發布日期：2026-10-03
當你想出門在外只帶一支手機或平板，直接在行動裝置上呼叫 AI 幫忙寫程式時，現在只要打開 Termux 就能立即啟動。官方現在提供了預先編譯好的 Android 原生執行檔，讓命令列介面（CLI，用文字輸入指令與電腦互動的工具）能夠直接在 Termux 裡面運作。以前你必須先在手機裡設定 proot-distro（在 Android 上模擬完整 Linux 作業系統的環境工具）才能勉強跑起來，過程繁瑣且佔用額外空間。
現在只要安裝好 Termux，直接下載執行檔就能開始對話。
當你一邊看著 AI 飛快輸出長篇回答，一邊想要上下捲動歷史紀錄或同步輸入新文字時，整個介面都不再感到延遲。這次顯著減少了工具在每次畫面更新與重新繪製時所配置的記憶體（RAM，電腦運行時暫存資料的記憶空間）。以前每次文字串流（內容一邊生成一邊即時顯示的輸出方式）或畫面變動，都會重複申請大量暫存記憶體，容易讓終端機在長對話中反應變慢。
現在無論輸出多長，打字與捲動都保持俐落順暢。
當你想查詢自己的使用額度何時恢復、在終端機輸入 -p "/usage" 檢查時，看到的重置時間會自動換算成你所在的本地時間。以前終端機印出來的都是國際標準時間，你總得自己在腦中手動加減時差，才能算出確切幾點幾分能繼續使用。為了兼顧自動化程式，只要你使用管線串接輸出或加上 --json 參數，工具依然會輸出 UTC（世界協調時間，跨時區通用的標準時間基準）時間戳記。
這項調整既方便肉眼閱讀，又完全不影響既有腳本運作。
當你授權 AI 執行版本控制工具的特定指令，或是直接拒絕了某項操作時，系統對權限邊界的拿捏更加嚴謹。現在審核 repo 工具時，「總是允許」的建議只會鎖定在具體的子指令（例如只放行 repo status 查詢狀態），不再一口氣放行所有危險指令，行為與 git、hg 和 jj 完全一致。若權限請求遭到拒絕，例如無法主動向你詢問的子代理（在背景協助處理特定分工的代理人）碰壁時，AI 會嚴格遵守拒絕決定。它絕不會私自改用其他指令、腳本或替代工具繞道偷跑。
在長對話中快速調整終端機視窗大小時，對話畫面不再卡死不動。以前多次連續縮放視窗觸發畫面重繪後，整個工作階段畫面上常常不再出現新的執行步驟，即使 AI 仍在背後工作，你也只能看著凍結的畫面乾等。現在工具會在連線中斷時自動嘗試重新連線，只有在重試失敗時才會跳出警告；若連線真的丟失，只要發送一則新訊息就能立刻恢復畫面連線。
當方案配額耗盡且帳戶內的 AI 點數餘額不足以支付時，系統會立即給出清楚的提示訊息。以前在餘額不足的情況下送出請求，終端機會停在「Working...」狀態不斷重試約兩分半鐘，最後才拋出模糊的「Agent execution terminated due to error.」錯誤；現在會當場明確告知「Your AI credits balance is too low to continue.」，幫你省下無謂的等待時間。
透過 SSH（用來安全連線至遠端電腦的通訊協定）使用 CLI 命令列工具，退出 CLI 後，終端機畫面不再被奇怪的字元代碼弄亂。以前在 macOS 內建的 Terminal.app 透過 SSH 使用 CLI，退出 CLI 後，螢幕上常會殘留像 Ga=q,f=32,s=1,v=1,i=31;AAAAAA== 這類探測圖片協定的亂碼。現在工具在偵測圖片支援前會先要求終端機表明身分，在 SSH 連線下只將 Kitty 與 Ghostty 視為具備圖片顯示能力的終端機，避免 WezTerm 和 Konsole 產生亂碼佔位字串。
在終端機開啟 Vim 編輯模式處理含有表情符號或多重組合標記的文字時，游標移動與編輯變得更加精準。以前游標移動到像 👩💻、⚠️ 這種由多個編碼組合而成的 Emoji，或是天城文等帶有組合標記的字元時，游標會卡在字元中間。現在使用 h、l 等移動鍵會一次跨越整個完整字元，按下 x 刪除或 r 取代時，也會直接操作整組完整字元，不再將複合字形拆散切碎。
當你輸入 /resume 挑選並恢復先前的歷史對話時，搜尋關鍵字較長也較不會折行。以前對話選擇清單裡的搜尋輸入框寬度有限，查詢字串較長時就會自動折行，導致最前面的文字被擠出可視範圍之外。現在搜尋框會自動延展並填滿整個終端機的寬度，較長的查詢較不會折行，開頭也不會因此被擠出畫面。
當你在設定目錄管理自訂規則時，AI 現在能正常讀取 ~/.gemini/config 裡的 rules/ 資料夾、AGENTS.md、GEMINI.md、skills.json、rules.json、plugins.json 與 agents.json，且修改前會先徵詢同意，其餘檔案維持禁止存取。若 ~/.gemini/GEMINI.md、~/.gemini/AGENTS.md 或其 ~/.gemini/config/ 對應檔案為指向同一個實體檔案的符號連結（指向另一個位置的捷徑檔），系統會自動排除重複，避免規則被多次載入。
當你嘗試登入某些透過標準協定串接的外部工具時，不再因為伺服器回傳的狀態碼而登入失敗。在登入支援 MCP（Model Context Protocol，讓 AI 連接外部工具的協定）的伺服器時，若對方的 OAuth 客戶端註冊流程回傳了代表成功的 HTTP 200 而非預期的 201，以前工具會誤判為錯誤並拋出「registration failed with status 200」；現在這個問題已經修好，能正常相容這兩種成功狀態碼並順利完成登入。
當你請 AI 分析高解析度圖檔時，不再受格式差異影響。以前 AI 讀取大型 PNG 與 JPEG 圖片時都會自動縮小以符合尺寸上限，唯獨漏掉了 WebP 格式；現在修復後，WebP 也會比照辦理自動縮小，避免超大圖檔超出規格限制。
When you want to travel light with only a phone or tablet and call on AI to write code directly on your mobile device, you can now launch right inside Termux. Pre-compiled native Android binaries are now provided, allowing the CLI to run directly within Termux. Previously, you had to set up proot-distro (an environment tool that emulates a full Linux operating system on Android) on your phone just to get it running, a tedious process that consumed extra storage.
Now, simply install Termux, download the binary directly, and start chatting.
When you watch AI rapidly stream long responses while scrolling through history or typing new text simultaneously, the interface no longer lags. Memory allocated during each screen update and redraw has now been significantly reduced. Previously, every text stream or visual change repeatedly requested large amounts of temporary memory, which tended to slow down terminal responsiveness during lengthy conversations.
Now, regardless of output length, typing and scrolling remain crisp and smooth.
When checking when your quota resets by entering -p "/usage" in the terminal, the reset time is now automatically converted to your local time. Previously, the terminal always printed Universal Coordinated Time, forcing you to manually calculate timezone offsets in your head to know the exact time you could resume usage. To preserve compatibility with automated programs, piping the output or passing the --json flag still outputs standard UTC timestamps.
This adjustment enhances human readability while leaving existing scripts completely unaffected.
When authorizing AI to run specific version control commands or denying an operation, permission boundaries are now handled much more strictly. When prompting permissions for repo, "Always Allow" suggestions target only the specific subcommand (such as granting only repo status for status queries) instead of approving every risky command at once, aligning behavior with git, hg, and jj. If a permission request is denied—such as when a subagent running in the background encounters a block without the ability to prompt you—the AI strictly respects the denial. It will never secretly bypass the block by switching to other commands, scripts, or alternative tools.
Rapidly resizing the terminal window during lengthy conversations no longer freezes the screen. Previously, consecutive resizes triggering redraws often caused the session screen to stop displaying new execution steps, leaving you staring at a frozen view while AI continued working in the background. The tool now automatically attempts to reconnect whenever the connection drops, raising a warning only if retries fail; if the connection is truly lost, sending a new message immediately restores the visual connection.
When plan quota is exhausted and account AI credits are insufficient, the system now provides an immediate and clear notification. Previously, submitting a request with insufficient balance left the terminal stuck on "Working..." retrying repeatedly for about two and a half minutes before throwing a vague "Agent execution terminated due to error." error; it now promptly informs you on the spot that "Your AI credits balance is too low to continue.", saving you unnecessary waiting time.
When using the CLI tool over SSH, exiting the CLI no longer clutters your terminal screen with strange character codes. Previously, using the CLI over SSH in macOS Terminal.app often left behind garbled image protocol probing strings like Ga=q,f=32,s=1,v=1,i=31;AAAAAA== after exit. The tool now queries terminal identity before checking image support, recognizing only Kitty and Ghostty as image-capable terminals under SSH connections to prevent WezTerm and Konsole from generating mojibake placeholder strings.
Cursor navigation and editing are now more precise when handling emojis or complex combining marks in terminal Vim editing mode. Previously, moving the cursor over multi-codepoint emojis like 👩💻 or ⚠️, or scripts with combining marks such as Devanagari, caused the cursor to get stuck midway through characters. Navigation keys like h and l now traverse entire composite characters in a single step, and pressing x to delete or r to replace acts on the complete grapheme cluster without breaking glyphs apart.
When entering /resume to pick and restore previous session history, longer search queries are much less likely to wrap. Previously, the search input box in the conversation picker had limited width, causing longer queries to wrap automatically and push the beginning of the text out of view. The search box now expands to fill the entire terminal width, making longer queries less likely to wrap and preventing the start of text from getting pushed off-screen.
When managing custom rules in the configuration directory, AI can now properly read the rules/ directory, AGENTS.md, GEMINI.md, skills.json, rules.json, plugins.json, and agents.json within ~/.gemini/config, asking for confirmation before making edits while keeping all other files restricted. If ~/.gemini/GEMINI.md, ~/.gemini/AGENTS.md, or their counterparts under ~/.gemini/config/ are symlinks pointing to the exact same physical file, the system automatically deduplicates them to prevent rules from loading multiple times.
When logging in to certain external tools integrated through standard protocols, login attempts no longer fail due to server response status codes. When logging in to servers supporting MCP (Model Context Protocol, a protocol connecting AI to external tools), if the server's OAuth client registration flow returned a successful HTTP 200 instead of the expected 201, the tool previously misidentified it as an error and threw "registration failed with status 200". This issue is now fixed to properly support both successful status codes and complete the login process smoothly.
When asking AI to analyze high-resolution images, format differences no longer cause issues. Previously, when AI read large PNG and JPEG images, it automatically downscaled them to fit image size limits, but WebP format was overlooked. Following this fix, WebP images are also automatically downscaled accordingly, preventing oversized image files from exceeding specification limits.
Added prebuilt native Android binaries of the CLI that can run directly in Termux, without requiring a proot-distro environment.Improved responsiveness while typing, scrolling, and streaming agent output by sharply reducing the memory the CLI allocates on every update and redrawImproved `-p "/usage"` to show quota reset times in your local time zone when printing to a terminal; piped and `--json` output still use UTC timestamps, so existing scripts are unaffectedImproved the always-allow suggestion when approving `repo` commands: it is now limited to the subcommand, such as `repo status`, instead of allowing every `repo` command, matching `git`, `hg`, and `jj`
Improved how the agent handles a permission request that is not approved, such as one raised by a subagent that cannot ask you: it now respects the denial and no longer tries to work around it with other commands, scripts, or toolsFixed the conversation view freezing for the rest of the session, with no new steps appearing while the agent kept working, after redraws such as several quick terminal resizes in a long conversation; the CLI now reconnects automatically, warns only if reconnecting fails, and sending a new message recovers a lost connectionFixed requests made after your plan quota is used up and your AI credits balance cannot cover them: instead of retrying on `Working...` for about two and a half minutes and then showing the generic "Agent execution terminated due to error.", the CLI now immediately says "Your AI credits balance is too low to continue."Fixed text like `Ga=q,f=32,s=1,v=1,i=31;AAAAAA==` being left behind in macOS Terminal.app after exiting the CLI over SSH; the CLI now asks the terminal to identify itself before probing for image support, and over SSH only Kitty and Ghostty are treated as image-capable, so WezTerm and Konsole no longer show garbled image placeholdersFixed the Vim mode cursor stopping inside emoji such as 👩💻 or ⚠️ and characters with combining marks such as Devanagari: `h`, `l`, and other motions now move over whole characters, and `x` and `r` delete or replace the whole character instead of a hidden part of itFixed the search box in the `/resume` conversation picker wrapping long queries so that their beginning scrolled out of view; the box now uses the full terminal widthFixed the agent being blocked from reading your global rules and customization files in `~/.gemini/config`, including the `rules/` folder, `AGENTS.md`, `GEMINI.md`, `skills.json`, `rules.json`, `plugins.json`, and `agents.json`; it can now read them and asks before editing them, while other files in that folder stay off-limits
Fixed global rules being added to the agent's context more than once when `~/.gemini/GEMINI.md`, `~/.gemini/AGENTS.md`, or their `~/.gemini/config/` counterparts are symlinks to the same fileFixed signing in to MCP servers whose OAuth client registration responds with HTTP `200` instead of `201`, which previously failed with `registration failed with status 200`Fixed large WebP images opened by the agent not being scaled down to fit image size limits the way PNG and JPEG images are