發布日期：2026-07-16
你能用全新的 /codesearch（或 /cs）指令快速翻找整個專案的程式碼了。它預設支援正規表達式（regex，一種用符號匹配字串的進階語法），你也能加上 -F 進行字字精準的比對，甚至用 f: 來過濾特定檔案。在介面體驗上，系統排版變聰明了：它會明確畫出一條線，告訴你先前的對話在哪裡被壓縮過（context-compaction，把舊記憶摺疊起來省空間的功能）；行內程式碼周圍多餘的空白被移除了；含有錢字號的終端機指令，也不會再被 Markdown 解析器誤認成數學公式而變成亂碼。此外，遇到 eligibility error（資格驗證錯誤）時，錯誤訊息和驗證網址現在會乖乖安插在輸入框裡，不會再把你的畫面往上頂。就像你買了一個新書櫃，不只附贈了超強的檢索系統幫你精準找書，還順手把原本會爆出櫃子的通知單好好夾在書頁裡、撫平書本間的折頁。
現在你打開 Antigravity（Google 的終端機 AI 助手）時，不用再枯等它暖機了。以前在啟動階段，系統會為了尋找技能（skills，提供額外能力的設定）而卡死在漫長的檔案掃描上；現在它改成非同步載入（asynchronously loading，讓主程式先跑，背景慢慢找），大幅縮短了啟動時間。不只如此，當系統在找技能、規則（rules）、代理或掛鉤（hooks）這類客製化設定時，現在懂得把所有路線整合起來並快取（caching，把查過的結果暫存起來直接用）尋找紀錄，砍掉了多餘的磁碟讀寫。同時，它也修好了一個小蟲：如果你的規則資料夾是用符號連結（symlink，指向另一個位置的檔案）連過去的，系統不會再笨笨地把它載入兩次。就像你開車出門，以前每次都要全車地毯式安檢，現在只要一發動就能上路，而且導航系統會記住去過的地方，不用每次都在同一個路口找路。
在使用不閃爍模式（no-flickering mode，讓終端機畫面更新時不會一直閃爍的顯示模式）時，你的電腦不會再因為超長對話而卡死了。以前對話累積太多，每次更新畫面都會重新把對話索引（index）越長越大，導致 CPU 狂飆；現在重建索引改成了冪等（idempotent，不管跑幾次結果都一樣，不會無限疊加），讓負載穩定下來。此模式下的行數計算也修好了，狀態列不再跑出視窗外；關閉獨立文件預覽（artifact，系統生成的獨立文件預覽）詳細畫面後，也不會再有幽靈留言殘留在螢幕上。現在更支援了選取即複製（copy-on-select）：滑鼠拖曳反白，放開就能把純文字（不含 ANSI 顏色代碼）複製到剪貼簿，且虛擬卷軸會自動隱藏不礙事。就像畫超長的清明上河圖，以前每畫一筆書僮都要從頭數到尾，現在他只看最新變化；當你想臨摹其中一段，一框起來就印好給你，畫軸也會自動閃避。
當你把系統放在背景自動跑任務時，它不會再自作主張開後門。以前在無頭模式（headless runs，無終端機視窗的執行模式）下，遇到需確認權限的工具時不是卡死就是自動同意；現在會溫和拒絕並提示需設定哪條白名單（allow-rule）。在「一律同意（always-proceed）」模式下，也不會再錯誤批准寫入工作區之外的檔案。Linux 使用者以前常因作業系統鑰匙圈（OS keyring，存密碼的保險箱）讀取太慢而被切斷，逼得重新登入；現在若無 D-Bus 工作階段匯流排（D-Bus session bus，用來在工作階段溝通的管道）就直接繞過鑰匙圈，超時先跳過一小時並放寬讀取時間。就像派機器人去倉庫，遇到鎖住的門它不硬闖也不傻等，只回報缺哪把鑰匙；保全動作慢，它也願意多等，不會動不動要你重簽通行證。
當你讓 AI 呼叫外部工具時，它不會再因為對方的失誤而跟著同歸於盡了。以前如果你的 MCP（Model Context Protocol，讓 AI 連結外部工具的通訊協定）伺服器沒有回應，代理（agent）就會像石化了一樣無限期死等；現在系統為連線、列出工具清單與呼叫工具的動作都設下了超時停損點（timeouts），時間一到就會果斷放棄。此外，有些特定的工具呼叫以前會弄壞對話紀錄（corrupted history），導致 AI 之後完全無法回應，整個對話直接壞死；這個嚴重的蟲現在也被抓出來修好了。就像派助理去拿資料，以前沒人他會永遠死等並弄丟聯絡簿；現在給他設了超時規定，時間一到就直接回報，也不會再搞丟聯絡簿。
You can now quickly dig through your entire project's code using the new /codesearch (or /cs) command. It interprets your queries as regular expressions (regex, an advanced syntax for pattern matching) by default, but you can use -F for strictly literal matches and filter specific files using f: globs. Visually, the interface has grown smarter: it now draws a clear indicator line to show exactly where earlier conversation history was squeezed down (context-compaction, the feature that folds old memories to save space); extra padding spaces around inline code have been stripped away; and shell commands containing dollar signs will no longer be mistaken for math formulas. Also, when an eligibility error occurs, the message and verification URL now sit politely inline in your input loop, rather than pushing your whole screen upwards. It's like getting a new bookshelf that comes with a powerful cataloging system to pinpoint any book, neatly tucks annoying notification slips into the pages, and smooths out the awkward gaps between the books.
You no longer have to stare at a stalling screen while Antigravity (Google's terminal AI assistant) warms up. Previously, during bring-up, the CLI would completely block your interactive startup while it laboriously scanned your file system to discover skills (settings providing extra capabilities); it now loads them asynchronously (letting the main interface start instantly while loading happens in the background), drastically cutting down startup latency. Moreover, when the system hunts for customizations like skills, rules, agents, or hooks, it now consolidates these directory walks and caches (temporarily saving results for instant reuse) its filesystem lookups, eliminating redundant I/O reads. It also fixed a minor hiccup where customization rules were loaded twice if the directory was reached via a symlink (a file pointing to another location). It's like driving your car: instead of a full-body security sweep every time you turn the key, you can just start the engine and go, and your GPS remembers the routes so it doesn't have to recalculate at every intersection.
Your computer will no longer choke on massive conversations when using no-flickering mode (a display mode that prevents the terminal screen from flashing on every update). Previously, as the chat grew, every single screen refresh would rebuild the conversation index (the internal map of the chat) larger and larger, causing high CPU usage and unbounded rendering costs. This index rebuild is now idempotent (meaning it converges to a stable state no matter how many times it runs), keeping system load stable. Row math in this mode was also corrected so the status line stays neatly inside your viewport, and dismissing an artifact (a standalone document preview generated by the system) detail view will no longer leave ghost comments on the screen. Even better, it now supports copy-on-select: dragging to highlight text and simply releasing the mouse instantly copies the clean text (stripped of ANSI terminal color codes) to your clipboard, hiding the virtual scrollbar so it doesn't interfere. It's like working on a mile-long scroll — previously, your assistant would recount the entire scroll every time you added a stroke; now, they only process what's needed. And when you want to trace a section, circling it instantly gives you a clean photocopy, while the scroll handles tuck themselves away.
When running tasks in the background, the system no longer makes unauthorized decisions. Previously in headless runs (modes without a terminal window, like -p), tools requiring permission either hung or auto-approved; the CLI now soft-denies them and prints a stderr notice showing the allow-rule needed to permit them. In "always-proceed" mode, outside-of-workspace file writes are no longer incorrectly auto-approved. Linux users are saved from login loops caused by slow OS keyring reads; the CLI now bypasses the keyring if no D-Bus session bus is present, skips it for an hour after a timeout, and extends timeout windows. Like sending a robot to a warehouse: encountering a locked door, it reports the key it needs; if the guard is slow, it waits longer instead of triggering alarms.
When you ask the AI to invoke an external tool, it will no longer get dragged down by unresponsive servers. Previously, if your MCP (Model Context Protocol, a standard connecting AI to external tools) server simply never responded, the agent would freeze like a statue and hang indefinitely; the system now actively bounds connection attempts, tool-listing requests, and per-tool-call attempts with strict timeouts, allowing it to gracefully give up and move on. Furthermore, certain tool calls used to catastrophically break the conversation by corrupting the history (the underlying memory of the chat), which blocked all further responses and killed the session instantly; this fatal bug has been fully squashed. It's like sending an assistant to fetch a document — they used to stand there forever if the room was empty and lose their walkie-talkie; now you've given them a strict timeout rule, so they come back if nobody answers, and keep their radio perfectly safe.
Added a `/codesearch` command (aliases `/cs` and `/search`) to interactively search code across your workspace, interpreting queries as regex by default with `-F`/`--literal` for exact matching and `f:`/`file:` globs to include or exclude paths.
Added an indicator at each context-compaction boundary so you can see where earlier compaction happened.
Improved eligibility error handling by showing errors with a verification URL inline in the input loop instead of stacking them above the screen.
Removed the padding spaces around inline code for tighter rendering.
Fixed code-block corruption where `$..$` math expansion desynced from the Markdown parser and mangled fenced shell snippets such as `git fetch "$GIT_REMOTE"` by detecting fenced code blocks line-by-line.Improved interactive startup latency by loading skills asynchronously so the CLI no longer blocks on a synchronous, filesystem-heavy skill-discovery pass during bring-up.
Improved customization loading latency for skills, rules, agents, and hooks by consolidating directory walks and caching filesystem lookups to cut redundant I/O during discovery.
Fixed customization rules being loaded twice when a rules directory is reachable through a symlink.Added copy-on-select in no-flickering mode so dragging highlights text and releasing the mouse copies the ANSI-stripped selection to the clipboard, and hides the virtual scrollbar so it no longer interferes with copying multi-line output.
Fixed high CPU and unbounded render cost on large conversations in no-flickering mode by making index rebuilds idempotent so the conversation index converges instead of growing on every rebuild.
Fixed lingering artifact comments after dismissing the artifact detail view and corrected no-flickering-mode row math so the status line renders correctly within the viewport.Fixed headless (`-p`) runs hanging or silently auto-approving tools that require a permission confirmation, so the CLI now soft-denies such tools and prints a stderr notice naming the allow-rule needed to permit them.
Fixed outside-of-workspace file writes being incorrectly auto-approved in always-proceed mode.
Fixed repeated sign-in prompts on Linux caused by the OS keyring: the CLI now bypasses the keyring when no D-Bus session bus is present (headless hosts and containers), skips it for an hour after a timeout, and uses longer keyring timeouts so a slow-but-successful credential read is no longer cut short and forced into a fresh login.Fixed MCP servers hanging the agent indefinitely when a server never responds by bounding connection, tool-listing, and per-tool-call attempts with timeouts.
Fixed conversations breaking after certain tool calls, which previously corrupted the conversation history and blocked all further responses.