Codex 0.145.0

做了一半的長對話，最怕隔天想回頭找某段結論，卻得從頭捲到尾。實驗中的分頁對話歷史把記錄切成好翻的頁面，讓你有效率地接續工作、搜尋內容，名稱也會保存下來；它還支援子代理與記憶。這像把一疊散落的會議筆記裝進有索引的活頁夾：你不只找得到那一頁，還能從原處繼續寫下去。仍標示為 experimental，表示功能在試行，介面或行為可能再調整。

換用 Codex 時，不必把舊工具的習慣一項項重建。/import 現在可搬入 Cursor 與 Claude Code 的設定、MCP（Model Context Protocol，讓 AI 連接外部工具的協定）伺服器、外掛、工作階段、指令，以及專案專屬記憶。就像新手機不只搬通訊錄，連常用捷徑與工作進度也一起帶走。你在終端機輸入 /import 後，才依畫面選擇來源與要遷移的內容；原文沒有保證所有設定都能完全相容。

公司若透過 Amazon Bedrock 使用模型，現在多了一條可走的登入路，也能接上自訂端點（endpoint，服務接收請求的網址）與驗證方式；Bedrock 的預設模型則改為 GPT-5.6 Sol。好比辦公大樓新增了識別證入口，也允許公司接自己的專線。這些是實驗功能，適合需要把 Codex 接進既有 Bedrock 架構的人；一般直接使用 Codex 的讀者不必特別設定。

手邊是一段錄音而不是文字時，現在可以把常見本機音檔當輸入，也能在工具輸出裡取得音訊。新版另加入串流即時 V3 對話：回覆不是等全部完成才送出，而是邊產生邊傳來。它像電話通話，不必等對方把整段話錄完才聽見第一句。把音檔交給工具後，結果也可帶回音訊，讓聲音能進也能出。原文沒有列出支援格式或操作指令，因此實際可用格式仍以介面提示為準。

一件工作要分給多個代理並行處理時，這套 opt-in 多代理 V2 體驗變得更穩定。你可設定子代理使用的模型、推理強度與同時工作的數量，先前的角色也恢復，代理之間的瀏覽與切換更順手。像導演終於能替每位組員選工具、分配同時上場的人數，並清楚看見誰在做哪一幕。若任務要快或要深，可依工作性質挑不同模型與推理程度。opt-in 表示需要你主動開啟，不是所有人自動套用。

看見終端機輸出的圖表或互動視覺化時，現在能直接點安全的行內連結開啟，不必複製網址再猜要貼去哪裡。這像文件裡出現一張縮小地圖，你點一下就能進入完整地圖，而不是先抄下座標。連結放在輸出文字中，閱讀結果時就能順手打開對應視覺化；終端機也不必把圖表硬塞成難讀的純文字。原文只說連結安全且可點擊，沒有承諾所有終端機、所有視覺化種類都支援。

想把早先的提問改個字、或重試被安全機制暫緩的回合，現在不必拿整段原對話冒險。Codex 會建立帶上下文的分支，保留原本對話、附件與提及的綁定。這像在書稿影印一份後改稿：新稿能走另一條路，原稿與夾在裡面的資料仍完整留著。附件和提及的對象也會跟著原路徑保存，避免分支後還要重新交代背景。你可以比較兩種方向，而非只能覆寫前一個選擇。

外部工具很多時，MCP（Model Context Protocol，讓 AI 連接外部工具的協定）一啟動就同時登入、刷新授權，容易慢下來甚至互相衝突。新版替啟動設逾時、不讓 OAuth 探索（尋找登入授權入口的流程）阻塞整體、把刷新排隊，並安全重用工具目錄。像機場把登機口檢查分流：太久沒來的人不會堵住全隊，同一張航班看板也不用每人重新印。結果是 MCP 啟動與驗證較穩，並非新增工具功能。

在 Windows 上執行 Codex 時，這次集中補強執行與沙盒（sandbox，隔離程式可碰範圍的安全空間）可靠性：加入原生 exec-server 沙盒化、強制網路代理規則、隱藏輔助主控台，也正確處理 hook（掛鉤，特定事件發生時被觸發的小程式）指令的引號。像工地把圍籬、出入口與工作人員通訊一起整理好；你看不見的細節，正是程式能照預期跑、網路規則不被繞過的基礎。

工具準備執行強制 rm 刪除時，安全判斷更能辨認危險情況；取得完整存取權前的確認也更一致。若你拒絕某個工具的操作，拒絕原因不會在工具間遺失。這像門禁不只看「有人拿著鑰匙」，還會看他是否正要拆掉整扇門，並把你寫下的拒絕理由交接給下一位警衛。之後換另一個工具處理同件事，也能知道你先前為何說不。它強化的是防護與溝通，不是放寬刪除權限。

想查 OpenAI Docs skill（協助查詢 OpenAI 文件的內建技能）或選模型時，隨附指南已更新 GPT-5.6 的模型解析、提示詞與遷移建議，涵蓋 macOS、Linux、Windows。原先打包的 GPT-5.4 選項與內部用途，也改對應到 GPT-5.6 Terra 與 Luna。這像公司更新產品型錄與貨架標籤：你依舊在挑合適的工具，但名稱與建議跟上新一代。

Long conversations no longer have to feel like one endless scroll. Experimental paginated history lets you resume efficiently, search, keep names, and work with sub-agents and memories. Think of a loose stack of meeting notes moved into an indexed binder: find the page, then continue from it. Experimental means it may still change.

Moving to Codex no longer means rebuilding every old habit. /import can migrate Cursor and Claude Code settings, MCP servers, plugins, sessions, commands, and project-scoped memories. It is more like moving a whole phone than just its contacts. Run /import, then choose the source and what to bring over; compatibility is not promised for every setting.

Teams using Amazon Bedrock now get experimental login, custom endpoint, and authentication support, with GPT-5.6 Sol as the default Bedrock model. It is like adding a badge entrance and allowing a company’s own phone line into the building. This matters mainly when Codex must fit an existing Bedrock setup.

When the source is a recording rather than text, Codex can now take common local audio formats as input and return audio from tools. Streaming realtime V3 conversations also arrive: replies begin traveling while they are generated, like a phone call instead of a finished voicemail. Exact formats and commands are not specified here.

For work split across several agents, the opt-in multi-agent V2 experience is now more stable. You can configure sub-agent models, reasoning levels, and concurrency; roles return and navigation improves. It is like a director choosing each crew member’s gear and number of simultaneous tasks. Opt-in means it must be enabled deliberately.

Visualizations in terminal output can now include secure, clickable inline links. Instead of copying an address and guessing where to paste it, you can open the full view from the result itself. The release only promises secure clickable links, not universal support for every terminal or visualization.

Editing an earlier prompt or retrying a safety-buffered turn now makes a contextual branch. The original conversation, attachments, and mention bindings stay intact. It is like photocopying a manuscript before editing: the new draft can diverge while the original and its materials remain available.

When many external tools start at once, MCP sign-ins and credential refreshes can collide. Startup timeouts, non-blocking OAuth discovery, serialized refreshes, and safe reuse of tool catalogs prevent slow or conflicting flows. Think of an airport that keeps one delayed passenger from blocking every gate and reuses one flight board instead of printing it repeatedly.

Windows execution and sandbox reliability are strengthened with native exec-server sandboxing, enforced network-proxy rules, hidden helper consoles, and correctly quoted hook commands. It is like fixing a worksite’s fence, entrance, and radio protocol together: invisible details that keep the job inside its boundaries and running as intended.

Safety handling now detects forced rm deletion more carefully, confirms full access more consistently, and preserves rejection reasons across tools. It is like a guard who notices not only a key but also an attempt to remove the whole door, then hands your refusal note to the next guard. This strengthens protection; it does not loosen deletion rights.

Bundled OpenAI Docs guidance now covers GPT-5.6 model resolution, prompting, and migration across macOS, Linux, and Windows. Packaged GPT-5.4 selections and internal uses also move to the corresponding GPT-5.6 Terra and Luna variants. It is like updating both the catalog and shelf labels so model choices reflect the current generation.