發布日期：2026-10-13
當你安裝了許多擴充套件，想一次盤點目前啟用了哪些工具與專屬助理時，資訊常會顯示不全。
現在執行 agy plugin list 時新增了 --json 參數，能輸出方便程式解析的結構化資料。該指令不再只顯示手動匯入的套件，而是將從市集安裝的所有外掛、版本編號以及啟用狀態完整列出。在對話中輸入 /plugin 查看外掛詳情時，底下的代理人清單過去會漏掉在 agents.json 根目錄宣告的子代理人。這個問題已經修好，現在會依名稱完整顯示所有載入的代理人，不再只抓取資料夾名稱。
想切換對話模型或調整推理解題深度時，以前常因為少打參數或選錯項目而卡住。
現在使用 --model 或 /model 指定模型時，可以直接填寫模型基礎名稱，不必強制加上 --effort 參數。系統會自動套用預設的思考層級（伺服器預設值，否則依序採用中等或最低層級），不再報錯中斷。若透過 Google Cloud 按量計費的商業帳號登入，選單會主動隱藏專案無權存取的第三方模型。另外使用應用程式預設憑證（ADC，讓本機直接取得雲端授權的機制）時，思考預算與思考層級衝突的報錯也已修復。
出門在外用手機網頁查看終端機進度時，常常碰上偏好選項改不了、或是撤銷規則卻依然生效的窘境。
瀏覽器遠端控制（Remote Control，讓使用者透過網頁介面操作終端機工作階段的功能）的設定面板現在已正式開放。視窗中會直接列出專案與全域權限、排隊訊息、AI 額度、遙測數據與工件審查等實際運作項目。透過 agy remote-control serve 啟動時會載入 settings.json，在網頁上的修改也會即時寫回本機。而在瀏覽器中核准的「永遠允許」規則，過去在規則編輯器或 /permissions 分頁移除後依然會繼續生效；現在規則只會存入共用的 config.json，移除後便立即失效。
在指令列啟動工具時，如果不小心打錯代理人名稱或組合了互相矛盾的參數，很容易因系統沒有明確阻止而產生誤解。
當使用 --agent 指定代理人卻拼錯名稱時，過去系統會默默改用預設代理人執行。現在無頭模式（headless，沒有終端機視覺畫面的執行模式）若搭配 -p 遇到名稱不符，會直接以狀態碼 1 結束並列出所有可用代理人。若在一般互動視窗遇到相同狀況，系統則會先顯示 Agent Not Found 警告才退回預設代理人。若將 --print 與 --prompt-interactive 混用，回報衝突後的退出狀態碼也已從原本代表成功的 0 改為狀態碼 2。
當 AI 伺服器連線不穩或請求連續失敗時，以前終端機常常毫無反應。在模型請求遭遇連續失敗時，系統在自動重試之間的等待時間過去最長可達 4 分鐘，導致整個工作階段長時間停滯。這個問題現在已經修好，自動重試之間的等待間隔上限已縮短為最多 30 秒，減少連線異常時的乾等時間。
使用需要網頁授權的遠端外掛時，最怕網路稍有波動就被強制登出。當連線至使用 OAuth 驗證的遠端 MCP（Model Context Protocol，讓 AI 連接外部工具的協定）伺服器時，過去只要遭遇短暫網路中斷、逾時或伺服器錯誤，系統就會清除登入資訊。現在只有在授權伺服器確實拒絕驗證時才會清除紀錄，短暫斷線不再需要重新登入。
在對話中想查看程式碼修改進度時，終端機有時會明明有改動卻顯示一片空白。在 Git、Mercurial 或 Jujutsu 版本庫中輸入 /diff 查看變更時，若版本控制設定了強制彩色輸出（例如 color.ui = always），過去系統會誤判並顯示 No uncommitted changes。這個問題現在已經修好，即使環境開啟了色彩標記，指令也能正確讀取並顯示所有尚未提交的檔案修改。
在輸入框打字引入舊對話時，如果搜尋速度變慢，游標常常會被卡在原地無法挑選。在終端機輸入 @conversation: 尋找過往對話時，若對話搜尋速度較慢，選單過去會一直卡在 Fetching matches for ... 的讀取畫面。這個問題現在已經修好，只要等候超過 2 秒，選單就會自動退回顯示本機儲存的歷史對話，不必再漫長乾等。
讓代理人分析專案資料夾裡的媒體檔案時，若檔案未下載完整，可能導致後續整段交談全部報銷。當代理人讀取到內容不完整或只有標頭資訊的 WebM 影音檔案時，過去會將該檔案附加至工作階段中，導致後續的每一次對話請求全部發生錯誤。這個問題現在已經修好，系統在讀取到此類殘缺檔案時會直接拒絕載入，不再將損壞的附件帶入對話，確保後續操作能正常進行。
When you have installed many extensions and want to inventory which tools and dedicated assistants are currently active all at once, the information is often incomplete.
Running agy plugin list now adds a --json flag to output structured data for easy programmatic parsing. The command no longer displays only manually imported packages, but completely lists all plugins installed from the marketplace along with their version numbers and enabled status. When typing /plugin in a chat to view plugin details, the agent list below previously omitted subagents declared at the root of agents.json. This issue is now fixed, and all loaded agents are fully displayed by name instead of only grabbing directory names.
When switching chat models or tuning reasoning depth, users often got stuck due to missing parameters or selecting invalid options.
When specifying a model with --model or /model, you can now enter just the base model name without being required to add the --effort flag. The system automatically applies a default thinking effort (the server default, or otherwise medium or lowest in order) without throwing an error and halting. If logged in through a Google Cloud pay-as-you-go commercial account, the menu proactively hides third-party models that the project has no permission to access. In addition, an error where thinking budgets conflicted with thinking effort when using Application Default Credentials (ADC, a mechanism allowing local environments to directly obtain cloud authorization) has also been fixed.
When checking terminal progress via a mobile browser while away, users often ran into dilemmas where preference options could not be changed or revoked rules remained in effect.
The settings panel for browser Remote Control (a feature allowing users to operate terminal sessions through a web interface) is now officially open. The window directly lists active items such as project and global permissions, queued messages, AI quotas, telemetry data, and artifact reviews. Starting via agy remote-control serve loads settings.json, and changes made on the web page are immediately written back to the local machine. Previously, "always allow" rules approved in the browser remained in effect even after being removed in the rule editor or on the /permissions tab. Rules are now stored exclusively in the shared config.json, becoming invalid immediately upon removal.
When launching tools from the command line, accidentally mistyping an agent name or combining contradictory flags could easily cause confusion because the system did not clearly intervene.
When specifying an agent using --agent with a misspelled name, the system previously fell back silently to the default agent. In headless mode (an execution mode without a visual terminal interface) combined with -p, encountering an unmatched name now exits directly with status code 1 and lists all available agents. Under the same condition in a standard interactive window, the system now displays an Agent Not Found warning before falling back to the default agent. When mixing --print with --prompt-interactive, the exit code after reporting the conflict has also been changed from 0 (which originally indicated success) to exit code 2.
When connections to AI servers are unstable or requests fail repeatedly, the terminal previously often became unresponsive. During consecutive model request failures, the system previously waited up to 4 minutes between automatic retries, leaving the entire session stalled for extended periods. This issue has now been resolved, and the upper limit for wait intervals between automatic retries has been shortened to at most 30 seconds, reducing idle waiting time during connection abnormalities.
When using remote plugins that require web authorization, the biggest frustration was being forcibly logged out over minor network fluctuations. When connecting to a remote MCP (Model Context Protocol, a protocol allowing AI to connect to external tools) server that uses OAuth authentication, the system previously cleared login credentials whenever it encountered brief network interruptions, timeouts, or server errors. Now credentials are cleared only when the authorization server definitively rejects authentication, so brief disconnections no longer require logging in again.
When attempting to review code modification progress during a conversation, the terminal sometimes displayed blank output even when modifications were present. When entering /diff to inspect changes in a Git, Mercurial, or Jujutsu repository, if version control was configured to force color output (such as color.ui = always), the system previously misinterpreted the output and displayed No uncommitted changes. This issue has now been resolved, so the command properly reads and displays all uncommitted file modifications even when color formatting is enabled in the environment.
When typing in the input box to reference an earlier conversation, slow searches often froze the cursor in place and prevented selection. When entering @conversation: in the terminal to search for past conversations, if the conversation lookup was slow, the menu previously stayed stuck on the Fetching matches for ... loading screen. This issue has now been resolved: after waiting more than 2 seconds, the menu automatically falls back to showing locally stored conversation history, eliminating the need to wait indefinitely.
When having an agent analyze media files in a project folder, an incompletely downloaded file could ruin the entire subsequent conversation. When an agent read a WebM media file that was incomplete or contained only header information, the system previously attached the file to the session, causing every subsequent conversation request to fail with an error. This issue has now been resolved: the system directly refuses to load such incomplete files and no longer brings corrupted attachments into the conversation, ensuring subsequent operations proceed normally.
Added a `--json` flag to `agy plugin list` for machine-readable output, and fixed the command showing only imported plugins: it now lists every installed plugin, including ones installed from a marketplace, with its version and whether it is enabled.
Fixed the `Agents:` row in `/plugin` plugin details leaving out subagents a plugin declares in its root `agents.json`; it now lists every loaded agent by name instead of by its `agents/` folder name.Changed `--model` and `/model` to accept a base model name without `--effort`: instead of failing with an error asking for `--effort`, the CLI now picks the model's default thinking level (the server's default for that model, otherwise medium, otherwise the lowest available).
Improved the `/model` picker for Business sign-in with a Google Cloud project on pay-as-you-go: third-party models your project can't access are now hidden instead of failing when you send a message.
Fixed requests failing with `thinking_budget and thinking_level are not supported together` when signed in with Application Default Credentials and using a model that has a thinking level.Improved Remote Control so the browser's Settings modal opens for CLI sessions and shows the settings the CLI actually uses (project and global permissions, queued messages, AI credits, telemetry, and artifact review) instead of a notice that settings can't be changed in the browser, and `agy remote-control serve` sessions now use your saved `settings.json` and save changes made in the browser back to it.
Fixed "Always allow" rules approved from a Remote Control browser session still applying after you removed them in the browser's rules editor or the shared tab of `/permissions`; these rules are now saved only to the shared `config.json`.Fixed `--agent` silently running the default agent when the name matches no agent: headless `-p` runs now exit with code 1 and list the available agents, and interactive sessions show an `Agent Not Found` warning before falling back to the default agent.
Fixed combining `--print` with `--prompt-interactive` exiting with status 0 after reporting the conflict; it now exits with status 2 like other print-mode usage errors.Fixed long stalls when model requests fail repeatedly: the agent now waits at most 30 seconds between automatic retries instead of up to 4 minutes.Fixed remote MCP servers that use OAuth signing you out after a temporary network failure, timeout, or server error; saved sign-ins are now cleared only when the server or its authorization server actually rejects them.Fixed `/diff` showing `No uncommitted changes` in Git, Mercurial, and Jujutsu repositories configured to always color output, for example with `color.ui = always`.Fixed `@conversation:` suggestions getting stuck on `Fetching matches for ...` when conversation search is slow; after 2 seconds the list now falls back to your local conversations.Fixed a conversation failing on every later request after the agent read a truncated or header-only WebM file; such files are now rejected when read instead of being attached.