發布日期：2026-07-01 ~ 07-08
為了讓你隨時掌握背景任務的動向，系統在輸入框正下方設計了全新的互動式狀態列。它可以即時顯示 active subagents（在背景幫你分析程式碼或寫報告的子助理 AI）與 background tasks（丟給電腦在背後慢慢執行的耗時指令，例如跑測試或編譯）的即時動態，讓你在輸入指令時一眼看清背後是誰在忙、有沒有卡住，不需手動打命令查詢。這就像你在辦公室裝了透明玻璃，抬頭就能看到小助手們是不是在賣力打字、測試軟體，不用每隔五分鐘就跑去敲門問進度。
為了省去你反覆手動滾動日誌畫面的時間，/tasks 詳情視窗現在會配合日誌流更新自動下捲。每當新的背景任務 log（程式執行時印出的詳細過程與輸出文字）吐出時，視窗會自動捲到底部呈現最新一行；若你手動向上滾動檢視舊內容，自動下捲就會貼心暫停，絕不強行拉回底部打斷閱讀。這就像你用手機通訊軟體聊天，畫面會隨著對方的回信自動往下推進、顯示最新訊息；但當你想往上翻舊對話時，畫面也不會調皮地一直跳回最底下。
為了讓 Windows 使用者能順暢貼上剪貼簿中的截圖，系統新增了專屬貼上快捷鍵。如果你的 Ctrl+V 被終端機軟體（如命令提示字元或 PowerShell 視窗）攔截而無法貼上，現在可以直接改按 Alt+V，即可順利將圖片或文字貼進輸入框，不再因鍵位衝突而卡關。這就像是大樓的旋轉門（Ctrl+V）因為安全設定把你的大件行李（圖片）卡在門外，現在保全在旁邊幫你開了一扇專屬的側門（Alt+V），讓你提著行李也能輕鬆通過。
為了讓你能直接使用最習慣的開發工具修改 AI 產出的內容，現在在成果畫面可以直接呼叫外部編輯器。當你在檢視 AI 寫好的 artifacts（AI 產出的程式碼或完整文件檔案）時，按一下 Ctrl+G 就會以你本機設定的 $EDITOR（你的預設文字編輯器，比如 VS Code 或 Vim）開啟該檔案；若當前有未送出的草稿備註，系統也會貼心詢問是否保留。這就像你原本用公用電腦看文件，現在按一下按鈕，這份文件就直接拷貝一份用你自己的筆記本電腦打開，還順便幫你帶上剛剛寫了一半的草稿紙。
為了防止網路波動打斷你的思考，系統提升了模型生成時的容錯能力。當 AI 回答遇到 transient errors（因為網路不穩暫時斷線，或是雲端伺服器瞬間過載造成的暫時性錯誤）時，系統會自動在背後嘗試重新連線並接續發送，不會像以前一樣直接噴錯中斷，逼你把指令重新打一遍。這就像你打電話給客服時突然訊號不好斷線，系統在背景幫你自動重撥，接通了直接讓你跟原本的客服人員繼續聊，不用重新撥號排隊。
為了能安心執行所有類型的終端指令，背景任務現在能正確處理完全沒有內容的輸出。以前當你執行的 background tasks（丟到背景悄悄執行的任務，例如延遲幾秒的 sleep 指令）沒有印出任何文字時，系統會因為讀到 empty outputs（完全沒有任何文字的空輸出結果）而觸發崩潰錯誤；現在已修妥此問題，即使指令靜悄悄地執行完畢也絕不當機。這就像你以前叫保鏢去辦事，如果他沒拿回任何文件，你就認定他出了意外而直接抓狂；現在你懂了「沒拿回文件也是一種正常辦完的結果」，大家相安無事。
為了讓你能量身訂做特殊助手，臨時建立客製化助理的流程已被完全修復。以前你使用 define_subagent 臨時在系統中產生一個具有特定指示與工具權限的 dynamically defined subagents（你臨時為特殊工作量身設定、具有獨特指示與工具權限的小助手）時，會因為內部以脆弱的 JSON 格式傳遞而導致無法被順利喚醒；現在系統已全面改用 Markdown 格式，你的客製化助理可以穩定就位、隨喚隨到。這就像你原本用繁雜晦澀的手寫表格（JSON）去申請臨時雇員，表格填錯一個空格就被駁回；現在改成用一般白話紙條（Markdown）交代，助理一看就懂，馬上就能開工。
A new interactive status line has been added directly under the input prompt. It reports on active subagents (helper AIs spawned in the background to analyze code or write docs) and background tasks (long-running processes like code compilation or test suites running behind the scenes) in real-time, allowing you to see if anything is stuck without typing queries. It's like replacing solid walls with glass windows in an office; now you can see if your assistants are typing away, without having to knock on their doors every five minutes.
To save you from constantly fighting the scrollbar to see progress, the /tasks detail panel now auto-scrolls down in sync with log streams. As new background task log output (log, text streams outputted by programs as they run to describe their progress) arrives, the panel scrolls to show the latest lines, but pauses if you manually scroll up to inspect earlier entries. It's like typing in a chat app where the view scrolls to keep up with new incoming messages, but stays put if you scroll up to reread what was said earlier.
App features an alternative paste hotkey. If your standard Ctrl+V is hijacked by your terminal emulator (the shell window programs like CMD or PowerShell), you can press Alt+V instead to insert clipboard images or text directly, avoiding shortcut conflicts. It's like a revolving entrance door (Ctrl+V) that keeps locking because your large luggage (images) gets stuck; now security has opened a side door (Alt+V) so you can carry your luggage straight in.
You can now open files directly in your preferred editor. While viewing generated artifacts (artifacts, documents or files of code generated by the AI for you to read), pressing Ctrl+G opens the file in your configured $EDITOR (your default text editor, such as VS Code, Vim, or Notepad), checking to save any unsent comment drafts first. It's like viewing a document on a library tablet and pressing a button that instantly transfers the file to your personal laptop, along with the half-written sticky notes you had placed on it.
Response generation is now more resilient. When the AI model runs into transient errors (transient errors, temporary failures caused by unstable Wi-Fi disconnects or brief cloud server overloads), the app automatically handles client-side retries behind the scenes, rather than aborting and making you resubmit your prompt. It's like being on a support call when your signal drops; instead of hanging up and making you dial and queue again, the phone system automatically redials in the background and reconnects you to the same agent.
Background executors now handle silent processes properly. Previously, if background tasks (tasks sent to run silently in the background, like a sleep command that pauses for a few seconds) yielded empty outputs (empty outputs, console command outputs containing absolutely no text), the internal reader panicked and crashed; this is now fixed so silent finishes close cleanly. It's like hiring a messenger and panicking that he was kidnapped because he came back with no letters; now you understand that returning empty-handed is a perfectly normal result.
Custom agent creations are now fully functional. Spawning custom helper agents on the fly via define_subagent (dynamically defined subagents, custom AIs created on the fly with specialized prompts and tool permissions for unique tasks) used to fail due to fragile JSON format checks; the app now maps these in Markdown, allowing custom helpers to wake up and run immediately. It's like filling out a complex, poorly designed form (JSON) to hire a temp worker where one wrong checkmark rejects the request; now you just write a simple sticky note (Markdown) that they read and start working from immediately.
Introduced a new interactive status indicator below the input box that displays active subagents and background tasks in real-time.Improved the /tasks detail panel to automatically scroll to the bottom as new background task logs stream in, and default to the latest output when opened while preserving scroll position if scrolled up manually.Added alt+v as an alternative paste shortcut on Windows to resolve issues where ctrl+v is intercepted by the terminal emulator, enabling reliable image pasting.Added ctrl+g on the artifact view to open $EDITOR, with a warning confirmation prompt before opening the editor in the artifact detail view if there are unsent comments.Improved model generation resilience by adding automatic client-side retries when encountering transient errors.Fixed a crash occurring when executing background tasks or terminal commands that produce empty outputs (such as sleep).Fixed dynamically defined subagents failing to invoke by transitioning definitions from JSON to Markdown format.