Antigravity 1.2.2

為終端機擴充各式外部工具，外掛隨附的伺服器重名衝突現在迎刃而解。當你使用 MCP（Model Context Protocol，讓 AI 連接外部工具的協定）時，不同外掛套件隨附的伺服器，過去若恰好取了相同的名稱，彼此之間、或是與你在 mcp_config.json 中自訂的伺服器便會發生衝突與覆蓋。這種命名撞車會導致其中一方無法正常運作。現在系統會自動替外掛隨附的伺服器加上命名空間，統一改以 <plugin>_<server> 的格式獨立註冊，徹底解決重名互相覆蓋的問題。

頻繁開啟歷史對話或在任務間切換，終端機不再悄悄囤積過多電腦記憶體。

以前只要在命令列開啟或掃描對話紀錄，背景負責清理步驟快取（用來暫存各執行步驟紀錄的內部機制）的背景協程常會殘留下來，並在整個工作階段中持續運行，造成不必要的記憶體流失。這個記憶體流失問題現在已經修好，執行 /resume 重新載入對話或在不同對話間頻繁切換後，系統記憶體佔用顯著減少，長久運作依然維持輕快。

把檔案交給 AI 閱讀時，工具現在能主動擋下會塞爆上下文視窗的非預期輸入。以前使用 view_file 工具查看檔案時，若遇上非 UTF-8 編碼的二進位檔案，往往會硬當成純文字解析；甚至在碰上體積大於 100 MB 的超大檔案時，也會直接嘗試塞進模型的上下文視窗（模型在單一輪對話中所能讀取與容納的文字容量上限）。這種過度載入常導致對話中斷或發生錯誤。現在這項缺陷已經修好，面對不支援的二進位格式與超大檔案，工具會在讀取前直接予以拒絕並回傳明確錯誤，避免塞爆模型上下文視窗。

透過金鑰與模型展開多輪深度交談，上一輪的思考脈絡現在能完好延續。

以前透過 GEMINI_API_KEY 啟動 Gemini API 工作階段時，每當使用者送出後續的新訊息，系統往往會遺失先前對話輪次中的模型思考區塊。同時，文字段落與思考段落上的思考簽章（用來驗證推論過程完整性與身分的數位記號）也無法正常傳遞。這兩項缺陷現在皆已修正，模型的思考紀錄與簽章能在多輪對話中順暢保留並正確傳播。

審查程式產出物或建立對話分支，專案內部目錄的管理現在更加乾淨俐落。以前若在 .gemini/ 之上的父目錄路徑中包含 scratch 等被忽略的路徑名稱，產出物審查（檢視程式碼生成成果的檢查面板）常會意外失效而無法觸發。此外，在執行對話分支或快照復原時，過去常會多餘複製內部專用的 .system_generated/subagents 與 .system_generated/worktrees 資料夾，且在刪除對話時也未清理子代理人的後設資料。如今這些管理問題均已修復，分支與復原不再夾帶內部目錄，刪除對話時也會同步清除子代理人紀錄。

手動清空不需要的舊對話紀錄，已經被刪除的歷史檔案絕不再死灰復燃。

以前在使用者刪除對話紀錄之後，若有背景查詢作業嘗試重新連線，往往會在底層硬生生重新建立起一筆空白且缺乏資料結構的 SQLite 資料庫檔案。這些幽靈檔案不僅徒增混亂，也無法正常讀取。這個幽靈重建問題現在已經徹底修復，背景查詢重新連線時不會再擅自生成空白資料庫，刪除對話確實清爽乾淨。

不在任何專案目錄下隨手開啟終端機，獨立交談環境現在具備完善的隔離邊界。以前在未指定工作區資料夾的狀態下啟動新對話，該工作階段往往會意外繼承其他正在運行的開啟工作階段中的專案路徑與自訂設定。這種跨工作階段的設定滲漏，容易導致臨時對話誤讀了無關專案的客製化規則或檔案目錄。這個隔離缺陷現在已經修好，未指定工作區的對話會維持純粹獨立，不再擅自套用其他工作階段的環境路徑與自訂設定。

When extending the terminal with various external tools, name collisions among plugin-bundled servers are now completely resolved. When using MCP (Model Context Protocol, a protocol for connecting AI to external tools), bundled servers across different plugins that happened to share the same name previously collided with each other or overwrote custom servers in your mcp_config.json. Such naming clashes would prevent one of the servers from working properly. The system now automatically namespaces plugin-bundled servers, registering them independently in the <plugin>_<server> format to eliminate overlapping naming conflicts.

When frequently opening conversation histories or switching between tasks, the terminal no longer silently accumulates excessive system memory.

Previously, whenever conversation histories were opened or scanned from the command line, background goroutines responsible for evicting step cache (an internal mechanism for temporarily caching execution step records) often lingered and ran throughout the rest of the session, causing unnecessary memory leaks. This memory leak has now been fixed, significantly reducing system memory usage after running /resume or switching between conversations, keeping long-running sessions nimble and responsive.

When handing files over for AI reading, the tool now actively intercepts unexpected inputs that would overflow the context window. Previously, when using view_file to inspect files, non-UTF-8 encoded binary files were often forcibly parsed as plain text, and files larger than 100 MB were directly loaded into the model's context window (the upper limit of text capacity that the model can ingest and hold in a single turn). Such overloading frequently led to conversation aborts or errors. This defect is now fixed; unsupported binary formats and oversized files are directly rejected with a clear error before loading, preventing context window overflows.

When engaging in multi-turn in-depth conversations with models via API keys, the thinking context from prior turns is now cleanly preserved.

Previously, when launching Gemini API sessions via GEMINI_API_KEY, sending subsequent user messages often caused the system to drop model thinking blocks from previous conversation turns. Additionally, thought signatures (digital markers used to verify the integrity and identity of reasoning processes) on text and thought parts failed to propagate properly. Both defects have now been fixed, allowing model thinking records and signatures to be smoothly retained and correctly propagated across multiple conversation turns.

When reviewing code artifacts or branching conversations, internal project directory management is now much cleaner. Previously, if a parent directory path above .gemini/ contained skipped path names such as scratch, artifact review (an inspection panel for reviewing generated code outputs) often failed to trigger. In addition, when executing conversation forks or snapshot reverts, internal .system_generated/subagents and .system_generated/worktrees directories were unnecessarily copied, and deleting conversations failed to clean up subagent metadata. These management issues have now been resolved: forks and reverts no longer include internal directories, and deleting a conversation cleanly purges subagent records.

When manually clearing unwanted conversation histories, deleted history files will never resurface.

Previously, after a user deleted conversation histories, any background queries attempting to reconnect would recreate empty, schema-less SQLite database files underneath. These ghost files added clutter and could not be read properly. This ghost recreation bug has now been completely resolved; background queries reconnecting after deletion no longer generate empty databases, keeping conversation deletions truly clean.

When launching the terminal outside of any project directory, standalone conversation environments now maintain proper isolation boundaries. Previously, when starting a new conversation without a workspace folder, the session often inadvertently inherited project paths and customizations from other open sessions running concurrently. This cross-session leak caused ad-hoc conversations to mistakenly read irrelevant project rules or file directories. This isolation defect has now been fixed: conversations launched without a workspace folder remain purely standalone, no longer inheriting environment paths and custom settings from other sessions.