繁體中文

新功能

企業帳號多了三條登入路

公司把 Gemini Enterprise 帳號交給你用時,現在可以從 Google Cloud 專案登入,使用組織分配或自動配發的 GE-Standard、GE-Plus 授權席次,並挑選推論執行的地區;管理員原本設定的控制項目也會跟著生效。公司若把登入身分交給外部身分供應商,可在 Google Cloud 登入畫面選 Use advanced SSO config,走 Workforce Identity Federation。另一條路是 Application Default Credentials,讓你能使用 Agent Platform。它像同一棟辦公樓新開三個對應不同證件的入口,各人仍照公司的門禁規則進門。

新功能

同一對話開兩次,先提醒你分岔

你在同一台電腦又開了一個 CLI(命令列介面,用文字下指令的程式)視窗,剛好載入同一段對話時,畫面會出現不會打斷工作的提醒。它會指向 /fork 指令,讓你把對話分成一條新支線再繼續。以前兩個視窗會同時往同一份對話紀錄寫內容,結果像兩個人共用一支筆在同一行上書寫,句子容易交錯;現在先看到路牌,你可以選擇分開寫,也能照原本的對話繼續工作。

新功能

沙盒能看 Git 資訊,不能改 Git

讓代理在 Git 專案裡工作時,.git 資料夾現在只開放唯讀(可以查看、不能修改)。這個資料夾存著提交紀錄、分支等版本控制資料,所以代理仍能查到專案的來龍去脈,卻不能從沙盒(隔離執行區域)裡重寫它。讀取版本控制的中繼資料和保住它,現在終於可以同時成立。像把檔案櫃的玻璃門換成展示櫃:裡面的卷宗看得到,手卻伸不進去改頁面。

新功能

掛鉤終於接得到最後一棒

你替流程加了 hook(掛鉤,特定事件發生時被觸發的小程式)來收尾或記錄時,最後一步不再從眼前溜走。hooks.json 裡定義的掛鉤現在會先於內建的終止檢查執行,因此 PostInvocation 能觀察一回合的最後一次呼叫,Stop 也不再被內建檢查擋在後面、永遠到不了。這像接力賽終點前先讓計時員站到跑道旁;最後一棒抵達時,他才能按下碼表。

新功能

排程工具看得懂數字格式了

模型呼叫 schedule 工具排工作時,若把 DurationSeconds 或 MaxIterations 寫成 JSON(資料交換用的文字格式)數字,不會再因為少了引號而整個呼叫失敗。整數值現在會被接受;例如工作持續秒數或最多反覆次數直接寫數字即可。帶小數的數字則會收到清楚的錯誤訊息,因為這兩欄只接受整數。它像訂位表終於能看懂你寫「3 人」而不強迫你把 3 圈成一段文字,同時仍會退回「3.5 人」這種不合理填法。

問題修復

選定模型與 effort 不再被悄悄換掉

你明明指定了 --model 或 --effort,互動式工作階段和無頭 -p 執行卻可能偷偷用回已儲存或預設的模型;這個問題已修正,旗標會在模型設定初始化前套用。只寫 --effort 時也會依你目前選的模型解析,不再依預設模型算完後把你帶去另一個模型。effort 是模型投入推理力道的選項,像在點餐時同時選主餐和份量;現在廚房會照你點的那一份做,不會因為後台預設而端錯餐。

問題修復

背景工作終於能確實停下或放到背景

協調多個子代理時,按下停止不會再只停掉眼前那段對話,卻讓所有後代子代理和它們的背景工作繼續跑;整棵子代理樹現在會一起停止。協調者等待仍在執行的工作時,也不會再一直塞入空白 continue 步驟,耗掉 token(AI 處理與計費的文字單位)又卡住進度。另一項修正讓 MCP 伺服器標記為永遠在背景執行的工具,真的不會堵住整回合。像關掉一串延長線時,所有接上的電器都會斷電,而不是只熄掉第一盞燈。

問題修復

選項、游標與模型選單都能正常操作

在多選提示裡,空白鍵又能勾選選項了,包含 ask_question 對話框和初次設定時的匯入勾選框;提示列也會明說可用空白鍵,不必只靠 x。輸入框顯示建議下拉選單時,左右方向鍵不再被選單攔走,因此能移動游標修改文字。模型挑選器若暫時沒有模型可選,畫面也會保留標頭、頁尾與按 esc 返回的提示。這像三個常用按鈕終於裝回控制台,不會在你伸手時消失。

問題修復

MCP 斷線不再留下偷偷跑的程序

MCP(Model Context Protocol,讓 AI 連接外部工具的協定)伺服器的連線意外中斷後,不會再留下沒關掉的程序。那種程序雖然不在畫面上,仍可能繼續占著電腦資源;這次修正會在連線掉落時處理這個外洩。發生這類斷線時,程序不會再在背景遺留。就像客服通話突然斷線後,電話線另一頭不會還有一位接線員一直掛著耳機等人回話。

問題修復

文件畫面與網路拒絕紀錄不再被藏起來

你在 artifact viewer(成品檢視器)開一般文件時,內容不會再被一律橫向裁掉。原本只有圖表類 artifact(可檢視的輸出成品)需要這種裁切,現在普通文件會正常顯示。另一個修正補上沙盒的網路紀錄:即使命令本身成功結束,只要其中的網路請求被沙盒擋下,也會留下記錄。像展覽館不再把所有作品都削窄成海報比例,門禁也會登記每次被拒絕進門的人。

English

FEATURE

Three new sign-in routes for enterprise accounts

If your company gives you a Gemini Enterprise account, you can now sign in through a Google Cloud project, use an assigned or automatically assigned GE-Standard or GE-Plus license seat, choose the region where inference runs, and retain the controls set by your administrator. If your organization uses an outside identity provider, choose Use advanced SSO config on the Google Cloud sign-in screen for Workforce Identity Federation. Application Default Credentials is another route for using Agent Platform. Think of one office building adding three entrances for different badges while keeping the same company access rules.

FEATURE

A warning before two sessions write into one conversation

If you open another CLI window on the same machine and load the same conversation, you now get an advisory that does not interrupt your work. It points to /fork, which creates a new branch of the conversation before you continue. Previously, both windows could write into the same conversation trajectory, like two people sharing one pen on the same line and mixing their sentences together. Now there is a signpost before that happens, so you can choose separate lines.

FEATURE

The sandbox can inspect Git, but cannot rewrite it

When an agent works in a Git repository, the .git directory is now read-only: it can be inspected but not changed. That directory holds version-control information such as commits and branches, so the agent can still understand the project's history while being unable to rewrite it from inside the sandbox, an isolated execution area. It is like replacing a filing cabinet door with a display case: you can read the records inside, but cannot reach in and alter a page.

FEATURE

Hooks can finally catch the final handoff

If you add hooks, small programs triggered by particular events, to record or finish a workflow, the final step no longer slips past them. Hooks defined in hooks.json now run before the built-in termination checks. That lets PostInvocation observe the final call in a turn, and lets Stop run instead of remaining unreachable behind those checks. It is like placing the timekeeper beside the track before the last relay runner arrives; only then can the stopwatch be pressed at the finish.

FEATURE

The scheduler now understands numeric values

When a model calls the schedule tool, DurationSeconds and MaxIterations no longer fail just because they arrive as JSON, a text format for exchanging data, numbers without quotation marks. Integral values are now accepted, so a duration in seconds or a maximum iteration count can be written as a number. Decimal values get a clear error because these fields require whole numbers. It is like a reservation form finally understanding “3 people” without demanding that 3 be turned into a text label, while still rejecting the nonsensical “3.5 people.”

BUGFIX

Your selected model and effort no longer get quietly replaced

You may have specified --model or --effort yet found an interactive session or headless -p run quietly using the saved or default model. That is fixed: the flags are applied before model configuration is initialized. A bare --effort is also resolved against the model you selected, rather than the default model that could move you elsewhere. Effort is the setting for how much reasoning work a model uses. It is like ordering both a main dish and its portion size; the kitchen now follows the order you gave instead of serving the back-office default.

BUGFIX

Background work can now actually stop or stay in the background

When coordinating several subagents, stopping no longer ends only the current conversation while descendant subagents and their background tasks keep running. The whole subagent tree now stops together. A coordinator waiting for active work also no longer keeps injecting empty continue steps, wasting tokens, units of AI processing and billing, and blocking progress. Another fix makes tools that an MCP server marks as always-background genuinely avoid blocking a turn. It is like switching off a power strip and cutting power to every connected appliance, rather than darkening only the first lamp.

BUGFIX

Options, cursor keys, and the model picker work normally again

In multi-select prompts, the spacebar can select an option again, including in the ask_question dialog and the onboarding import checkbox; the hint bar now says so instead of leaving x as the only toggle. When the input box shows a suggestion dropdown, Left and Right no longer get captured by that dropdown, so you can move the cursor and edit text. If the model picker has no models available, it now still shows its header, footer, and an esc hint to return. It is like putting three everyday buttons back on a control panel so they do not disappear when you reach for them.

BUGFIX

A dropped MCP connection no longer leaves a process behind

When an MCP server, using the Model Context Protocol that lets AI connect to external tools, drops its connection unexpectedly, it no longer leaves an unclosed process behind. Such a process may be invisible on screen yet keep using computer resources; this fix handles that leak when the connection falls away. It is like a customer-service call dropping without leaving an operator at the other end still wearing a headset and waiting for an answer.

BUGFIX

Documents and denied network requests are no longer hidden

When you open a plain document in the artifact viewer, it no longer gets horizontally clipped by default. Only diagram artifacts, viewable output pieces, need that treatment, so ordinary documents now display normally. Another fix fills in the sandbox network record: even if the command itself ends successfully, a network request denied by the sandbox is recorded. It is like a gallery no longer trimming every work to poster proportions while the entrance desk also logs every visitor denied admission.

原始 CHANGELOG

Original changelog企業帳號多了三條登入路
Added Business sign-in for Gemini Enterprise accounts, so you can authenticate with a Google Cloud project under Google Cloud terms, use a license seat allocated or auto-assigned from your organization's GE-Standard or GE-Plus subscription, run inference in a chosen region, and have your organization's administrator controls applied to various features.
Added Workforce Identity Federation sign-in for enterprise users, available as the `Use advanced SSO config` option on the Google Cloud sign-in screen, so organizations that federate identity through an external provider can authenticate with their own identity provider.
Added sign-in with Application Default Credentials so you can use Agent Platform.
Original changelog同一對話開兩次,先提醒你分岔
Added a non-blocking advisory banner when the same conversation is already open in another CLI instance on the same machine, pointing at `/fork` so two sessions no longer interleave writes into one trajectory.
Original changelog沙盒能看 Git 資訊,不能改 Git
Improved the terminal sandbox by granting read-only rather than writable access to a Git repository's `.git` directory, so the agent can inspect repository metadata without being able to rewrite it from inside the sandbox.
Original changelog掛鉤終於接得到最後一棒
Improved hook ordering so hooks defined in `hooks.json` run before the built-in termination checks, which lets `PostInvocation` hooks observe the final invocation of a turn and lets `Stop` hooks run at all instead of sitting unreachable behind the built-ins.
Original changelog排程工具看得懂數字格式了
Improved the `schedule` tool to accept `DurationSeconds` and `MaxIterations` when a model emits them as bare JSON numbers rather than strings, accepting integral values and rejecting non-integral ones with a clear error instead of failing the call.
Original changelog選定模型與 effort 不再被悄悄換掉
Fixed `--model` and `--effort` being ignored in interactive sessions and in headless `-p` runs, where the flags were applied after model configuration had already been initialized so the run silently fell back to the persisted or default model.
Fixed a bare `--effort` resolving against the default model instead of the model you actually have selected, which could silently move you to a different model.
Original changelog背景工作終於能確實停下或放到背景
Fixed stopping a subagent tree stopping only the conversation it was invoked from, while every descendant subagent and the background tasks they owned kept running and the CLI still reported them as killed.
Fixed a forced-continuation deadlock where a coordinator waiting on active subagents or background tasks would loop injecting empty continue steps until it hit the invocation limit, wasting tokens and blocking progress.
Fixed tools that an MCP server marks to always run in the background executing as blocking calls that stalled the turn.
Original changelog選項、游標與模型選單都能正常操作
Fixed the spacebar not toggling an option in multi-select prompts, including the `ask_question` dialog and the onboarding import checkbox, which left `x` as the only working toggle; the hint bar now advertises it.
Fixed the Left and Right arrow keys being captured to navigate the input box suggestion dropdown, so you can move the cursor and edit text again while suggestions are showing.
Fixed the model picker's "No models available" state rendering without its header and footer, so it now shows the standard chrome and an `esc` hint to go back.
Original changelogMCP 斷線不再留下偷偷跑的程序
Fixed an MCP process leak when a server connection dropped unexpectedly.
Original changelog文件畫面與網路拒絕紀錄不再被藏起來
Fixed the artifact viewer corrupting plain documents by horizontally clipping every document rather than only the diagram artifacts that need it.
Fixed the sandbox not recording blocked network requests when the command itself exited successfully, which hid the fact that a request had been denied.