發布日期：2026-07-31
你想把檢查工作交給不開視窗的列印模式（print mode，跑完一個指令就輸出結果的模式）時，現在可以直接輸入 -p "/my-skill review this diff"。開頭的斜線指令會被辨認，技能也會被展開套用，不再整串當成普通文字丟給 AI。即使指令和參數之間按了換行或 Tab，仍會正確拆開；若你要原封不動送出斜線文字，則加上 --disable-slash-commands。這像寄件櫃終於讀得懂收件人姓名與地址各在哪一欄，不會把整張地址貼紙當成一段留言。
剛打開互動對話，某個 MCP（Model Context Protocol，讓 AI 連接外部工具的協定）伺服器若很慢或卡住，第一輪不會再跟著停在門口。互動工作階段會在背景載入 MCP 伺服器；不過無頭模式和單次執行仍會等待，確保它們唯一那輪工作看得到完整工具組。連線中斷後恢復時，MCP 也不再強迫你重新做完整驗證。像登機時，櫃檯先讓旅客開始辦理，行李車在後頭跟上；但只飛一趟的包機仍得等所有行李裝妥才起飛。
同一段對話裡，你已核准過的指令樣式，之後遇到相同樣式不必每次再按一次允許。這份核准會記住到對話結束；不是把所有指令都放行，而是只讓符合已核准模式的後續指令通過。另一個常見卡點也拿掉了：預設可用的系統暫存目錄，現在連建立或更新檔案都包含在權限裡，不會只因要寫檔又跳出詢問。這像進出大樓的訪客證，警衛記得今天已查過的通行類型；而原本可進卻不能放東西的置物櫃，終於配了鑰匙。
替自動化流程裝了停止掛鉤（hook，特定事件發生時被觸發的小程式）後，萬一它每次都要求繼續，工作不會再被困到永遠。連續續跑次數到達你設定的上限後，這個掛鉤便不能再阻擋，該輪會正常結束。PostToolUse 掛鉤也修正了：它以前連使用者輸入和模型回覆這些非工具步驟都會觸發，還因此忽略原本設定的匹配條件；現在只照該照的條件動作。像舞台監督以前讓謝幕演員一再回台，戲散不了場，還把觀眾說話當成燈光提示；現在謝幕有次數上限，提示也只認對的暗號。
一次貼進很長內容後，如果輸入框把它收成一個折疊的貼上區塊，現在按刪除會整塊消失。以前畫面上會一個字一個字地刪，留下看得到的碎片，送出時卻仍把完整內容交出去，讓人難以判斷自己到底刪掉了什麼。成品檢視器也一起修好：從編輯畫面回來時，程式碼的語法上色不會消失；離開成品細節時，也會回到正確面板。像撕掉便利貼時終於整張撕下，不會桌上殘一小角、抽屜裡還藏著原件；看完展品也能回到剛才那個展廳。
用無頭模式的 stream-json 輸出（把執行過程用 JSON 一行行送出的格式）接到自己的程式時，開頭的 init 事件現在只會列出你這個版本實際可用的工具。以前它可能先宣告某些工具，實際執行卻找不到，接收端若照那張清單安排工作便會踩空。這次是把起跑線上的工具牌換成正確版本；不是新增工具，而是讓宣告和手上的工具箱一致。像活動場地的入口看板不再列出已拆掉的櫃位，排隊的人不會走到門前才發現那裡什麼都沒有。
When you want to hand a review to print mode (a mode that runs one command and returns its result without opening a window), you can now type -p "/my-skill review this diff" directly. The opening slash command is recognized, and its skill is expanded and applied instead of being handed to AI as literal text. A newline or Tab between command and arguments also works; if you need to send slash text unchanged, add --disable-slash-commands. It is like a mailing counter finally separating recipient and address fields instead of treating the whole label as a note.
When you open an interactive chat and an MCP server is slow or stuck, your first turn no longer waits at the door with it. MCP (Model Context Protocol, a protocol that connects AI to outside tools) servers load in the background for an interactive session. Headless and one-shot runs still wait, so their only scripted turn sees the complete toolset. When a connection returns after dropping, MCP no longer makes you complete authentication all over again. It is like starting passenger check-in while the baggage carts catch up, while a one-flight charter still waits until every bag is aboard.
Within one conversation, a command pattern you have already approved no longer makes you press Allow again when that pattern appears later. The approval lasts for the conversation; it does not open the gate to every command, only later commands matching the approved pattern. Another common interruption is gone too: the default system temporary directory now includes permission to create and update files, not only read them. It is like a visitor badge: security remembers the kind of entry checked today, while a locker you could open but not put anything into finally gets its key.
After you add a stop hook (a small program triggered by a particular event) to an automated flow, it can no longer trap work forever if it asks to continue every time. Once consecutive continuations reach your configured limit, that hook can no longer block and the turn ends normally. PostToolUse hooks are fixed too: they used to fire on non-tool steps such as user input and model replies, then ignore their configured matchers. They now act only under the right conditions. It is like a stage manager who kept calling curtain-call actors back and mistook audience chatter for a lighting cue; now curtain calls have a limit and cues require the right signal.
After pasting a long piece of text, the input can collapse it into one paste placeholder. Deleting it now removes the whole block. Before, the display removed one character at a time and left a visible fragment, while sending still submitted the full content, making it hard to know what had been deleted. The artifact viewer is fixed as well: syntax highlighting remains when you return from editing, and leaving artifact details returns you to the correct panel. It is like removing a sticky note in one piece instead of leaving a corner on the desk while its original remains hidden in a drawer, then returning to the same gallery after viewing an exhibit.
When your program consumes stream-json output from a headless run (a format that sends the run as JSON lines), its opening init event now lists only tools that are truly available in your build. Before, it could advertise tools that could not be found when work began, leaving a receiver that planned from the list on empty ground. This changes the tool sign at the starting line; it does not add tools, it makes the announcement match the toolbox at hand. Like an event map that no longer lists booths already removed, so visitors do not queue up only to find nothing at the door.
Added slash-command and skill expansion to print mode, so a headless run such as `-p "/my-skill review this diff"` now resolves and applies the skill instead of sending it as literal text, with `--disable-slash-commands` to opt out.
Fixed slash commands not being recognized when separated from their arguments by a newline or tab, so a prompt starting with a command followed by a newline is now parsed as a command instead of being sent verbatim.Improved interactive startup so a slow or hanging MCP server no longer stalls the first agent turn, loading MCP servers in the background for the interactive session while headless and one-shot runs keep blocking so their single scripted turn still sees the full toolset.
Fixed MCP servers forcing a full re-authentication after a dropped connection.Improved permission grants so a pattern approved at a prompt is recorded for the rest of the conversation, letting later commands that match it run without prompting again.
Improved the default system temporary-directory grant to cover writes as well as reads, so agents no longer trigger a permission prompt when creating or updating files there.Fixed stop hooks that always block hanging the agent forever; after a configurable number of consecutive continuations, the hook can no longer block and the turn ends normally.
Fixed `PostToolUse` hooks firing on non-tool steps such as user input and model responses, which also caused them to ignore their configured matchers.Fixed deleting into a collapsed paste placeholder removing one character at a time, which left a visible fragment in the prompt while the full pasted content was still submitted; the block is now deleted atomically.
Fixed the artifact viewer losing syntax highlighting when returning from the editor view, and returning to the wrong panel when exiting the artifact detail view.Fixed the headless `stream-json` `init` event advertising tools that are not available in your build.