 MCP（讓 AI 連接外部工具的協定）伺服器。你可以直接在終端機執行帶有專屬參數的指令來新增伺服器：

codex mcp add --oauth-client-secret

面對要求事先註冊密鑰的驗證機制，也能順利連接並整合至日常工作中。

在背景執行伺服器與主程式進行即時通訊時，傳輸通道的安全防護再度升級。

Codex 現在使用 bearer token（持有者出示即可驗證身分的授權令牌）來保護直接連向 exec-server 的 WebSocket 連線。這項安全保護也包含經由 app-server 所配置的連線管道。透過令牌驗證機制，連線能防止未授權的外部存取，保障通訊過程的安全性。

請 AI 製作或修改圖像素材時，產出的圖片能更方便地直接放入簡報與網頁中。

現在你在下達生成或修改圖片的指令時，可以明確要求輸出透明背景。圖片編輯功能也支援直接讀取對話紀錄中已存為檔案的圖像，讓你不用重新匯出上傳就能接續調整畫面。

當你在等待指令審核的過程中想發問或確認狀態時，操作不再會意外中斷。

現在若在待審核狀態下輸入新文字，審核流程會自動重試，詢問狀態不會直接中斷等待執行的操作。針對以提升權限執行的指令，終端機輸入審核預設開啟，僅限執行階段授予的權限也不再引發多餘的審查。在 macOS 上套用修正時，系統能識別既有權限已涵蓋的路徑別名（alias，指向相同位置的替代路徑名稱），避免重複跳出確認提示。

請 AI 繪製流程圖或架構圖時，終端機顯示圖表的相容性變得更完整。

Mermaid（用純文字繪製圖表的語法）流程圖現在能正常渲染帶有雙引號的標籤與「&」符號。如果遇到目前尚未支援的圖表類型，畫面會明確說明退回備援（fallback，無法顯示時改用原始文字呈現的後備機制）的原因，不再讓人摸不著頭緒。

透過客戶端應用程式監看指令執行狀況時，原本可能漏看開頭輸出，也難以得知行程為何沒啟動。

現在指令完成事件中會包含執行最初期產生的前期輸出內容。當啟動行程（process，作業系統中正在運作的程式個體）遭遇失敗時，也會主動將錯誤回報給客戶端，讓排查問題更有效率。

Browsing or organizing conversation history in the full-screen terminal interface now aligns more closely with everyday habits.

You can now enable copy-on-select and right-click paste in settings. Content copied from chat history fully preserves Markdown syntax, so it no longer loses formatting when pasted into other note-taking tools.

When you want to connect external services that require special authentication, you can now directly provide a dedicated key to establish the connection.

Codex now supports MCP servers that require a pre-registered OAuth client secret. You can add a server directly by running a command with dedicated flags in the terminal:

codex mcp add --oauth-client-secret

This allows you to smoothly connect and integrate authentication mechanisms that require pre-registered secrets into your daily workflow.

When the background execution server communicates in real time with the main program, security protection for the transmission channel is upgraded once again.

Codex now uses a bearer token to protect direct WebSocket connections to exec-server. This security protection also covers connection channels configured through app-server. Through token authentication, connections prevent unauthorized external access and ensure communication security.

When asking AI to create or modify image assets, generated images can be placed directly into presentations and webpages more conveniently.

You can now explicitly request a transparent background when issuing prompts to generate or modify images. The image editing feature also supports reading images saved as files directly from the conversation history, allowing you to continue adjusting visuals without re-exporting and uploading.

Asking questions or checking status while waiting for command review no longer accidentally interrupts the operation.

If new text is entered while in a pending review state, the review process automatically retries, ensuring status inquiries do not directly interrupt operations awaiting execution. For commands executed with elevated permissions, terminal input review is enabled by default, and runtime-only permissions no longer trigger redundant audits. When applying fixes on macOS, the system recognizes path aliases covered by existing permissions, avoiding duplicate confirmation prompts.

Terminal chart compatibility is now more comprehensive when asking AI to generate flowcharts or architecture diagrams.

Mermaid flowcharts now properly render labels containing double quotes and "&" symbols. When encountering currently unsupported chart types, the screen clearly explains the reason for the fallback, preventing any confusion.

When monitoring command execution through a client application, initial output could previously be missed, and diagnosing why a process failed to launch was difficult.

Command completion events now include early output generated at the very beginning of execution. When launching a process fails, the error is actively reported to the client, making troubleshooting much more efficient.