發布日期：2026-09-10
出門在外想從遠端連回工作主機，最怕主機登出或重開機後連線中斷。
工具新增了 remote-control start、remote-control status 與 remote-control stop 子指令，可將命令列工具註冊為作業系統背景服務（常駐在系統底層運行的程序）。即使登出帳號或重新開機，主機依然保持可連線狀態。執行 remote-control start --name <label> 能在實體清單自訂辨識標籤，加入 --session 旗標則能將服務限制在當前登入的工作階段。
# 註冊為作業系統常駐服務，重開機或登出皆保持連線
agy remote-control start --name "office-mac"
# 僅限定在目前登入的工作階段
agy remote-control start --session
# 查詢服務狀態與停止服務
agy remote-control status
agy remote-control stop在終端機中瀏覽落落長的回覆或比對異動時，畫面移動更加俐落。所有可捲動檢視全面改善了半頁捲動體驗，包含備用螢幕模式（altscreen，終端機獨立覆蓋主畫面的全螢幕互動介面）、差異比對檢視器與各類唯讀詳細資訊面板。在唯讀面板中按下 ctrl+d 快捷鍵，現在只會專心向下翻動半頁，不再誤觸退出確認對話框。同時新增了 shift+up 與 shift+down 作為預設的半頁導覽快捷鍵，分別對應 navigation.half_page_up 與 navigation.half_page_down 設定鍵，使用者也能依個人習慣自由重新綁定。
{
"navigation": {
"half_page_up": "shift+up",
"half_page_down": "shift+down"
}
}當你輸入的提示詞或模型產生的回覆觸動內容安全防線時，終端機不再用含糊的錯誤訊息讓人摸不著頭緒。
以前只要提示或生成內容被內容安全過濾器（防止有害或違規內容的審查機制）阻擋，系統常會跳出無關的 no candidate found 假錯誤、整輪對話靜默結束不留文字，甚至反覆重試被拒絕的輸入。現在命令列工具會直接顯示明確的 content-filter 停止原因，讓使用者一眼看清是哪裡觸發了安全審查。
讓代理人寫臨時程式碼或運算草稿時，你的工作面板與版本紀錄不再被暫存檔案洗版。當 AI 代理人在名為 scratch/ 的草稿暫存資料夾中建立臨時工作檔案時，過去會意外觸發檔案系統的遞迴監聽機制（持續追蹤目錄內所有子檔案變動的背景監控）。這導致許多僅用於中途運算的臨時檔案，全被當成正式成果塞進成品審查面板（審視程式碼產出物的檢視介面）。存檔檢查點更因此留下一堆無關的雜訊紀錄。這個問題已經修好，草稿檔不再灌進成品面板與檢查點。
擴充外部工具時，伺服器的啟動連線與工具檢驗現在更加可靠。使用 MCP（Model Context Protocol，讓 AI 連接外部工具的協定）擴充功能時，過去安裝在全域外掛中的隨附伺服器常在命令列工具啟動時初始化失敗，在外掛啟用或停用時也無法即時同步更新運行狀態。此外，部分第三方工具的輸入規格若未明確寫出 additionalProperties 設定，在舊版架構驗證時會被判定不合規而無法呼叫。如今這兩項瑕疵皆已修正，全域外掛隨附的伺服器能如期啟動並隨開關切換狀態，省略該欄位的第三方工具也能正常通過檢驗。
平常在對話框裡敲入日常問題，你現在能更快看見 AI 開始作答。
過去只要使用者送出一則訊息，即使完全沒有輸入 / 斜線開頭的特殊指令，系統依然會在磁碟中進行客製化探索遍歷（在檔案系統逐層搜尋技能與規則設定檔的掃描程序）。這項多餘的搜尋每次都平白耗費時間，拖慢了代理人動手回應的速度。現在系統只在真正呼叫指令時才進行搜尋，有效降低了代理人開始作答前的等待延遲。
查看記錄檔或排查系統狀態時，滿螢幕的假警報不再干擾你的判斷。檢視 cli.log 記錄檔或執行背景守護程序（daemon，在系統背景運作的常駐服務）的子指令輸出時，終端機與記錄檔現在恢復了乾淨面貌。過去底層程式庫在完成初始化之前，每一行輸出的日誌文字開頭都會被強制加上 ERROR: logging before google.Init 這行虛假錯誤字串。這把正常的運作紀錄全染成刺眼的錯誤訊息，嚴重誤導除錯方向。現在這行假錯誤已經不再出現。
在不同平台與帳號環境使用 Remote Control 背景服務時，各項運作細節現在更加穩健。
在 macOS 上由系統服務管理員啟動 remote-control 背景服務時，過去常在啟動後不久便異常終止，現在能穩定常駐運行。同時修正了執行 remote-control start 與 remote-control stop 時，仍會殘留舊版安裝腳本所啟動的背景服務之問題，確保背景服務乾淨交接。此外，若使用者是透過付費 Google Cloud 專案取得權限，先前無法開啟 Remote Control 設定的問題也已解決，現在能順暢啟用功能。
想接續幾天前或更新前留下的對話記錄時，對話歷程不再因為格式不認得而拒絕載入。以前使用恢復對話功能時，若該筆歷史紀錄是在較早版本中建立，系統常因無法辨識記錄檔裡的舊版步驟格式，而跳出 unknown step type 錯誤中斷載入，導致整段對話完全無法開啟。現在載入解析器強化了對舊格式的向下相容能力，即使對話中包含過去定義的步驟型別，也能順利讀取並完整還原對話現場。
When connecting back to your workstation remotely while away, the biggest headache is losing the connection after logging out or rebooting.
The tool has added the remote-control start, remote-control status, and remote-control stop subcommands, allowing you to register the CLI as an operating system background service. Even if you log out of your account or restart the machine, the host remains reachable. Running remote-control start --name <label> lets you set a custom identification label in the instance list, while adding the --session flag restricts the service to your current login session.
# Register as an OS background service to stay reachable across reboots and logouts
agy remote-control start --name "office-mac"
# Scope strictly to the current login session
agy remote-control start --session
# Check service status and stop the service
agy remote-control status
agy remote-control stopNavigating lengthy responses or inspecting diffs in the terminal is now much smoother. Half-page scrolling has been improved across all scrollable views, including altscreen mode, the diff viewer, and various read-only detail panels. Pressing ctrl+d in read-only panels now strictly scrolls down half a page instead of accidentally triggering the exit confirmation prompt. Additionally, shift+up and shift+down have been introduced as customizable default keybindings for half-page navigation, mapping to navigation.half_page_up and navigation.half_page_down respectively.
{
"navigation": {
"half_page_up": "shift+up",
"half_page_down": "shift+down"
}
}When an input prompt or model-generated response triggers safety guardrails, the terminal no longer leaves you baffled with vague errors.
Previously, when prompts or generated content were blocked by content safety filters, the system often threw spurious no candidate found errors, ended the entire turn silently with an empty response, or repeatedly retried rejected input. The CLI now surfaces an explicit content-filter stop reason, letting you clearly see what triggered the safety review.
When an agent drafts scratchpad code or intermediate computations, your workspace panels and version history will no longer be cluttered with temporary files. Previously, when an AI agent wrote temporary working files to the scratch/ directory, it accidentally triggered recursive filesystem watchers. As a result, intermediate computational files were treated as finished deliverables and flooded into the artifact review panel. Checkpoints were also polluted with noisy, irrelevant entries. This is now fixed, and scratch files no longer show up in the artifact panel or checkpoints.
Expanding external tools is now much more reliable for both server startup and tool schema validation. When using MCP extensions, bundled servers inside globally installed plugins previously failed to initialize at CLI startup and failed to update their running status when plugins were enabled or disabled. Additionally, when third-party tool input schemas omitted the additionalProperties setting, legacy schema validation rejected them as non-compliant and blocked tool invocation. Both issues have now been resolved: bundled servers in global plugins initialize as expected while syncing state changes, and third-party tools omitting that field pass validation normally.
When typing ordinary questions into the chat box, you will now see the AI begin responding much faster.
Previously, whenever a user sent a message—even one without any / slash command—the system still performed filesystem customization discovery walks. This redundant search wasted time on every turn, adding noticeable delay before the agent began answering. The system now performs disk searches only when a command is actually invoked, effectively reducing response startup latency.
When inspecting log files or troubleshooting system status, screens full of false alarms will no longer mislead your investigations. Viewing cli.log or running background daemon subcommands now presents clean output in both logs and the terminal. Previously, before the underlying library completed initialization, every logged line was forcefully prefixed with a spurious ERROR: logging before google.Init string. This painted normal runtime output as alarming error messages and severely derailed debugging efforts. That spurious prefix is now gone.
Operating the Remote Control background service across different platforms and account environments is now much better behaved.
When launched by the system service manager on macOS, the remote-control background service previously terminated shortly after startup; it now runs stably in the background. Running remote-control start and remote-control stop no longer leaves background services running from legacy installer scripts, ensuring clean service handoffs. Additionally, for users whose access comes from a paid Google Cloud project, the previously unavailable Remote Control setting has been fixed, allowing straightforward feature activation.
When picking up past conversations from days ago or before an update, your chat history will no longer fail to load due to unrecognized formats. Previously, when resuming a conversation created in an older version, the system often failed with an unknown step type error because it could not recognize legacy step formats in the log, making the entire conversation unopenable. The loader parser now features improved backward compatibility for older formats, smoothly parsing legacy step types and fully restoring the conversation session.
Added the `remote-control start`, `remote-control status`, and `remote-control stop` subcommands to run the CLI as a background service registered with your operating system's service manager so your machine stays reachable across logouts and reboots; `remote-control start --name <label>` sets a custom machine label in the Remote Control instance list, and `--session` scopes the service to your active login session.Improved half-page scrolling across all scrollable views — including altscreen mode, the diff viewer, and read-only detail panels where `ctrl+d` no longer triggers the exit prompt — and set `shift+up` and `shift+down` as customizable default keybindings for half-page navigation (`navigation.half_page_up` and `navigation.half_page_down`).Fixed prompts or model responses blocked by content safety filters failing with a spurious "no candidate found" error, ending silently with an empty turn, or repeatedly retrying the rejected input; the CLI now surfaces a clear content-filter stop reason.Fixed temporary working files written to the agent `scratch/` directory triggering recursive filesystem watchers and appearing as noisy entries in the artifact review panel and checkpoints.Fixed MCP servers bundled inside globally installed plugins failing to initialize at CLI startup or failing to update their running status when plugins are enabled or disabled.
Fixed third-party MCP server tools failing schema validation when their input schemas omit `additionalProperties`.Fixed unnecessary filesystem customization discovery walks running on every user message even when no slash command is invoked, reducing latency before the agent begins responding.Fixed every log line in `cli.log` and background daemon subcommand output being prefixed with a spurious `ERROR: logging before google.Init` message.Fixed the macOS `remote-control` background service terminating shortly after startup when launched by the system service manager, `remote-control start` and `remote-control stop` leaving legacy installer-script background services running, and the Remote Control setting being unavailable when your access comes from a paid Google Cloud project.Fixed older conversations failing to load with an `unknown step type` error when resumed.