發布日期：2026-09-19
在終端機中開啟產出物詳細檢視器時，現在能直接看到圖形化數學公式與流程圖。
新機制在支援 Kitty 協定的終端機中，能直接在檢視器畫面中渲染 LaTeX 數學公式（包括 $$...$$ 與 math、latex、tex 程式碼區塊），以及 Mermaid 流程圖和循序圖。若終端機不支援該協定，畫面會自動切換為 Unicode ASCII 純文字作為備援。遇到較寬的圖表時，你可以使用左右方向鍵進行水平平移；隨時按下 m 鍵，還可在影像、ASCII 以及原始文字三種檢視模式之間循環切換。
AI 一口氣丟來好幾個釐清問題時，填答畫面現在讓你前後翻題，也看得出哪幾題已經答過。
提示介面現在支援使用左右方向鍵在多個問題之間自由切換與回顧。若選擇自行填寫答案，確認送出後的文字會直接在行內即時預覽；已經完成選取的選項旁會標註勾選符號，讓整體作答狀態一目了然。當你瀏覽到最後一個問題時，直接按下確認即可一步完成整份送出。
載入龐大的專案規範與自訂規則時，其他擴充功能不再因為空間不足而被無預警捨棄。
系統現在為使用者與工作區規則分配了專屬的 20,000 個 token（AI 處理文字的計量單位）預算配額。超出上限的規則會在換行邊界處自動截斷，並透過檔案路徑與描述清楚列出超額清單。這項獨立配額確保了龐大的規則清單不會再擠掉技能、工作流程、子代理或 MCP（讓 AI 連接外部工具的協定）等重要工具。
代理一啟動就帶在身上的工具箱，這版拿掉了三個舊的搜尋與列目錄工具。
系統改進了工具配置基準線，將舊版的 find_by_name、grep_search 以及 list_dir 工具從預設工具組中淘汰。這項調整能精簡代理預設攜帶的工具清單，避免多餘工具造成干擾。若你的自訂代理依然需要這些舊工具，只要在設定檔的 tools 清單中明確列出，依然可以照常呼叫與使用。
外掛裝多了以後，指令名稱互相打架、升級完舊程式還賴在背景，這幾個毛病這版都處理了。
系統修復了外掛技能的斜線指令在設定名稱自帶前綴時會出現雙重前綴（如 /<plugin>:<plugin>:<skill>）的錯誤，並解決了多個外掛定義相同簡稱技能時指令遭到遮蔽的問題。外掛升級時，系統不再讓舊的 MCP（讓 AI 連接外部工具的協定）伺服器與 sidecar 行程在已刪除的安裝目錄中持續殘留運作。啟動時也會主動清理市集快取（把下載檔案暫存起來的空間）中已被取代的舊版外掛與暫存目錄。此外，已停用的外掛不會再因只查詢啟用項目的指令而自 /plugin 清單中消失，移除外掛時也強化了外掛識別碼的路徑驗證。
兩個常看的畫面這版變清楚了：啟動時的歡迎橫幅，和查額度的 /usage 面板。
企業帳號在啟動命令列工具時，啟動橫幅現在會在登入帳號與方案層級下方，清楚標示當前作用中的 Google Cloud 專案識別碼（Project ID）。同時，/usage 用量查詢面板也移除了每個進度條下方重複顯示的剩餘配額百分比文字，讓整體畫面更為清爽緊湊。
模型 API 一時連不上時，你不必再盯著停住的終端機乾等好幾分鐘。以前在觸發重試機制時，每次重試之間的退避等待時間會逐步拉長，單次最長甚至需要等待多達 4 分鐘，導致終端機長時間停滯。
現在系統將單次重試的退避等待時間上限嚴格限制在 30 秒，縮短連線遇阻時的等待空檔，讓對話能更快恢復順暢交流。
如果你常用 -p 讓它在背景自己跑，或習慣在終端機裡清畫面、退出，這幾項修正都跟你有關。
系統將 Bubble Tea 框架升級至 v2.0.9，修復了終端機畫面渲染問題，以及在離開工具或清除畫面時 Kitty 鍵盤通訊協定堆疊的處理異常。在無頭模式（headless，沒有終端機視覺介面的執行模式）使用 -p 或 --prompt 執行時，修復了偶爾漏掉背景任務等待通知的問題。背景 SDK 工具的執行進度現在會確實記錄至任務記錄檔中，並透過複製截斷的指令輸出預覽來降低記憶體用量。
以前在電腦上啟動命令列工具時，可能會意外干擾桌面應用程式（desktop app）的狀態，導致原本釘選的對話被取消釘選，或是已讀過的對話被重新標記為未讀。現在這個問題已經修好，啟動命令列工具不會再誤改桌面版應用程式中的對話標記與釘選狀態。
When opening the artifact detail viewer in the terminal, you can now directly view rendered graphical math equations and flowcharts.
On terminals supporting the Kitty protocol, the new mechanism renders LaTeX math equations (including $$...$$ and math, latex, and tex code blocks) alongside Mermaid flowcharts and sequence diagrams inline in the viewer. If a terminal does not support the protocol, the display automatically falls back to plain Unicode ASCII text. For wider diagrams, you can use the Left and Right arrow keys to pan horizontally. Pressing the m key cycles between Image, ASCII, and Raw text view modes at any time.
When the AI asks several clarifying questions at once, the answer screen now lets you move back and forth between them and see which ones you have already answered.
The prompt interface now supports using the Left and Right arrow keys to freely switch between and review questions. If you choose to write in an answer, confirmed text is previewed inline in real time. Answered options display check marks so your response status is clear at a glance. When you navigate to the final question, pressing confirm submits the entire set in a single step.
When loading extensive project guidelines and custom rules, other extensions are no longer discarded without warning due to space constraints.
The system now allocates a dedicated 20,000-token budget for user and workspace rules. Oversized rules are automatically truncated at newline boundaries, with over-budget rules clearly listed by file path and description. This independent quota ensures that large rule sets no longer evict essential tools such as skills, workflows, subagents, or Model Context Protocol (MCP) integrations.
The toolbox agents carry from the moment they start drops three legacy search and directory-listing tools in this release.
The system refines the baseline tool configuration by retiring the legacy find_by_name, grep_search, and list_dir tools from the default toolset. This change streamlines the list of tools carried by agents by default, avoiding distractions caused by redundant tools. If your custom agents still require these legacy tools, you can continue calling and using them as usual by explicitly declaring them in the tools list of your configuration.
With several plugins installed, command names could clash and old processes could linger after an upgrade; this release deals with both.
The system fixes an issue where plugin skill slash commands received double prefixes (such as /<plugin>:<plugin>:<skill>) when the frontmatter name already included the plugin prefix, and resolves command shadowing when multiple plugins define skills with the same short name. During plugin upgrades, legacy MCP servers and sidecar processes are no longer left running from deleted installation directories. Startup now actively prunes superseded plugin versions and staging directories from the marketplace cache. Additionally, disabled plugins no longer disappear from the /plugin list after enabled-only customization queries, and path validation for plugin IDs has been hardened during uninstallation.
Two screens you look at often got clearer in this release: the startup banner and the /usage quota panel.
When enterprise accounts launch the command-line tool, the startup banner now clearly displays the active Google Cloud Project ID beneath the signed-in account and plan tier. Meanwhile, the /usage panel removes the duplicate remaining-quota percentage text beneath each progress bar, resulting in a cleaner and more compact interface.
When the model API briefly fails, you no longer have to stare at a stalled terminal for minutes. Previously, when triggering retries, the backoff delay between attempts grew progressively longer, taking up to 4 minutes for a single wait and leaving the terminal stalled for extended periods.
The system now strictly caps the per-attempt retry backoff at 30 seconds, reducing idle waiting intervals during connectivity hiccups so conversations can resume smoothly much faster.
If you often run it headless with -p, or routinely clear the screen and exit in the terminal, these fixes apply to you.
The system upgrades the Bubble Tea framework to v2.0.9, resolving terminal rendering issues and anomalies in Kitty keyboard protocol stack handling upon exiting the tool or clearing the screen. When running in headless mode using -p or --prompt, an issue where background-task waiting notices were occasionally skipped has been fixed. Execution progress for background SDK tools is now reliably recorded in task log files, and memory usage is reduced by cloning truncated command output previews.
Previously, launching the CLI on your computer could inadvertently interfere with desktop app state, causing pinned conversations to be unpinned or read conversations to be marked as unread. This issue has now been resolved, ensuring that starting the CLI will no longer unintentionally modify conversation read flags or pinned states in the desktop application.
Added inline Kitty graphics rendering for LaTeX math equations (`$$...$$` and fenced `math`, `latex`, and `tex` blocks) and Mermaid flowcharts and sequence diagrams in the artifact detail viewer, with Unicode ASCII fallback, Left/Right arrow horizontal panning for wide diagrams, and `m` to cycle between Image, ASCII, and Raw viewsImproved the `ask_question` interactive prompt with Left/Right arrow navigation between questions, inline previews of confirmed write-in answers, check marks on answered options, and single-step submission on the final questionImproved customization token budgeting by giving user and workspace rules a dedicated 20,000-token budget (cutting oversized rules on newline boundaries and listing over-budget rules by path and description) so large rule sets no longer evict skills, workflows, subagents, or MCP toolsImproved the default agent and subagent toolset by retiring the legacy `find_by_name`, `grep_search`, and `list_dir` tools from the default baseline while keeping them available to custom agents that explicitly list them in `tools`Fixed plugin skill slash commands being double-prefixed (`/<plugin>:<plugin>:<skill>`) when a skill's frontmatter name already includes the plugin prefix, or being shadowed when multiple plugins define skills with the same short name
Fixed plugin upgrades leaving old MCP server and sidecar processes running from deleted install directories, and pruned superseded plugin versions and staging directories from the marketplace cache on startup
Fixed disabled plugins disappearing from the `/plugin` list after enabled-only customization queries, and hardened plugin ID path validation on uninstallImproved the CLI startup banner for enterprise accounts to display the active Google Cloud Project ID beneath the signed-in account and plan tier
Improved the `/usage` panel by removing the duplicate remaining-quota percentage line beneath each progress barImproved model API retry responsiveness by capping per-attempt retry backoff at 30 seconds instead of waiting up to 4 minutes between attemptsFixed terminal rendering and Kitty keyboard protocol stack handling on exit and screen clear by upgrading Bubble Tea to v2.0.9
Fixed headless (`-p` / `--prompt`) runs occasionally skipping the background-task waiting notice, logged background SDK tool progress to task log files, and reduced memory usage by cloning truncated command output previewsFixed starting the CLI unpinning conversations or marking them unread in the desktop app