發布日期：2026-09-01
你在終端機切換模型或輸入指令時，操作變得更加乾淨俐落。
現在輸入 /model <name> 時，只要按下 Tab 鍵就能直接採納灰色的模型名稱建議文字；執行帶有子指令的命令（如 models 或 agents）時，不再因繼承未關閉的標準輸入管線卡死，而是立即執行。同時，程式能妥善攔截提示詞掛鉤（hook，特定事件發生時被觸發的小程式）引發的異常，並拒絕模型請求中的空完成設定，防止命令列意外當機；輸入框在打字並清空後，引導模式的佔位文字也不會再莫名跳出干擾。
剛登入帳號或進行身分驗證時，現在不會再遇到提示詞被吞掉或對話壞掉的問題。先前若 Google Cloud 認證與新手引導工作階段未設定完成或意外中斷，畫面會直接掉進故障的對話視窗，現在則會正常顯示登入介面。針對電腦時鐘偏差造成的偶發認證錯誤，瀏覽器與 WIF 的 OAuth 權杖（token，驗證身分的授權令牌）現在會在到期前 5 分鐘主動提前換發。此外，剛完成登入就送出的提示詞不會再因帳號驗證提示而被丟棄，而是會自動排隊等待背景資格驗證完成後依序執行。
指派多個子代理（subagent，負責分工處理特定子任務的輔助助理）處理複雜工作時，整體執行變得更輕快可靠。
系統現在改為每個子軌跡僅發送一次軌跡詮釋資料，不再每一步都發送，大幅降低串流傳輸的額外負擔；針對設定了 enable_mcp_tools=true 的子代理，執行器現在能正確取得 MCP（讓 AI 連接外部工具的協定）分派器中繼工具，排除找不到工具的錯誤；而在上層對話中被取消或終止的子代理，工作抽屜介面也會即時更新，不再卡在執行中的狀態。
使用各類擴充工具時，現在確認權限與讀取設定都更加清楚精確。當發起直接工具或 MCP（讓 AI 連接外部工具的協定）工具呼叫時，彈出的權限確認提示會直接顯示宣告的人類可讀動作描述，不再只印出模糊的通用標題。在為 Gemini 模型重組請求歷史紀錄時，系統已修正工具呼叫與結果遺漏識別碼的問題，確保上下文能精準對照。此外，在 Windows 系統中若把 MCP 的 JSON 設定檔儲存為帶有 UTF-8 BOM 標記的格式，現在也能正常讀取而不報錯。
When switching models or entering commands in the terminal, your workflow is now cleaner and smoother.
Now when entering /model <name>, you can press the Tab key to directly accept the grey ghost text suggesting the model name; executing commands with subcommands (such as models or agents) no longer hangs on an inherited, unclosed standard input pipe and runs immediately. At the same time, the program properly catches exceptions raised by prompt hooks (helper routines triggered by specific events) and rejects nil completion configurations in model requests to prevent accidental CLI crashes; placeholders in instructional cycle mode will also no longer unexpectedly reappear after typing and clearing the input prompt.
When signing into an account or verifying identity, you will no longer encounter dropped prompts or broken conversations. Previously, if Google Cloud authentication and onboarding sessions were unconfigured or interrupted, the interface would drop into a broken chat session; now it properly displays the sign-in screen. To prevent transient authentication errors caused by computer clock skew, browser and WIF OAuth tokens (authorization credentials used to verify identity) are now proactively refreshed 5 minutes before expiration. In addition, prompts submitted immediately after login will no longer be discarded due to account verification alerts, but are queued until background eligibility verification completes.
When assigning multiple subagents (auxiliary assistants that handle specific subtasks) to tackle complex work, overall execution is now faster and more dependable.
The runner now sends subtrajectory metadata once per subtrajectory instead of at every step, significantly reducing streaming overhead; for subagents configured with enable_mcp_tools=true, the executor can now properly access the MCP (protocol connecting AI to external tools) dispatcher meta-tool, eliminating unknown tool errors; and for subagents cancelled or killed in ancestor conversations, the task drawer interface updates promptly rather than remaining stuck in the running state.
When using various extension tools, reviewing permissions and loading configurations is now clearer and more precise. When initiating direct or MCP (protocol connecting AI to external tools) tool calls, the permission confirmation prompt now directly shows declared human-readable action descriptions instead of displaying generic prompt titles. When reconstructing request history for Gemini models, the missing tool-call IDs in tool invocations and results have been fixed to ensure precise context matching. Additionally, saving MCP JSON configuration files with a UTF-8 byte order mark (BOM) on Windows is now parsed properly without errors.
Improved `/model <name>` autocompletion to accept the proposed model name ghost text with `Tab`.
Fixed commands with subcommands (such as `models` or `agents`) hanging on an inherited, unclosed standard input pipe instead of executing immediately.
Fixed CLI crashes caused by prompt hooks by catching hook panics and rejecting nil completion configurations in model requests.
Fixed instructional cycle mode placeholders reappearing in the input prompt after typing and clearing content.Fixed unconfigured or interrupted Google Cloud authentication and onboarding sessions dropping into broken chat sessions instead of prompting with the sign-in screen.
Fixed transient authentication errors caused by token expiry clock skew by proactively refreshing browser and WIF OAuth tokens five minutes before expiration.
Fixed prompts entered immediately after login being discarded with an account verification alert by queueing them until background eligibility verification completes.Reduced subagent streaming overhead by sending subagent trajectory metadata once per subtrajectory instead of with every step.
Fixed subagents defined with `enable_mcp_tools=true` failing with unknown tool errors by ensuring the MCP dispatcher meta-tool is available to the subagent executor.
Fixed cancelled or killed subagents remaining stuck in the "Running" state in the UI task drawer across ancestor conversations.Fixed tool invocations and results omitting tool-call IDs when reconstructing request history for Gemini models.
Fixed tool permission prompts for direct and MCP tool calls displaying generic prompt titles instead of their declared human-readable action descriptions.
Fixed parsing MCP JSON configuration files on Windows when saved with a UTF-8 byte order mark (BOM).