Codex 0.160.0

當你讓 AI 執行跨步驟任務或在不同代理人間輪流協作時，最怕後續接手的環節漏掉最初的叮嚀。新加入的 Guardian 審查（Guardian review，檢驗操作是否符合安全規範的防護機制）是一項自由啟用的功能。它現在能主動擷取使用者在對話更早之前給出的指令，並將代理人交接時的完整脈絡一併納入檢查。

過去交接時若只傳遞當前進度，防護檢查可能因缺乏前因後果而做出誤判。現在只要開啟這項功能，審查模組就能掌握全貌，讓多代理人協作時的執行方向始終對齊你的原始要求。

有時候你只是想臨時發問或處理零散檔案，手邊的目錄並不是任何專案資料夾。現在只要系統政策允許，你可以在專案目錄之外直接啟動工作階段（session，你與 AI 連續互動的對話環境），並套用預設的工作區設定。不僅如此，當你之後想要接續處理並恢復工作階段時，系統也會一併還原先前儲存的授權與權限。

以前不在專案內可能無法正常啟用，中斷後重新連線還得重新授權一次。現在即使脫離專案目錄，操作彈性與權限記憶依然完整。

在終端機裡頻繁切換任務或檢視紀錄時，操作流暢度會直接影響工作節奏。代理人控制中心現在新增了可用鍵盤直接操作的「Show more」動作，讓你輕鬆翻查更早之前的任務紀錄。在支援此操作的本機 Linux X11 終端機上，全螢幕模式下也已支援選取對話紀錄文字並透過滑鼠中鍵快速貼上。在設定與狀態保存上，終端機介面現在會完整保留伺服器提供者、推論摘要與詳細程度等偏好，並在恢復對話或分支出新歷史時顯示正確的工作階段。

各項介面細節修整後，日常操作更加順手。

當你在網路不穩的環境下對話，最怕連線中斷時排隊等著送出的訊息被重複發送。現在佇列（queue，排隊依序等待處理的待辦清單）中的未發送訊息有了更完善的重連機制。當網路斷線重新連上時，系統會先核對並確認先前處於狀態不明的提交是否已經成功，釐清之後才繼續處理後續訊息，徹底避免同一句話被重複發送。

重連後會繼續處理未送出的佇列訊息，避免重複發送。

當你讓 AI 分派子任務給多個子代理人處理時，如果執行環境還在準備，過去容易出現狀態追蹤不明的情況。子代理人（subagent，由主程式派發去獨立執行特定子任務的輔助代理人）現在會妥善保留仍在啟動中的執行環境。如果在設定或準備階段發生任何錯誤，子代理人也能確實接收到具體的失敗原因。

這讓多代理人協作的初始化流程更加穩健，遇到問題時也能立即知道卡在何處。

在長時間執行或頻繁寫入日誌時，底層資料庫的穩定度與磁碟空間管理十分關鍵。這版防止了 SQLite（一種輕量級的嵌入式資料庫系統）在建立連線與寫入記錄時發生卡頓的情況。如果資料庫在初始化時發生錯誤，系統會直接顯示真正的錯誤訊息，不再將其模糊地掩蓋成連線超時。同時，系統新增了背景自動回收機制，會自動釋放日誌資料庫中未使用的剩餘空間，減少硬碟容量佔用。

當你連接外部自訂模型或載入多個外掛擴充時，設定的明確度與載入效率直接影響啟動速度。明確指定的模型提供者清單現在不再包含不支援的內建項目，重新整理失敗時也不會再沿用過期資料。官方也進一步釐清了提供者憑證如何搭配儲存後端運作，並說明如何利用 env_key 設定來指定 API 金鑰所屬的環境變數。在外掛方面，系統透過快取（把算過或查過的結果暫存起來直接用）已解析的資訊清單，並在發送遠端請求時重複使用既有的 HTTP 連線，減少了重複載入外掛的時間。

整體整合更乾淨，運作更俐落。

When you have AI execute multi-step tasks or rotate collaboration across different agents, the biggest concern is that subsequent stages might overlook initial instructions. The newly added Guardian review (an opt-in safety mechanism that verifies whether operations comply with security rules) is an optional feature. It can now proactively retrieve user instructions provided earlier in the conversation, incorporating full agent handover context into its checks.

In the past, passing only the current progress during handovers could lead to misjudgments in safety checks due to missing background context. Once enabled, the review module captures the complete picture, ensuring that multi-agent execution remains aligned with your original requirements.

Sometimes you simply want to ask quick questions or handle miscellaneous files when the current directory is not a project folder. As long as system policy permits, you can now launch a session (a conversational environment for continuous interaction with AI) directly outside project directories while applying default workspace settings. Furthermore, when you resume the session later to continue work, the system also restores previously stored authorizations and permissions.

Previously, starting outside a project might fail, and reconnecting required re-authorizing permissions. Now, even outside project directories, operational flexibility and permission persistence remain intact.

When frequently switching tasks or inspecting history in the terminal, operational smoothness directly affects your workflow pace. The Agent Control Center now introduces a keyboard-accessible "Show more" action, letting you easily browse earlier task records. On local Linux X11 terminals that support this operation, full-screen mode now also allows selecting conversation transcript text and pasting quickly with a middle click. For settings and state persistence, the terminal interface fully preserves preferences such as server provider, reasoning summaries, and verbosity, displaying the correct session when resuming conversations or branching into new history.

With these refined interface details, daily operations become much smoother.

When conversing over an unstable network, the biggest concern is that pending messages waiting to be sent might be duplicated during disconnections. Unsent messages in the queue (a pending list waiting to be processed in order) now benefit from an improved reconnection mechanism. Upon reconnecting after network drops, the system first verifies whether earlier submissions in an uncertain state succeeded before continuing with subsequent messages, completely preventing the same message from being sent twice.

After reconnecting, the system resumes processing unsent queued messages without duplication.

When you have AI delegate subtasks to multiple subagents, tracking could become unclear if the execution environment was still being prepared. A subagent (a helper agent dispatched by the main program to run specific subtasks independently) now properly retains execution environments that are still starting up. If any errors occur during configuration or setup, the subagent accurately receives the concrete reason for failure.

This makes multi-agent initialization much more reliable, allowing you to instantly see where things stalled when issues arise.

During long runs or frequent logging, underlying database stability and disk space management are important. This release prevents SQLite (a lightweight embedded database system) from experiencing jank when establishing connections and writing records. If an error occurs during database initialization, the system displays the genuine error message directly instead of masking it as a connection timeout. Meanwhile, a background automatic reclamation mechanism has been added to free unused space in the log database, reducing disk consumption.

When connecting external custom models or loading multiple plugin extensions, configuration clarity and loading efficiency directly impact startup speed. The explicitly specified model provider list no longer contains unsupported built-in entries, nor does it reuse stale data when refreshes fail. The update further clarifies how provider credentials work with storage backends and explains using the env_key setting to designate the environment variable for API keys. For plugins, the system caches parsed manifests and reuses existing HTTP connections during remote requests, reducing plugin reloading time.

Overall integration is cleaner and operates more smoothly.