Antigravity 1.1.24

當你在多個設定來源或外掛中定義了同名代理（agent，替你執行特定任務的 AI 助手），打開 /agents 選單時不會再看見重複項目的干擾了。過去系統在探索時，只要從不同設定來源或外掛發現同名代理，就會把每一筆都重複列在挑選器中；現在選單已經會自動去重，確保清單乾淨清爽。

就算你在一處已被刪除或無法存取的工作目錄中啟動命令列，工具初始化與代理啟動程序也不會再莫名崩潰。以前系統在處理記憶體中的參數架構時，會因目錄失效而找不到相對參照；現在全面改用絕對 URI（統一資源識別碼，以標準格式標記位置的字串）格式，確保各項工具與代理在任何目錄狀態下都能順利初始化。

撰寫自動化腳本時，你在無頭模式（headless，沒有終端機畫面的背景執行模式）下透過管線接收標準輸出或標準錯誤，程式結束時不再無故卡死。過去背景派生出來的子行程會持續霸佔呼叫端的管線，導致主程式遲遲無法順利關閉退出；現在系統已在保留的串流上加上 FD_CLOEXEC 標記，在結束時徹底釋放所有管線資源。

手動編輯工具設定檔時，你不必再為隨手寫下的筆記或多打的逗號膽戰心驚了。過去在 mcp_config.json 檔案中加入單行註解 //、多行註解 /* */ 或是在結尾多留尾隨逗號，都會導致 MCP（讓 AI 連接外部工具的協定）設定解析失敗；現在解析器已具備更佳的容錯能力，能順利讀取帶有註解與多餘逗號的設定內容。

清理舊對話時，本機硬碟空間現在會被收拾得乾乾淨淨。以前每當你刪除對話紀錄，系統雖然移除了主對話，卻常把關聯的註記檔案遺留在本機儲存空間中，隨著時間推移默默堆積成無用的孤立檔案；這次修正補齊了刪除邏輯，會在刪掉對話的同時自動清除對應的註記檔案。

在 AI 全力衝刺大專案的中途，你終於能安心插話問個小問題了。當你開啟 /goal 指令讓系統執行長期目標時，若中途打 /btw（順帶一提，用來向 AI 提出旁支問題的指令）插話發問，先前系統常會把這則閒聊誤當作目標的一部分強行繼續推動，甚至亂呼叫工具；現在兩者狀態已徹底拆開，插問只會解答當前疑問，絕不干擾進行中的主任務。

When you define agents (assistants that perform specific tasks) with identical names across multiple configuration sources or plugins, opening the /agents picker no longer shows distracting duplicates. Previously during discovery, finding an agent with the same name across different sources or plugins listed every instance in the picker; the menu now automatically deduplicates entries, keeping your selection list clean.

Launching the command line from a working directory that has been deleted or is inaccessible will no longer cause tool initialization or agent startup to crash. Previously, parameter schemas held in memory lost relative references when the directory was invalid; switching fully to absolute URI schemes ensures all tools and agents initialize reliably regardless of directory state.

When running automation scripts in headless mode (background execution without a visual interface) with piped standard output or standard error, invocations no longer hang on exit. Previously, spawned child processes kept the caller's pipes open and prevented the main process from shutting down cleanly; setting FD_CLOEXEC on preserved streams now ensures all pipe resources are fully released on exit.

When manually editing tool configurations, you no longer need to worry about quick notes or extra commas breaking your setup. Previously, adding single-line comments //, multi-line comments /* */, or trailing commas in mcp_config.json caused MCP (protocol connecting AI to external tools) configuration parsing to fail; the parser is now more forgiving and reads configurations containing comments and trailing commas smoothly.

When clearing old conversations, your local disk space is now kept completely tidy. Previously, deleting a conversation removed the main chat but often left associated annotation files in local storage, gradually accumulating useless orphaned files over time; the deletion logic has been updated to automatically prune matching annotations alongside the conversation.

While AI is actively working through a major project, you can now safely ask quick side questions without derailment. When running a long-term goal with /goal, asking a side question with /btw (by the way, a command for asking auxiliary questions) previously led to treating the tangent as part of the active goal, forcing unwanted progression and unintended tool calls; their execution states are now decoupled so side questions resolve your query without interrupting the main task.