新功能
語音聽寫直轉提示詞，支援遠端麥克風轉發
現在你想直接用講話代替鍵盤打字，只要按下 f5 鍵或輸入 /voice 指令，就能啟動或停止語音聽寫，將話語直接轉成提示詞。如果透過 SSH 連線到遠端機器，還能利用 mic-serve 子指令把本機麥克風聲音轉發過去。每個帳號同時只允許一條即時聽寫串流，若開啟第二條串流，提示訊息會清楚告知已達連線上限，不再誤報成額度（quota）耗盡。
繁體中文
新功能
狀態列新增成本欄位，精確呈現即時花費
想在自訂介面掌握 token（AI 處理與計費的文字單位）開銷時，資訊變得更加即時而精細。狀態列（status line，終端機最底下的即時資訊列）資料模型現在新增了 cost 欄位，直接公開當前工作階段未經四捨五入的預估成本。這讓自訂狀態列的使用者能精確計算並顯示累積花費，不必再猜測背後消耗的額度。
效能優化
內建搜尋工具加速檢索，跨環境維持相同行為
搜尋專案程式碼時，速度與穩定度都大幅提升。代理現在直接透過內建的 ripgrep 執行檔進行程式碼搜尋，不再呼叫作業系統本身提供的外部指令。這項改進不僅加快了搜尋速度，也確保在任何沒有安裝 ripgrep 的主機環境下，搜尋行為與輸出格式都能完全一致，不會因為環境差異而發生錯誤。
新功能
免確認模式放行外部工具，任務執行不再中斷
想完全不被打擾地連續執行任務，always-proceed 權限模式現在更加乾脆。過去即使開啟了這項專為免打擾而設計的模式，遇到 MCP（Model Context Protocol，讓 AI 連接外部工具的協定）工具呼叫以及網頁讀取時，畫面依然會頻繁跳出確認提示中斷流程；現在系統會自動核准這些操作，真正做到一路順暢執行到底。
新功能
腳本授權鎖定特定指令，避免全專案權限外洩
批准腳本執行權限時，安全防護變得更為精準嚴密。當你使用 npm run、yarn、pnpm 或 cargo run 等工具時，系統在跳出「永遠允許」的授權建議時會固定綁定特定的腳本名稱。例如當你同意讓代理執行 npm run dev 啟動開發伺服器時，系統只會放行該指令，不會一口氣把整個專案內所有的自訂腳本都列入免審查名單，有效防範未授權的危險指令被隨意觸發。
新功能
對話建立自動命名，接續選單一目了然
切換或挑選舊對話時，清單呈現的內容變得清晰好認。過去每當你建立新對話時，接續對話選單往往只會抓取第一則訊息的簡短佔位文字當作名稱，讓人難以一眼認出討論主題。現在系統在建立對話的當下，就會自動提煉出具有實質意義的完整標題，讓你在日後挑選歷史紀錄時能夠快速找到目標項目。
新功能
Google 憑證連線指引更清晰，快速排除外部工具授權問題
設定使用 Google 憑證的 MCP（Model Context Protocol，讓 AI 連接外部工具的協定）伺服器時，除錯變得更加容易。如果設定好的外部伺服器在本機找不到預設應用程式憑證（ADC，Application Default Credentials），介面現在會顯示具體清晰的錯誤引導，幫助你快速確認授權設定，省下在模糊報錯中盲目摸索的時間。
問題修復
非 ASCII 檔案精準編輯，修復中文與表情符號亂碼損毀
編輯包含中文、特殊重音符號或 emoji（表情符號）的檔案時，內容不再因為編碼計算失誤而損毀。過去代理在處理這類非 ASCII 字元檔案時，若遇到非精確字串比對，可能會在錯誤的位元組偏移處進行文字拼接，進而產生不合法的 UTF-8 編碼並改壞檔案；現在修復了這項偏移判定，確保多國語言與特殊符號都能被精準替換。
問題修復
異常編碼不再卡死串流，對話回應順暢不中斷
工具執行結果或檔案比對內容夾帶異常編碼時，工作階段（session）不再卡住停擺。過去只要工具回傳的內容或檔案差異（diff）中含有不合法的 UTF-8 字元，就會中斷代理背後的狀態串流傳輸，導致整個對話介面卡在回應途中無限等待；現在系統能正確消化這些異常字元，確保狀態持續傳遞，對話不中斷。
問題修復
大型檔案寫入回傳修正，避免無謂重複存檔
寫入大型檔案時，代理不會再白白浪費時間重複操作了。過去當代理處理像 Jupyter 筆記本（notebook）這類較大的檔案時，內容明明已經順利寫入硬碟，系統卻可能誤報寫入失敗，進而觸發無謂的重複重試流程；現在回傳狀態判定已修正，一旦寫入成功就會立刻確認並往下推進。
問題修復
明確指定外掛路徑優先，同名衝突不被覆蓋
自訂功能遇到同名衝突時，系統現在絕對尊重你的個人設定。過去如果你在設定檔中明確指定了某個技能（skill，給 AI 的專屬工作指令包）或是外掛擴充路徑，一旦鄰近目錄中存在同名且被系統自動探索到的自訂項目，你的手動設定可能會意外被覆蓋失效。現在載入優先順序已修正，你在設定檔中明確指定的路徑一定會在命名衝突中勝出。
English
FEATURE
Voice dictation types into prompts and supports SSH mic forwarding
When you want to speak your thoughts instead of typing, press f5 or run /voice to start and stop speech dictation straight into the prompt. When working on a remote machine over SSH, you can use the mic-serve subcommand to forward your local microphone. Each account is limited to one live dictation stream; attempting a second stream now explicitly explains the single-stream limit instead of falsely claiming your quota ran out.
FEATURE
Status line adds cost field to display exact session spend
When monitoring token spend on your own display, the numbers are now far more precise. A new cost field is available in the status line data model, exposing the exact, unrounded estimated cost of the current session. Custom status lines can now calculate and display ongoing token expenses in real time without guessing rounded numbers.
PERF
Embedded ripgrep speeds up search and guarantees consistent behavior
Code searches across your projects are now faster and completely consistent across machines. The agent now runs an embedded ripgrep binary directly rather than shelling out to whatever external search tool happens to be available on the host machine. This approach speeds up search operations and guarantees identical behavior even on systems where ripgrep is not installed at all.
FEATURE
Always-proceed mode auto-approves MCP tools and web reads
For uninterrupted workflows, the always-proceed permission mode now runs without friction. Previously, even though this mode was specifically designed to execute tasks without stopping, it still paused to ask for permission whenever an MCP tool call or page read occurred. Those operations are now auto-approved alongside regular actions, allowing your tasks to proceed smoothly.
FEATURE
Script permissions pin exact names instead of granting broad access
Granting permanent command permissions is now much safer and more precise. When you execute scripts with runners such as npm run, yarn, pnpm, or cargo run, the suggested "always allow" rule now pins down the specific script name. Approving npm run dev to start a dev server now authorizes that specific command alone, rather than granting blanket permission to run every single script in your project.
FEATURE
Conversations auto-name on creation for a clearer resume picker
Picking up past conversations is now much easier to navigate. When a new conversation begins, the system automatically generates a concise and meaningful title right away. Instead of showing an unhelpful raw placeholder snipped from your very first message, the resume picker presents clear topic names so you can instantly spot the session you want to continue.
FEATURE
Clearer error guidance for MCP servers missing Google credentials
Troubleshooting MCP servers configured with Google authentication is now much more straightforward. When a server fails to locate Application Default Credentials on your local system, the CLI presents a detailed and informative error message that points you directly to the missing setup rather than displaying an ambiguous failure.
BUGFIX
Precision editing for non-ASCII text fixes UTF-8 corruption
Editing files that contain non-ASCII text such as CJK characters, accented letters, or emoji is now completely safe from corruption. Previously, when performing an inexact string match on non-ASCII files, text splices could occur at the incorrect byte offset, producing invalid UTF-8 and mangling file contents. The offset calculation has been fixed to ensure multilingual text and special characters are spliced cleanly and accurately.
BUGFIX
Invalid UTF-8 no longer stalls sessions or freezes the interface
Tool outputs or file diffs containing unexpected character encoding will no longer freeze your session. Previously, if a tool result or diff payload contained invalid UTF-8 characters, it broke the agent's internal state stream, leaving the user interface hanging indefinitely mid-response. The stream parser now handles these encoding anomalies resiliently so your conversations continue without disruption.
BUGFIX
Accurate write reporting eliminates redundant retries on large files
File write operations on large documents now report their status accurately without redundant work. Previously, saving bulky files such as Jupyter notebooks could trigger a false failure report even after the data was safely committed to disk, causing the agent to attempt unnecessary retries. The status check has been corrected so confirmed writes proceed immediately to the next task.
BUGFIX
Explicit skill and plugin paths always win name collisions
When custom extensions share identical names, the system now strictly honors your explicit configurations. Previously, if you defined a path to a skill or plugin in your config file, nearby customizations that were automatically discovered in neighboring directories could take precedence and override your settings during name collisions. The resolution order is now fixed so that any path you explicitly configure always wins.
原始 CHANGELOG
Original changelog語音聽寫直轉提示詞，支援遠端麥克風轉發
Added `/voice` dictation, which transcribes speech straight into the prompt; press `f5` or run `/voice` to start and stop it, and use the `mic-serve` subcommand to forward a local microphone to a CLI running on another machine over SSH. Only one live dictation stream is allowed per account, and a second one now says so instead of reporting an exhausted quota.Original changelog狀態列新增成本欄位，精確呈現即時花費
Added a `cost` field to the status line data model, exposing the unrounded estimated cost of the current session so a custom status line can display running token spend.Original changelog內建搜尋工具加速檢索，跨環境維持相同行為
Improved the agent's code search by running an embedded `ripgrep` binary instead of shelling out to whatever the machine provides, so searches are faster and behave identically on systems with no `ripgrep` installed.Original changelog免確認模式放行外部工具，任務執行不再中斷
Improved the `always-proceed` permission mode to auto-approve MCP tool calls and page reads as well, which kept prompting even though the mode exists to run without interruptions.Original changelog腳本授權鎖定特定指令，避免全專案權限外洩
Improved allow-always permission suggestions for script runners such as `npm run`, `yarn`, `pnpm` and `cargo run` by pinning the specific script name, so approving `npm run dev` no longer grants every script in the project.Original changelog對話建立自動命名，接續選單一目了然
Improved conversation titles by generating them automatically when a conversation is created, so the resume picker shows a meaningful name instead of a placeholder from the first message.Original changelogGoogle 憑證連線指引更清晰，快速排除外部工具授權問題
Improved the error shown when an MCP server configured for Google credentials cannot find Application Default Credentials.Original changelog非 ASCII 檔案精準編輯，修復中文與表情符號亂碼損毀
Fixed corrupted edits to files containing non-ASCII text such as CJK characters, accented letters or emoji, where an inexact match was spliced at the wrong offset and produced invalid UTF-8.Original changelog異常編碼不再卡死串流，對話回應順暢不中斷
Fixed the session stalling mid-response when a tool result or file diff contained invalid UTF-8, which broke the agent state stream and left the interface waiting indefinitely.Original changelog大型檔案寫入回傳修正，避免無謂重複存檔
Fixed a file write being reported as a failure after the content had already been written to disk, which sent the agent into unnecessary retries on large files such as notebooks.Original changelog明確指定外掛路徑優先，同名衝突不被覆蓋
Fixed explicitly configured skill and plugin paths losing to customizations discovered automatically nearby, so a path you set in your configuration now wins a name collision.