輸入提示詞可用 @ 指名子代理,自動補齊快速對話
當你同時開了好幾個背景工作幫手,想要單獨插話交辦新任務,現在只要在輸入框打個 @ 符號就能辦到。現在新增了 @<subagent> <message> 提示詞語法,能直接把訊息發給指定的子代理(subagent,在背景獨立處理特定任務的助理)。打字時只要輸入 @ 就會跳出自動補齊選單,清單會列出目前正在執行以及已經完成的子代理名稱。例如研究助理剛查完資料,你直接打 @research 幫我整理成三點結論 就能繼續追問,不必手動切換對話。
發布日期:2026-09-23
當你同時開了好幾個背景工作幫手,想要單獨插話交辦新任務,現在只要在輸入框打個 @ 符號就能辦到。現在新增了 @<subagent> <message> 提示詞語法,能直接把訊息發給指定的子代理(subagent,在背景獨立處理特定任務的助理)。打字時只要輸入 @ 就會跳出自動補齊選單,清單會列出目前正在執行以及已經完成的子代理名稱。例如研究助理剛查完資料,你直接打 @research 幫我整理成三點結論 就能繼續追問,不必手動切換對話。
習慣在終端機裡用 Vim 快速修整文字的讀者,現在不用再狂按同一組快捷鍵了。編輯器現在支援數字倍數乘數,在一般模式(Normal mode,游標移動與執行指令的狀態)以及選取模式中,你可以在動作、移動與運算子前面加上次數。常用的整行刪除、貼上、復原、游標跳轉到指定行或選取多層括號區塊,都能直接加倍執行:
3dw # 刪除接下來的 3 個單字
2d3w # 執行 2 次刪除 3 個單字
3dd # 刪除 3 行
3x # 刪除 3 個字元
3rX # 將接下來 3 個字元替換為 X
3p # 連續貼上 3 次
3u # 連續復原 3 次動作
5G # 移動至第 5 行(支援 [count]G、gg 與 $)
2di( # 刪除外層第 2 層括號內的文字物件用腳本串接命令列工具來自動跑任務時,最怕背景工作沒跑完就被砍掉、或是執行結束後卡住不放。在以 -p 或 --prompt 執行的無頭模式(headless,沒有終端機畫面的自動執行模式)下,以前程式退出後常殘留背景守護程序。這會導致讀取輸出的外部腳本一直等不到結尾訊號而卡死,現在常駐程序會在執行結束時確實終止。另外,以前助理只要閒置約 5 秒,就會誤把還在跑的背景任務直接取消;現在會持續等待背景任務完成,最長可等到 --print-timeout 指定的期限,上限為 30 分鐘。
當你同時開啟多個終端機視窗、或是進行超長篇對話時,底層資料庫與連線都不再無預警出錯。以前多個處理程序同時存取同一個對話歷史資料庫時,檢查寫入權限的動作可能會悄悄丟失檔案鎖(file lock,防止多個程式同時寫入同一檔案的鎖定機制),造成資料庫損壞風險。當對話過長觸發上下文壓縮(context compaction,自動濃縮歷史以節省空間的機制)時,部分情境下用於建立檢查點的工具設定曾遭到拒絕而導致壓縮失敗。模型串流回傳資料時,要是封包缺少外層回應信封,先前會觸發後端當機並以連線錯誤終止工作階段。現在這些資料庫鎖定、壓縮檢查點與串流封包解析問題都已修復。
在終端機裡查看模型產出的表格與程式碼清單時,畫面排版現在變得整齊乾淨。以前表格儲存格如果包含換行折疊的檔案連結,整張 Markdown 表格的欄位對齊就會歪掉破版;現在即使連結文字跨行折疊,欄位依然能維持精確對齊。另外,指向函式或類別等程式碼符號的檔案連結,過去常遺失自訂顯示文字,直接把長長一串原始檔案路徑印出來;現在修復後會正常顯示原本的描述文字。
在不同登入方式之間切換、或是驗證中途斷線時,帳號狀態處理現在更順手。使用 GEMINI_API_KEY 環境變數啟動的工作階段,過去與一般登入路徑存在些許行為差異,現在已修復並減少兩者的行為落差。另外在進行企業帳號登入時,若在挑選授權或專案之前就關閉瀏覽器視窗,以前會殘留卡住的局部憑證(credential,記錄登入身分的授權資料)而無法重試。現在會自動清除未完成的殘餘憑證,讓你直接乾淨重試登入。
在終端機裡查看生成產物或回頭翻找對話歷程時,多處介面細節變得更順手。當大型圖表超出視窗寬度時,成品檢視器(artifact viewer,瀏覽產出圖表與文件的介面)會在底部顯示向左或向右平移的提示。開啟 /rewind 回溯面板時,各步驟會標上相對時間,並在面板打開時自動聚焦在最新步驟。在指令授權方面,白名單建議現在能精準辨識 jj config 與 jj op 子指令,方便你直接設定永久允許。瀏覽器隨行頁面的分頁標題也修正為大寫縮寫的 Antigravity CLI。這些介面與指令辨識的修整,讓日常瀏覽圖表與回溯操作更俐落。
When running multiple background helpers and wanting to chime in with a new task, you can now do so simply by typing an @ symbol in the input box. The newly added @<subagent> <message> prompt syntax lets you send messages directly to a specific subagent (an assistant handling dedicated tasks in the background). Typing @ brings up an autocomplete menu listing the names of both running and completed subagents. For example, right after a research assistant finishes gathering info, you can type @research summarize this into three key takeaways to follow up without manually switching conversations.
If you are used to editing text quickly with Vim in the terminal, you no longer have to mash the same shortcuts repeatedly. The editor now supports numeric count multipliers. In Normal mode (the state for moving the cursor and running commands) and Visual mode, you can prefix actions, motions, and operators with a count. Common operations like deleting whole lines, pasting, undoing, jumping to a specific line, or selecting nested bracket blocks can now run in multiples:
3dw # Delete the next 3 words
2d3w # Run delete 3 words 2 times
3dd # Delete 3 lines
3x # Delete 3 characters
3rX # Replace the next 3 characters with X
3p # Paste 3 times consecutively
3u # Undo 3 actions consecutively
5G # Jump to line 5 (supports [count]G, gg, and $)
2di( # Delete the inner text object inside the 2nd outer bracket levelWhen scripting command-line tools to automate tasks, you never want background jobs killed prematurely or hanging indefinitely after execution finishes. In headless mode (an automated execution mode without a terminal UI) run with -p or --prompt, background daemon processes previously lingered after exit. This caused external scripts reading CLI output to hang while waiting for an EOF signal, but daemons now terminate cleanly when the run ends. Additionally, when the agent sat idle for about 5 seconds, it previously cancelled still-running background tasks by mistake. Runs now keep waiting for background tasks until the --print-timeout deadline, capped at 30 minutes.
When opening multiple terminal windows at once or having very long conversations, underlying databases and connections no longer fail unexpectedly. When multiple processes accessed the same conversation history database concurrently, checking write permissions previously risked silently dropping file locks (a locking mechanism preventing multiple programs from writing to the same file at once), creating a database corruption risk. When long chats triggered context compaction (a mechanism that summarizes history to save space), tool configurations used for compaction checkpoints were rejected in certain scenarios, causing compaction to fail. When the model streamed response chunks without an outer response envelope, it previously triggered a backend crash that terminated the session with connection errors. These database locking, compaction checkpoint, and stream packet parsing issues are now fixed.
When viewing model-generated tables and code listings in the terminal, the layout is now clean and tidy. Previously, if a table cell contained a file link that wrapped across lines, the column alignment of the entire Markdown table would break. Even when link text wraps across multiple lines, table columns now maintain precise alignment. Additionally, file links pointing to code symbols like functions or classes often dropped their custom display text, rendering a long raw file path instead. This is now fixed so the intended descriptive text displays properly.
Handling account states is now smoother when switching between sign-in methods or disconnecting mid-authentication. Sessions started with the GEMINI_API_KEY environment variable previously had slight behavior discrepancies compared to the standard sign-in path, which are now resolved. Furthermore, during enterprise account sign-ins, closing the browser window before selecting a license or project previously left behind stuck partial credentials (authorization data recording sign-in identity) that blocked retries. Incomplete leftover credentials are now cleared automatically so you can retry sign-in cleanly.
Several UI details are now smoother when reviewing generated artifacts or browsing past conversation history in the terminal. When a wide diagram overflows the window width, the artifact viewer (an interface for browsing generated diagrams and documents) shows a Left/Right pan hint in the footer. Opening the /rewind panel displays relative timestamps for each step and automatically focuses on the most recent step when the panel opens. For command authorization, allow-listing suggestions now accurately recognize jj config and jj op subcommands when offering to always allow a command. The browser companion page title has also been corrected to the capitalized acronym Antigravity CLI. These interface and command recognition refinements make viewing diagrams and rewinding history much crisper.
Added `@<subagent> <message>` prompt syntax to send a message directly to a subagent conversation, with autocomplete listing running and completed subagentsAdded Vim numeric count multipliers so counts apply to operators, motions, and actions in Normal and Visual modes, including `3dw`, `2d3w`, `3dd`, `3x`, `3rX`, `3p`, `3u`, `[count]G`/`gg`/`$`, and counted text objects such as `2di(`Fixed headless (`-p` / `--prompt`) runs leaving daemon background processes running after exit, which could hang scripts reading the CLI's output until end-of-file; daemon processes now terminate when the run ends
Fixed headless (`-p` / `--prompt`) runs cancelling still-running background tasks about 5 seconds after the agent went idle; runs now wait for background tasks until the `--print-timeout` deadline, up to a 30-minute capFixed a conversation-history database corruption risk where checking the database file for write access could silently drop file locks held by concurrent CLI processes on the same file
Fixed context compaction failing when the tool configuration used for compaction checkpoints was rejected in certain scenarios
Fixed a backend crash when a streamed model response chunk arrived without its response envelope, which terminated the session with connection errorsFixed markdown table column alignment when a table cell contains file links that wrap across lines
Fixed markdown file links to code symbols dropping their display text and rendering the raw path insteadImproved `GEMINI_API_KEY` sessions to reduce behavior discrepancies with the non-API-key sign-in path
Fixed incomplete enterprise sign-ins (for example closing the browser window before finishing license or project selection) leaving behind a stuck partial credential; such credentials are now cleared automatically so sign-in can be retried cleanlyImproved the artifact viewer to show a Left/Right pan hint in the footer when a wide diagram overflows the window
Improved `/rewind` to show a relative timestamp for each step and to focus the most recent step when the panel opens
Improved command allow-listing suggestions to recognize `jj config` and `jj op` subcommands when offering to always allow a command
Fixed the browser companion page title to read `Antigravity CLI` instead of `Antigravity Cli`