`/resume` 新增目錄分組檢視,快速鍵切換對話清單更俐落
你在多個專案資料夾之間切換工作時,要在歷史對話裡找回之前的討論,現在多了更清楚的分類方式。
使用 /resume 指令挑選過去的對話時,新增了自由啟用的工作區分組檢視。只要按下 Ctrl+F 快速鍵,就能在原本平鋪的對話清單與依目錄分組的檢視之間即時切換,迅速找到特定專案底下的歷史紀錄。
發布日期:2026-09-03
你在多個專案資料夾之間切換工作時,要在歷史對話裡找回之前的討論,現在多了更清楚的分類方式。
使用 /resume 指令挑選過去的對話時,新增了自由啟用的工作區分組檢視。只要按下 Ctrl+F 快速鍵,就能在原本平鋪的對話清單與依目錄分組的檢視之間即時切換,迅速找到特定專案底下的歷史紀錄。
偏好直接使用官方 API 金鑰連線的開發者,模型清單中迎來了更新的生力軍。
只要你在環境中設定 GEMINI_API_KEY 連線,模型目錄中就會列出最新的 Gemini 3.8 Flash 模型。在進行日常程式碼查詢或快速編輯時,便能直接在選單中指定此模型,享有更充裕的模型調配彈性。
用 Markdown 檔案撰寫自訂代理時,現在能享有與預設代理完全一致的資源配置。過去自訂代理不會自動承接環境裡現有的各項技能、規則與子代理(subagents,分擔特定任務的輔助代理),讓自訂助理往往少了一整套基本能力;現在自訂代理一律預設繼承周遭的所有技能、規則與子代理,開箱就能直接調用環境資源。
你在自訂代理的 Markdown 定義檔中引用周邊模組時,相對路徑不再因為執行位置不同而迷路。過去在代理定義檔中若使用相對路徑來指定子代理與外掛程式,系統常以當前執行目錄而非定義檔所在位置進行解析,導致檔案找不到;現在相對路徑一律以該代理定義檔所在的資料夾為基準正確解析。
透過外部授權伺服器綁定帳號時,長序號不再卡死登入流程。使用 MCP(Model Context Protocol,讓 AI 連接外部工具的協定)進行 OAuth 登入驗證時,如果授權伺服器回傳長度超過 1024 個字元的授權碼,過去系統會因長度限制而判定失敗;現在這項門檻限制已經修復,長度超過 1024 字元的授權碼也能順利通過驗證並完成服務連接。
在 Windows 系統上檢視已安裝的擴充技能時,面板清單中的分類與層級變得更加精準。過去在 Windows 環境開啟 /skills 面板時,由於主機路徑分隔符號(反斜線 \)的比對問題,導致全域技能與工作區技能經常被錯誤分類;現在路徑比對與分組邏輯已完成修正,所有技能都會依其實際路徑正確歸類到所屬群組。
長時間執行多代理協作或重啟工作階段時,幾項影響執行階段與系統設定的底層問題已獲得修正。過去在重新載入工作階段與呼叫子代理時,設定檔會不斷累積重複的授權項目,現在已修正此累積現象。針對加速執行,改善了工作者工具配置以解決 /boost 在執行期失敗的問題。同時修復了背景摘要更新在軌跡關閉後因嘗試讀取快取步驟(暫存的歷史步驟)而引發的空指標嚴重崩潰。
從外部裝置遠端連線操控開發環境時,網路傳輸通道現在變得更加牢固。
Antigravity 提供的 Remote Control(讓使用者從遠端裝置操控本機對話的功能)依靠反向通道路由來轉發通訊;這次更新全面加固了反向通道路由的處理機制,大幅提升遠端連線時的連線穩定性與路由安全性。
When switching between multiple project directories, finding past discussions in your conversation history is now much more organized.
When using the /resume command to pick a past conversation, you can now opt into a workspace-grouped view. Simply press Ctrl+F to toggle between the flat conversation list and conversations grouped by directory, quickly locating history for a specific project.
Developers who prefer connecting directly with an official API key now have a fresh, responsive option in their model list.
Once you configure GEMINI_API_KEY in your environment, the latest Gemini 3.8 Flash model appears in the model catalog. When performing routine code lookups or quick edits, you can select this model directly from the menu to enjoy greater flexibility.
When defining custom agents with Markdown files, they now receive the exact same resource configuration as default agents. Previously, custom agents did not automatically inherit ambient skills, rules, and subagents (specialized auxiliary agents that handle specific tasks), often leaving them without essential capabilities. Now, custom agents inherit all surrounding skills, rules, and subagents by default, ready to use ambient resources out of the box.
When referencing nearby modules in a custom agent's Markdown definition file, relative paths will no longer get lost depending on where you run the tool. Previously, when specifying subagents and plugins using relative paths in an agent definition, paths were often resolved against the current working directory instead of the definition file's location, causing files to go missing. Relative paths now resolve correctly against the directory containing that agent definition file.
When connecting accounts via external authorization servers, lengthy tokens no longer block your sign-in flow. During MCP (Model Context Protocol) OAuth authentication, if an authorization server returned an authorization code longer than 1024 characters, the process previously failed due to a length limitation. This limit has now been fixed, allowing authorization codes exceeding 1024 characters to authenticate successfully and complete the connection.
When viewing installed skills on Windows, categories and groupings in the panel are now much more accurate. Previously, when opening the /skills panel on Windows, host path separators (the backslash \) caused global and workspace skills to be misclassified due to path-matching issues. The path-matching and grouping logic has now been corrected so that all skills are properly sorted into their respective groups based on their actual paths.
When running multi-agent workflows or reloading sessions over extended periods, several underlying runtime and configuration issues have now been resolved. Previously, reloading sessions or invoking subagents caused duplicate permission grants to accumulate in settings files, which has now been eliminated. For accelerated execution, worker tool configurations have been improved to prevent /boost from failing at runtime. Additionally, a fatal nil pointer crash caused by background summary updates attempting to read cached steps after trajectory closure has been fixed.
When connecting from external devices to control your development environment, network transmission channels are now significantly more resilient.
Antigravity's Remote Control feature relies on reverse-tunnel routing to forward communication. This update hardens the reverse-tunnel routing mechanism, substantially improving connection stability and routing security during remote sessions.
Added an opt-in workspace-grouped view to the `/resume` conversation picker, allowing users to toggle between a flat list and conversations grouped by directory with `Ctrl+F`.Added Gemini 3.8 Flash to the model catalog when connecting with a `GEMINI_API_KEY`.Changed custom agents defined in Markdown to inherit ambient skills, rules, and subagents by default, matching the configuration of default agents.Fixed relative path resolution in Markdown-defined custom agents so relative subagent and plugin paths resolve correctly against the agent definition's directory.Fixed MCP OAuth authentication failing when authorization servers return authorization codes longer than 1024 characters.Fixed skill path matching and grouping in the `/skills` panel on Windows where host path separators (`\`) caused global and workspace skills to be misclassified.Fixed duplicate permission grants accumulating in configuration settings across session reloads and subagent invocations.
Fixed /boost failing at runtime by improving worker tool configurations.
Fixed a fatal nil pointer crash in the agent runtime caused by background summary updates attempting to read cached steps after trajectory closure.Hardened Remote Control reverse-tunnel routing.