繁體中文
想替 AI 配一位專門校稿或查資料的助手,現在可以用 agent.md 這種 Markdown 檔(用標題與一般文字寫內容的純文字格式)來定義 custom agent(自訂 Agent,替 AI 指定另一套角色與做事規則)。檔案開頭的 YAML frontmatter(放在檔案開頭、用欄位寫設定的區塊)可決定它是主助手、子助手、是否隱藏、是否沿用 MCP(讓 AI 連接外部工具的協定),以及可否執行指令;用 H1 標題也能切出給它的系統提示詞。這像替不同職務做名牌和工作守則;動態建立的子助手也會拿到同款名牌,搬到外部版本仍認得出來。/agents 從自訂角色切回預設角色的失敗問題也修好,不會表面切換、其實還困在舊人格。
剛看完幾輪回答,才發現想貼出去的是前一則?/copy 2 現在會把倒數第二則回答複製到剪貼簿(暫時放著、可貼到別處的地方)。數字 n 代表第 n 則最近回答;不加數字的 /copy 和 /copy 1 仍是最新一則。你可直接把那段貼進訊息、文件或筆記。像翻相簿時直接抽出前一張照片,不必先把最新照片複製走,再回頭手動選文字。
在大專案裡用 /codesearch 找一個函式名稱時,不必盯著空畫面等全部結果。搜尋結果現在會隨資料到達逐步顯示,還會告訴你目前找到了幾筆;你可以先從已出現的檔案判斷是否找對地方。如果方向不對,按 Esc 就能取消還在跑的搜尋。這像在倉庫找貨時,搬運員一找到一箱就先送到你面前,數量也一路更新;你發現拿錯訂單,隨時喊停即可。
你請 Agent 看剛下載或剛產生的暫存檔時,現在通常不會先被權限視窗打斷。CLI 已預設給 system temporary directory(系統暫存目錄,程式短暫放檔案的地方)讀取權限,並依 Windows、macOS、Linux 各自的位置正確處理。它只多開了「看」的門,沒有多開修改或刪除的門;像飯店櫃台直接給你行李寄放區的查看證,讓助手能讀到它剛放進去的收據,不必每次先打電話確認。
同一套自訂規則和路徑,現在每次被找到時都會排成固定順序。以前順序偶爾飄動,即使內容沒變,送進 prompt cache(提示詞快取,把已處理過的指示暫存再利用)看起來也像另一份資料,得重新處理。這次把排序定下來,讓快取能認出「還是原來那張清單」,少做無謂的重複準備。像把食譜卡永遠照同一順序放回抽屜,廚師下次一眼就知道沒換菜單,不必重新備料。
螢幕上的小毛病,這版一次掃掉好幾個。切到圖片 artifact viewer(成品檢視器,用來看 AI 產出檔案的畫面)時,iTerm2 這類偵測到但不能畫 Kitty 圖形的終端機,不會再噴出亂碼;第一次開啟檢視器也不會吞掉你按的 Esc。回答串流時,Markdown(用簡單符號排版的文字格式)重新排版不再把輸入框往上推得抖動;倒帶或重設步驟後,訊息列表標題仍在。登入驗證頁結束後也不會只剩空白畫面。像電影院換幕時,以前字幕、座椅和片頭偶爾各自滑走;現在換場仍把畫面與你的按鍵留在該在的位置。
程式卡住時,最怕它只丟一句無關的錯誤。print mode(列印模式,用 -p 把回答直接輸出、不開互動介面的模式)建立對話失敗時,現在會顯示真正的建立失敗原因,不再假稱「沒有啟用中的對話」。sandbox(沙箱,把程式關在受限空間裡執行的保護機制)先擋住指令時,即使 network proxy(網路代理,代替程式轉送網路請求的中介)停用,也會留下受阻紀錄,核准與拒絕的訊息也整理得更清楚。背景自動更新器則修掉雙重啟動競賽:同一個更新時段不會兩個行程各自再叫一個更新器。沒有正在執行的背景 shell 指令時,你重新綁定的 ctrl+b 快捷鍵也會生效。像值班室終於寫對事故單,也確保同一班只派一位維修員出車。
English
If you want an AI helper dedicated to proofreading or research, you can now define a custom agent with an agent.md Markdown file: ordinary text with headings. Its YAML frontmatter, a settings block at the very top of the file, can say whether it is the main agent, a subagent, hidden, inherits MCP (the protocol that lets AI connect to outside tools), or may run commands. Think of it as a job badge plus work rules for each role. Dynamically created subagents now get the same badge, so external builds recognize them too. /agents also correctly returns from a custom role to the default one instead of appearing to switch while leaving the conversation in the old persona.
Finished a few rounds of answers and realized the one you need to paste is the previous one? /copy 2 now copies the second-most-recent response to the clipboard, the temporary place from which you paste elsewhere. The number n means the n-th most recent response; plain /copy and /copy 1 still mean the latest. It is like pulling the prior photo from an album directly instead of copying the newest photo first and then selecting old text by hand.
When you use /codesearch to find a function name in a large project, you no longer have to stare at an empty screen until every result arrives. Matches now appear progressively, with a live count; if you searched in the wrong direction, Esc cancels the search while it is still running. It is like a warehouse worker bringing each box to you as soon as it is found while updating the tally, and stopping the moment you realize it is the wrong order.
When you ask an agent to inspect a file it just downloaded or generated temporarily, a permission dialog usually will not interrupt first. The CLI now grants read access by default to the system temporary directory, the place where programs briefly park files, and resolves that location correctly on each platform. This opens only the door to look, not to edit or delete. It is like a hotel desk giving you a viewing pass for the luggage area so the helper can read the receipt it just put there without calling for approval every time.
The same set of custom rules and discovered paths now arrives in a fixed order every time. Previously, the order could drift; even unchanged content then looked new to the prompt cache, which temporarily stores processed instructions for reuse, and had to be handled again. Stable sorting lets the cache recognize the same list. It is like returning recipe cards to a drawer in the same order, so the cook sees at once that the menu has not changed and need not prep everything again.
Several screen glitches disappear together in this release. When switching an image artifact viewer, the view used for AI-generated files, iTerm2-like terminals that are detected but cannot draw Kitty graphics no longer spew garbage characters; the first viewer opening also stops swallowing your Esc key. During response streaming, Markdown, a plain-text formatting language, no longer reflows in a way that jitters the input box upward, and the message-list header survives rewinding or resetting steps. The screen also no longer goes blank after the sign-in page. Like a cinema changing scenes: subtitles, seats, and the opening title used to slide away separately; now the scene changes while the image and your keypress stay put.
When a program fails, the worst outcome is an error that points somewhere else. If print mode, the -p mode that outputs an answer without an interactive interface, cannot create a conversation, it now shows the real creation failure rather than claiming there is no active conversation. When the sandbox, the protective restricted space for running programs, blocks a command first, it records that block even with the network proxy off; approval and denial messages are cleaner too. The background auto-updater also fixes a double-spawn race, so two processes cannot each launch an updater in the same update window. When no background shell commands are running, a remapped ctrl+b shortcut is now respected. It is like a control room finally filing the right incident report and sending only one repair crew for one shift.
原始 CHANGELOG
Original changelog自訂 Agent 改用 Markdown,角色切換不再卡住
Custom Agents (Markdown Format). Added support for defining custom agents using Markdown files (`agent.md`) with YAML frontmatter and H1-delimited system prompts. Markdown agents support `mainAgent`, `subagent`, `hidden`, `inheritMcp`, and `commandExecutionPolicy` frontmatter fields for fine-grained control over agent behavior. Dynamically defined subagents (via `define_subagent`) now also write Markdown format so they resolve correctly on external builds.
Improved support for markdown-based custom agents so custom agent management and selection behave more consistently.
Fixed switching from a custom agent back to the default agent via `/agents`, which previously failed silently and left the conversation stuck on the custom agent's persona.
Original changelog`/copy` 能挑選較早的回答
Added an optional index argument to `/copy` so `/copy <n>` copies the n-th most recent response to the clipboard, while `/copy` and `/copy 1` still copy the latest.
Original changelog程式碼搜尋邊找邊顯示,也能按 Esc 停下
Improved `/codesearch` to render results progressively as they stream in, showing a live count while loading and letting you cancel an in-flight search with `Esc` instead of blocking until the whole search finishes.
Original changelog讀取暫存檔不再先跳權限詢問
Improved default file access by granting read access to the system temporary directory out of the box, resolved correctly per platform, so agents no longer trigger permission prompts when reading temporary files.
Original changelog自訂規則排序固定,快取不再白白失效
Improved customization discovery by sorting rules and discovered paths deterministically, preventing unstable prompt ordering and needless prompt-cache misses.
Original changelog終端機畫面修順,登入後不再一片空白
Fixed the artifact viewer emitting garbage escape bytes when cycling to image mode on terminals that are detected but cannot actually render Kitty graphics, such as iTerm2.
Fixed the first keystroke (such as `Esc`) being dropped when opening the first artifact view on some non-Kitty terminals.
Fixed conversation jitter and a stranded input box during streaming so transient markdown reflow no longer shifts the pending line and input box upward.
Fixed the message list dropping its header when rewinding or resetting conversation steps.
Fixed the screen going blank after the authentication page.
Original changelog背景更新與受阻指令的錯誤資訊更可靠
Improved overall reliability and stability across the CLI with additional hardening and fixes for intermittent failures in background tasks, print mode, and interactive flows.
Fixed a crash when a command was blocked by sandbox permissions before its output was captured, and cleaned up the permission approval and denial messages.
Fixed print mode (`-p`) surfacing the real conversation-creation failure instead of a misleading "no active conversation" error.
Fixed a background auto-updater double-spawn race where two processes could each spawn an updater within a single update window.
Fixed sandbox error reporting so blocked actions are recorded even when the network proxy is disabled.
Fixed the `ctrl+b` shortcut being hardcoded to background shell commands even when none were running, so a remapped `ctrl+b` is now respected whenever there are no running shell commands in the conversation.