發布日期：2026-09-12
設計專精特定任務的自訂代理人時，你現在能擁有一張完全乾淨的畫布。在代理人 Markdown 檔案頂端的前言設定（在文件開頭以橫線包夾的設定區塊）加入 excludeDefaultComponents: true，便能選擇不載入系統預設的提示詞區段與內建工具，只留下你親自指定的行為邏輯。這項設定能避免通用指令干擾專業發揮。而任務結束後的掛鉤（在特定事件發生時被觸發的小程式）依舊完整保留，讓自訂自動化流程順暢運作。
---
name: my-specialist
excludeDefaultComponents: true
---面對遠端模型伺服器偶發的網路壅塞或限流，工具現在具備更沉著的容錯韌性。
以前連線時若撞上暫時性的 genai.APIError 錯誤，執行流程往往直接中斷。現在無論是 502、503、504 伺服器異常、每分鐘觸發的 429 速率限制，或是傳輸串流中途中斷，程式都會在內部以指數退避（重試等待時間隨次數倍增的做法）自動重試。重試期間已完成的工具呼叫產出會完整保留，不再白白作廢。若重試後仍未能成功復原，系統會針對 503 與 429 狀況顯示清楚易懂的錯誤提示，不再丟出晦澀難懂的內部代碼。
串接各類外部擴充工具時，規格檢驗不再把合法的彈性輸入擋在門外。使用 MCP（Model Context Protocol，讓 AI 連接外部工具的協定）或各家模型提供者的擴充工具時，舊版結構驗證機制過於嚴格。即使工具規格本身允許開放物件結構描述（schema，規範資料格式與欄位型別的結構定義），例如宣告為 {"type": "object"} 或明確寫出 additionalProperties: true，系統過去仍會拒絕傳入任何未預先宣告的引數。現在檢驗邏輯全面改善，能妥善保留這些開放物件結構，讓合規允許額外參數的工具順利接收資料。
重新載入累積大量歷史紀錄的長篇對話時，終端機的操作手感現在依然輕快。
以前開啟對話紀錄特別長的工作階段（session，從啟動到結束的一整段對話歷程）時，龐大的歷史內容往往會迅速堆高記憶體用量。隨著對話越來越長，每輪互動的送出與回應也常因此變得遲緩。現在開啟大型對話時的每輪回應速度顯著改善，同時有效降低了記憶體消耗。即使面對篇幅極長的工作階段，每次對話往返都能維持順暢反應。
想要迅速接續上一回的中斷進度，恢復對話的指令現在更值得信賴。以前執行 --continue 旗標時，只要終端機位於專案的子目錄、先前發生過異常當機，或者同一工作區內恰好有另一個工作階段（session，從啟動到關閉的完整使用過程）正在運行，系統往往會誤判並強制開出一筆全新空白對話。這讓使用者找不到原本的上下文紀錄。現在這項缺陷已經修好，指令會啟動智慧備援（fallback，在主路徑失效時自動切換至替代方案的容錯設計）機制，自動搜尋當前工作區及其父目錄、子目錄，精準找回最近一次非空的歷史對話。
# 接續當前工作區最近一次非空對話
agy --continue長時間在終端機中緊盯對話輸出，刺眼的畫面閃爍不再打擾你的專注。在行內模式（inline mode，直接在終端機當前游標處嵌入文字的顯示方式）下，使用者過去常常會撞見令人不適的全螢幕空白閃爍。這種刺眼的閃動特別容易發生在文字輸出首次被推入回捲緩衝區（scrollback，終端機用來保留上方歷史畫面的暫存區）的那一刻、每一輪對話即將結束的收尾瞬間，或是按下 Ctrl+L 快捷鍵嘗試清空螢幕之時。這個畫面閃爍問題現在已經徹底修復，各種切換情境下都能維持穩定的視覺感受。
使用遠端裝置協同操作終端機時，隨附介面現在能準確同步你已登入的身分。
以前透過 Remote Control 將遠端隨附介面連線至互動式命令列工作階段（session，從啟動到結束的一整段操作歷程）時，介面端常出現異常狀況。它不會顯示當前活動工作階段的有效登入狀態，反而會錯誤彈出未通過驗證的登入畫面，阻礙使用者操作。這個狀態顯示問題現在已經修好，遠端介面連線後能直接反映正常的登入狀態。
開啟隔離防護執行不可信指令時，底部狀態提示不再引起虛驚。當你在啟動命令列時加入了 --sandbox 命令列旗標，程式明明已經替該工作階段（session，從啟動到關閉的完整執行過程）啟用了終端機沙盒防護。然而畫面底部的狀態列（顯示目前模式與環境資訊的即時通知列）過去卻會錯誤回報沙盒為停用狀態，容易讓人誤以為隔離環境未生效。這個顯示問題現在已經修正，只要使用該旗標啟動，狀態列便會如實反映沙盒的啟用狀態。
# 以終端機沙盒防護模式啟動工作階段
agy --sandbox在終端機檢視流程圖或架構圖時，畫面控制項不再出現用不上的快捷鍵提示。當你在產出物檢視器（artifact viewer，查看程式碼或圖表等生成成果的專用檢視面板）中瀏覽 Mermaid 圖表時，若圖表是以 ASCII 純文字模式而非 Kitty 圖片模式進行算繪，終端機過去依然會顯示圖片專用的縮放按鍵提示。這導致底部的頁尾提示欄列出無效的 Ctrl+= 與 Ctrl+- 縮放快捷鍵，按下去也沒有實際縮放效果。這個介面提示問題現在已經修正，純文字圖表不再顯示多餘的圖片縮放提示。
When designing custom agents tailored for specialized tasks, you now have a completely blank canvas. Adding excludeDefaultComponents: true to the frontmatter of an agent's Markdown file allows you to opt out of loading system default prompt sections and built-in tools, keeping only your explicitly specified behavior logic. This setting prevents general-purpose instructions from interfering with specialized capabilities. Meanwhile, post-task hooks remain fully preserved, keeping custom automated workflows running smoothly.
---
name: my-specialist
excludeDefaultComponents: true
---When facing occasional network congestion or rate limits from remote model servers, the tool now exhibits calmer fault resilience.
Previously, encountering transient genai.APIError failures during a connection would often abort execution immediately. Now, whether dealing with 502, 503, or 504 server errors, per-minute 429 rate limits, or mid-stream transmission interruptions, the program automatically retries in-process using exponential backoff. Completed tool call outputs during retries are fully preserved rather than wasted. If recovery still fails after retrying, the system displays clear and understandable error messages for 503 and 429 situations instead of throwing obscure internal error codes.
When integrating external extension tools, specification checks no longer lock out valid, flexible inputs. When using MCP or provider extension tools, the previous schema validation mechanism was overly strict. Even if a tool specification itself allowed open object schemas (such as declaring {"type": "object"} or explicitly setting additionalProperties: true), the system previously rejected any undeclared arguments. The validation logic is now improved to properly preserve these open object schemas, allowing tools that permit extra parameters to receive data smoothly.
When reloading lengthy conversations with accumulated history, terminal operations remain snappy.
Previously, opening sessions with extensive conversation histories would rapidly inflate memory usage. As conversations grew longer, sending and receiving responses per turn frequently became sluggish. Now, per-turn responsiveness when opening large conversations is significantly improved while effectively reducing memory consumption. Even in exceptionally long sessions, each conversational turn maintains smooth responsiveness.
When resuming interrupted progress from a previous session, the continuation command is now much more dependable. Previously, running the --continue flag would often mistakenly force a brand-new blank conversation if the terminal was located in a project subdirectory, after an abnormal crash, or while another session was running in the same workspace. This left users unable to find their original context history. This flaw has now been fixed; the command initiates a fallback mechanism that automatically searches the current workspace along with its parent and child directories to accurately retrieve the most recent non-empty conversation.
# Resume the most recent non-empty conversation in the current workspace
agy --continueWhen focusing on conversational output in the terminal for long periods, glaring screen flashes no longer disrupt your concentration. In inline mode, users previously often encountered jarring full-screen blank flashes. This glaring flash occurred especially when text output was first pushed into the scrollback buffer, at the closing moment of each conversation turn, or when pressing Ctrl+L to clear the screen. This screen flashing issue is now completely fixed, maintaining a stable visual experience across all transition scenarios.
When using remote devices to collaboratively operate the terminal, the companion UI now accurately synchronizes your signed-in identity.
Previously, connecting a remote companion UI to an interactive CLI session via Remote Control often caused an abnormal state on the UI side. Instead of displaying the active session's authenticated sign-in state, it erroneously showed an unauthenticated sign-in screen, hindering operations. This status display issue is now fixed, allowing remote interfaces to directly reflect the normal signed-in state upon connection.
When enabling isolated protection to execute untrusted commands, bottom status hints no longer cause false alarms. When launching the command line with the --sandbox flag, the program properly enables terminal sandbox protection for that session. However, the status line at the bottom of the screen previously erroneously reported the sandbox as disabled, easily leading users to believe the isolated environment was not active. This display issue is now fixed; launching with this flag ensures the status line accurately reflects the sandbox's enabled state.
# Launch session with terminal sandbox protection mode
agy --sandboxWhen viewing flowcharts or architecture diagrams in the terminal, UI controls no longer display unusable shortcut hints. When browsing Mermaid diagrams in the artifact viewer, if a diagram was rendered in ASCII plain-text mode rather than Kitty image mode, the terminal previously still showed image zoom key hints. This caused the bottom footer hint bar to list ineffective Ctrl+= and Ctrl+- zoom shortcuts that had no actual zooming effect when pressed. This UI hint issue is now fixed, so plain-text diagrams no longer display redundant image zoom hints.
Added support for `excludeDefaultComponents: true` in custom agent Markdown frontmatter, allowing custom agents to opt out of default prompt sections and built-in tools while preserving post-invocation hooks.Improved model API error resilience and diagnostics: transient `genai.APIError` failures (`502`, `503`, `504`, per-minute `429` rate limits, and mid-stream interruptions) automatically retry in-process with exponential backoff while preserving completed tool call outputs, and unrecovered `503` and `429` responses surface clear user-facing error messages.Improved MCP and provider tool schema validation to preserve open object schemas (such as `{"type": "object"}` or explicit `additionalProperties: true`) instead of rejecting undeclared arguments on schemas that allow them.Improved per-turn responsiveness and reduced memory usage when opening large conversations.Fixed `--continue` starting a brand-new conversation when launched from a subdirectory, after a crash, or while another session is open in the same workspace; it now falls back to the most recent non-empty conversation in the current workspace or its parent/child directories.Fixed full-screen blank flashes in inline mode when content first pushes to scrollback, at the end of a turn, or when clearing the screen with `Ctrl+L`.Fixed remote companion UIs connected to an interactive CLI session via Remote Control displaying an unauthenticated sign-in screen instead of the active session's signed-in state.Fixed the status line reporting the terminal sandbox as disabled when the session was launched with the `--sandbox` command-line flag.Fixed image zoom keys (`Ctrl+=` and `Ctrl+-`) and the footer zoom hint appearing in the artifact viewer when a Mermaid diagram is rendered in ASCII mode instead of as a Kitty image.