發布日期：2026-10-16
如果你平時依賴 Google Cloud 的應用程式預設憑證登入工具，現在環境變數 AGY_ADC_AUTH 已經正式支援 Claude Opus 5.5 與 Claude Sonnet 5.5 兩款模型。只要在終端機打 /model 切換模型，選單就不會再把所有選項一口氣倒給你，而是會先檢查你的 Google Cloud 專案權限。
當你正在處理需要高強度推理的系統架構，想切換到更合適的模型時，只要打 /model，就能直接確認目前專案是否獲准調用這兩款 Claude 5.5 模型。這避免了你選取模型後送出請求，才因為專案缺少權限而中途失敗。
當你開啟自動模式讓工具自行修改專案時，只要指令需要跨出終端機沙盒（隔離執行指令的受保護環境），現在審查流程變得更快而且不再無故逾時。
對於專案工作區自帶的程式碼格式化、檢查與自動修復指令，審查機制會直接判斷安全並放行，不會再三停下來跳出提示打擾你。相對地，要是指令試圖將專案裡的敏感資料上傳到外部，工具會第一時間阻擋該動作，保護你的隱私與憑證安全。
每次打 /diff 檢視版本提交樹時，你不再需要在一片灰白的純文字裡吃力尋找剛改動的內容。
畫面上提交紀錄的雜湊值（commit hash，Git 用來標記特定版本的一串代碼）與變更識別碼現在都有高亮色彩，讓每個版本節點一目了然。相對地，次要的提交說明文字則調暗顯示，代表當前位置的 @ 符號加粗呈現，未提交的工作複本（Working Copy）一整列更是清晰凸顯，讓版本代碼、時間戳記與說明文字徹底分開，不再黏成難以閱讀的文字區塊。
放手讓背景代理人執行繁重任務時，你最怕它無聲無息卡在半路、永遠等不到下一步。當模型連線停滯且超過 5 分鐘沒有傳回任何資料時，請求會自動取消並重新嘗試，中斷的請求也不會再於紀錄中留下一片空白的回覆步驟。子代理人與工作流程代理人主動發問時，提問不再悄悄遺失；介面現在會明確跳出提示顯示哪個代理人有問題，按快捷鍵 Ctrl+K 或 a 就能開啟選單作答。若是使用 -p 參數的無頭模式（headless，沒有終端機畫面的執行模式），則會自動略過這些提問以確保流程不中斷。遇到執行版本控制指令需要文字編輯器的情況（例如沒加 -m 的 git commit 或 jj describe），工具會直接將編輯器環境變數設為 false，讓指令當場報錯失敗，避免代理人卡在看不見的背景編輯器裡乾等。
在擴充自訂技能或管理外部外掛程式時，設定檔裡的微小語法差異不再導致載入失敗。為自訂技能、規則、工作流程或 Markdown 代理人編寫說明時，即使在 description: 後方未加引號的文字中包含冒號，也能順利解析，不再跳出格式錯誤訊息而中斷載入。外掛若在根目錄的 agents.json 宣告了子代理人，現在執行 agy plugin validate 驗證、agy plugin install 顯示安裝摘要，或在 /plugin 探索分頁瀏覽市集預覽時，都會如實列出，不再遭遺漏。外掛指令與資訊清單現在不再嚴格區分英文大小寫。打 /plugin List 等子指令不再因大寫報錯，資訊清單檔名若取為 Plugin.json，在安裝、驗證以及啟用或停用時也都能正確辨識。
在檢視代理人回覆或產出物檢視器的圖表時，Mermaid 圖形現在都能正常渲染呈現。以前只要圖表開頭帶有 YAML 前置資料、%%{init: ...}%% 指令、%% 註解，或是使用 Mermaid v11 節點語法（例如 id@{ shape: ..., label: ... }），畫面就會直接印出一整段原始碼文字。現在這些語法都能正確繪製成視覺圖表。
當你在輸入框寫下較長的需求、按下換行準備分段時，最怕編輯節奏莫名被打亂。以前只要在輸入內容中打出 @ 喚起檔案下拉選單，接著換行時選單仍然會持續卡在畫面上，強行攔截你的上下方向鍵，按下 Enter 鍵更會直接誤選檔案，而不是如期插入新的一行。這個選單殘留的問題已經修好，讓你換行編輯文字時不再被按鍵劫持。
在檔案與產出物檢視器中按 / 搜尋內文時，遇到特殊字母現在能精確標亮目標。以前若行內包含轉成小寫後字元長度會改變的字母（例如土耳其語的大寫「İ」），搜尋時高亮標記會對錯字元甚至引發當機，使用 n 與 N 跳轉時也常常停在根本沒有亮起的行。現在字元對應錯誤已修好，按 n 與 N 切換時也只會停在真正帶有高亮的行上。
在專案中執行 /codesearch 查找程式碼時，搜尋結果包含多位元組文字現在排版更為俐落。以前只要程式碼長行或檔案路徑含有 CJK（中日韓統一表意文字）等字元，在終端機寬度限制下截斷時，字元位元組常會被切碎，導致結尾顯示出亂碼符號「」。現在遇到長行與路徑都會在安全邊界乾淨加上省略號，不再產生亂碼。
If you rely on Google Cloud Application Default Credentials to sign in to the tool, the AGY_ADC_AUTH environment variable now officially supports Claude Opus 5.5 and Claude Sonnet 5.5. Whenever you type /model in the terminal to switch models, the menu no longer dumps every option onto you at once, but first checks your Google Cloud project permissions.
When you are working on a system architecture that requires intensive reasoning and want to switch to a more suitable model, typing /model lets you directly verify whether the current project is authorized to call these two Claude 5.5 models. This prevents you from selecting a model and sending a request, only to have it fail midway due to missing project permissions.
When you enable auto mode to let the tool modify your project, any command that needs to step outside the terminal sandbox (a protected environment for isolated command execution) now undergoes a faster review process that no longer times out for no reason.
For built-in code formatting, linting, and autofix commands in your project workspace, the review mechanism directly determines they are safe and allows them through without repeatedly pausing to prompt you. Conversely, if a command attempts to upload sensitive project data externally, the tool immediately blocks the action to protect your privacy and credentials.
Whenever you type /diff to inspect the commit tree, you no longer need to struggle through monochrome plain text to locate your recent changes.
Commit hashes and change identifiers across the display now feature highlighted colors, making each revision node immediately clear at a glance. In contrast, secondary commit descriptions are dimmed, the @ symbol representing your current position is bolded, and the entire row for an uncommitted working copy is sharply accented, cleanly separating revision codes, timestamps, and descriptive text so they no longer blur into an illegible block of text.
When letting a background agent handle heavy tasks, the biggest worry is having it stall midway in silence, leaving you waiting indefinitely for the next step. If the model connection stalls and returns no data for more than 5 minutes, the request is automatically canceled and retried, and interrupted requests no longer leave a blank response step in the log. When subagents or workflow agents ask questions, their inquiries are no longer quietly lost. The interface now clearly displays a prompt indicating which agent has a question, and pressing the Ctrl+K or a shortcut opens a menu to answer. When running in headless mode with the -p flag (an execution mode without a terminal window), these questions are automatically skipped to ensure the workflow continues uninterrupted. When a version control command requires a text editor (such as git commit or jj describe without -m), the tool directly sets the editor environment variable to false. This fails the command immediately with an error, preventing the agent from idling in an invisible background editor.
When extending custom skills or managing external plugins, minor syntax differences in configuration files no longer cause load failures. When writing descriptions for custom skills, rules, workflows, or Markdown agents, unquoted text following description: parses smoothly even if it contains a colon, rather than throwing a formatting error that aborts loading. If a plugin declares subagents in its root agents.json, running agy plugin validate to verify, running agy plugin install to view the installation summary, or browsing the marketplace preview on the /plugin Discover tab will now faithfully list them without omissions. Plugin commands and manifests are also no longer strictly case-sensitive. Typing subcommands such as /plugin List no longer errors out due to uppercase letters, and a manifest file named Plugin.json is properly recognized during installation, validation, activation, and deactivation.
When viewing diagrams in agent responses or the artifact viewer, Mermaid diagrams now render properly. Previously, whenever a diagram began with YAML frontmatter, a %%{init: ...}%% directive, %% comments, or used Mermaid v11 node syntax (such as id@{ shape: ..., label: ... }), the display would simply print out a raw block of source code. Now, all such syntax renders correctly into visual diagrams.
When writing longer requirements in the input box and pressing Enter to start a new paragraph, the last thing you want is having your editing rhythm disrupted. Previously, typing @ in the input summoned the file dropdown menu, and pressing Enter to insert a line break left the menu stuck on screen. It forcibly intercepted your up and down arrow keys, and pressing Enter accidentally selected a file instead of inserting a newline as expected. This menu persistence issue has been fixed so that line-breaking and editing text are no longer hijacked by keystrokes.
When pressing / to search file contents in the file and artifact viewers, the target is now accurately highlighted even when encountering special characters. Previously, if a line contained characters whose length changed upon lowercasing (such as the Turkish capital "İ"), search highlighting would misalign with characters or even trigger a crash. Navigating with n and N often stopped on lines that were not highlighted at all. Character mapping errors are now resolved, and cycling with n and N lands only on lines that actually contain highlights.
When running /codesearch to look up code across a project, search results containing multibyte text are now laid out much more cleanly. Previously, whenever long lines of code or file paths contained characters like CJK (Chinese, Japanese, and Korean unified ideographs), truncating them at the terminal width boundary frequently shredded character bytes and produced the replacement symbol "" at the end. Now, long lines and paths are cleanly truncated with ellipses at safe boundaries, preventing mojibake.
Added Claude Opus 5.5 and Claude Sonnet 5.5 for Application Default Credentials (`AGY_ADC_AUTH`) sign-in. Claude models now appear in `/model` only if your Google Cloud project can use them.Improved Auto mode's review of commands that would leave the terminal sandbox: the review is faster and no longer times out without a decision, it allows the workspace's own format, lint, and fix commands instead of asking about them, and it blocks commands that would upload sensitive data.Improved the `/diff` Commit Tree view with colors: commit hashes and change IDs are highlighted, descriptions are dimmed, the `Working Copy (Uncommitted Changes)` row stands out, and `@` is bold, so hashes, timestamps, and descriptions no longer run together.Fixed the agent hanging indefinitely when the model connection stalled: if no data arrives for 5 minutes the request is now canceled and retried automatically, and a stalled request no longer leaves an empty response step behind.
Fixed questions asked by subagents and workflow agents being lost, which left the subagent waiting forever: the alert now reads `<agent> has a question` and shows it, and `Ctrl+K` or `a` in the subagent's view opens a question picker that sends your answer to that agent. Headless `-p` runs skip these questions so the subagent can continue.
Fixed agent shell commands hanging on an invisible editor when a version control command wants one, such as `git commit` without `-m` or `jj describe`: the CLI now sets `GIT_EDITOR`, `HGEDITOR`, and `JJ_EDITOR` to `false` for agent commands, so they fail right away instead.Fixed skills, rules, workflows, and Markdown agents failing to load with `mapping values are not allowed in this context` when an unquoted `description:` contains a colon, such as `description: Runs checks: formatters and linters.`
Fixed `agy plugin validate`, the summary `agy plugin install` prints, and marketplace previews in the `/plugin` Discover tab leaving out subagents a plugin declares in its root `agents.json`.
Fixed plugin names and manifests being case-sensitive where the agent was not: `/plugin` subcommands such as `/plugin List` now work in any case instead of failing with `Unknown subcommand`, and a manifest named `Plugin.json` is now found by `agy plugin install`, `agy plugin validate`, and enable and disable.Fixed Mermaid diagrams that start with YAML frontmatter, a `%%{init: ...}%%` directive, or `%%` comments, or that use Mermaid v11 syntax such as `id@{ shape: ..., label: ... }` nodes, showing as raw source instead of rendering in agent responses and the artifact viewer.Fixed the `@` file-mention dropdown staying open after a newline in a multi-line prompt, where it took over `Up`/`Down` and made `Enter` pick a file instead of inserting the newline.Fixed `/` search in the file and artifact viewer highlighting the wrong characters, or crashing, on lines with letters whose lowercase form has a different length such as `İ`; `n` and `N` now stop only on lines that show a highlight.Fixed long lines and file paths containing CJK or other multi-byte text turning into `�` characters in `/codesearch` results; they are now cut cleanly with an ellipsis.