問題修復
畫面不再突然靜止，分頁標題也少抖一下
Claude 還在工作，眼前畫面卻整片不更新，最讓人拿不定主意：該等還是重開？少見的內部版面錯誤曾讓互動工作階段（session，使用 Claude 的一次連續對話）完全停止重畫，但程式仍在跑；現在會避開這個死角。這像直播攝影機卡在同一格，棚內的人仍在繼續說話，觀眾卻看不見後續。另一項小修正針對終端機標題的忙碌旋轉符號；部分終端機的分頁列會因此抖動，現在換了較穩的符號。
繁體中文
問題修復
Windows 從上層資料夾啟動，也找得到 Git
在 Windows 用 Claude Code 管理專案時，明明裝了 Git，卻被說找不到它，接下來的版本控制工作就全斷了。問題出在 Claude Code 從 Git 安裝資料夾的上層資料夾啟動時，git 或 Git Bash 可能找不到。這次修正的是尋路方式，不是替你安裝另一套 Git。像你知道工具放在某棟大樓，卻因為從大樓外的路口出發而沒找到入口；現在從那個上層路口出發，也能順著路走到 Git。
新功能
模型選擇不倒退，檔案寫入規則也講清楚
你用 /model 換過模型，再打 /tui 回文字介面，不該被偷偷送回上一個模型；現在選擇會保住。檔案操作也調整了：較新的模型使用 Write 工具時，即使這次工作階段還沒讀過既有檔案，也可覆寫它，規則已和 Edit 工具一致；舊模型仍得先讀再寫。這像你在車站換好月台，轉頭看時刻表卻被帶回舊月台，現在不再發生。Pro、Max、Team 第一次使用提示裡「自動模式會稍微更貴」的過時說明也移除了，別再把舊提醒當現況。
新功能
跨工作階段傳話看得見，遠端續聊不洩題
兩個工作階段（session，使用 Claude 的一次連續對話）互傳訊息時，第一個工作階段在安裝或升級後偶爾沒有收件匣，訊息也縮成一行；現在收件匣會建立，寄件者和內容直接顯示。傳到另一台機器的 Remote Control 工作階段，寄件者則會顯示你的遠端控制工作階段名稱。另一面更重要：連著 Remote Control 打 /resume 時，剛恢復的對話標題或歷史不會再漏進目前連線的工作階段。像辦公室便條終於有收件盤、署名和全文，但隔壁會議室的舊紀錄仍鎖在門內。
問題修復
自架 runner 不再無端失敗或提早收工
自己架 claude self-hosted-runner 跑工作時，一個根本不會推送的 repository，其 checkout 掛鉤（特定事件發生時被觸發的小程式）失敗，竟能讓每個全新 runner 的工作階段都失敗。現在這種 repository 會跳過並給警告，不再拖垮整次工作。背景任務完成到後續回合開始之間，runner 以前也可能誤把工作階段結束；這段交棒空隙已補上。像接力隊遇到不參賽的器材櫃壞了，過去全隊被迫退賽，現在繞過它並留下告示，接棒的人也不會還沒拿到棒就被判離場。
問題修復
清理時保住專案記憶與開發中的外掛
專案裡有 memory 資料夾，或你正用符號連結（symlink，指向另一個位置的捷徑檔）開發外掛時，最怕背景清理把該保留的東西當垃圾。先前工作階段清理可能刪掉專案 memory 資料夾裡的內容；背景的外掛快取（cache，把算過或查過的結果暫存起來直接用）清理，也可能在唯一版本是符號連結開發副本時刪掉該外掛快取。兩種誤刪都修正了。像管理員整理辦公室時，把貼在專案櫃裡的筆記和工作室裡正在使用的開發工具一起丟掉；現在不會再把這兩類東西清走。
問題修復
市集設定各自算帳，不再串到別層標頭
同一個 marketplace 項目在不同優先層的設定裡各有一份時，較高層重新定義它，本來應該整份換掉；舊行為卻可能把較低層的自訂 headers 一起帶過來。現在 marketplace 項目會以完整項目合併，不再東拼西湊。這對管理多組設定的人很重要：你改的是高優先層那張表單，就不會暗中繼承另一張表單的標頭。像公司把同名供應商的新版合約放到最上層，過去卻把舊合約頁首頁尾黏上去；現在會採用新版合約的完整內容。
問題修復
技能提示不重複，本機指令也不會被雲端蓋掉
叫用一個 skill（技能，教 Claude 怎麼處理特定工作的說明）後，deferred tools 提醒有時會送給模型兩次，現在只送一次。更關鍵的是，從 claude.ai 同步下來的技能收緊了：它不能遮蔽你本機指令或 MCP（Model Context Protocol，讓 AI 連接外部工具的協定）提示，描述會先清理並標示來源；在你的機器上，技能本文也不會執行 ! 指令或展開 @ 檔案。像外來的操作手冊可以放上書架參考，卻不能把你桌上的按鈕貼住，更不能自行翻出抽屜裡的文件。
新功能
憑證失效早點說，整理對話也看得到進度
Vertex AI 的 Google Cloud 憑證（credential，證明你可登入服務的資料）過期或缺漏時，過去會重試好幾分鐘才失敗，現在幾秒內就會回報。你不用盯著像壞掉的電梯按鈕一直亮著，卻不知道它到底有沒有在動。Claude 整理過長對話的 compaction（壓縮，把前面內容濃縮以騰出對話空間）也補了更實在的訊號：重試倒數與卡住提示會在進行中出現，不再只剩一條進度列。前者讓錯誤快點浮面，後者讓等待中的人知道它是在重試還是真的停住。
English
BUGFIX
The screen keeps redrawing, and terminal tabs wobble less
Claude may have kept working while the screen stopped updating, leaving you to wonder whether to wait or restart. A rare internal layout error could halt redraws in an interactive session even though its process was still running; that failure is now avoided. It was like a live camera frozen on one frame while the people in the studio kept talking. A smaller change replaces busy-spinner glyphs in the terminal title, reducing tab-bar jitter in some terminals.
BUGFIX
Windows can find Git even when Claude starts above its folder
On Windows, Claude Code could say Git was missing even when it was installed, cutting off version-control work. The failure happened when Claude Code was launched from a parent folder of Git's installation: git or Git Bash might not be found. This changes how it finds the existing installation; it does not install another Git. It was like knowing a tool is in a building but missing the entrance because you started at a nearby intersection. Now that route reaches Git too.
FEATURE
Model choice no longer rolls back, and Write rules are clearer
After changing models with /model, opening /tui should not quietly return your session to the earlier model; it now keeps your choice. File handling also changed: newer models may use Write to overwrite an existing file they have not read in this session, matching Edit; older models must still read first. Think of changing train platforms, then looking at the timetable and being sent back to the old one. An outdated first-use note claiming auto-mode sessions cost slightly more has also been removed for Pro, Max, and Team plans.
FEATURE
Cross-session messages are visible, without leaking resumed chats
When two sessions exchange messages, the first session after an install or upgrade could sometimes have no inbox, and messages were collapsed to one line. Now an inbox is created and both sender and body appear inline. For messages sent to a Remote Control session on another machine, the sender shows your Remote Control session name. More importantly, using /resume while connected through Remote Control no longer leaks a resumed conversation's title or history into the connected session. It is like office notes gaining an inbox, signature, and full text while old records in the next meeting room stay behind a locked door.
BUGFIX
Self-hosted runners stop failing for irrelevant repos or ending early
When running claude self-hosted-runner yourself, a checkout hook failing for a repository the session would never push to could make every fresh runner session fail. Such a repository is now skipped with a warning instead of taking down the work. A runner could also end a session in the handoff gap after a background task finished but before its follow-up turn began; that gap is now covered. It is like a relay team being disqualified because a broken equipment locker was never part of the race. Now they pass it with a notice, and the next runner is not dismissed before receiving the baton.
BUGFIX
Cleanup keeps project memory and in-development plugins intact
If you keep a project's memory folder so Claude retains context, or develop a plugin through a symlink (a shortcut file pointing somewhere else), cleanup must not mistake active material for trash. Session cleanup could delete contents inside the project's memory folder. Background plugin-cache cleanup could also delete a plugin's cache when its only version was a symlinked development checkout. Both accidental deletions are fixed. It was like an office manager clearing out notes in a project cabinet and active tools inside the workshop; now those live materials are recognized.
BUGFIX
Marketplace settings no longer inherit headers from another tier
When the same marketplace entry existed in settings tiers with different precedence, redefining it in a higher tier should have replaced the whole entry. Instead, custom headers from another tier could be inherited. Marketplace entries now merge as whole entries, rather than as a mix of pieces. This matters if you manage several layers of settings: changing the higher-precedence form will no longer quietly bring along headers from another form. It was like filing a supplier's new contract on top, then accidentally gluing the old contract's letterhead onto it. Now the full new contract is used.
BUGFIX
Skill reminders stop duplicating, and cloud skills cannot override local controls
After invoking a skill (instructions that tell Claude how to handle a particular kind of work), the deferred-tools reminder could sometimes be sent to the model twice; it now appears once. More importantly, skills synced from claude.ai are restricted: they cannot shadow local commands or MCP (Model Context Protocol, a protocol that connects AI to external tools) prompts, their descriptions are sanitized and labeled, and their bodies cannot run ! commands or expand @ files on your machine. An outside manual may sit on your shelf, but it cannot cover the buttons on your desk or open files from your drawers by itself.
FEATURE
Expired credentials fail fast, and compaction shows its progress
When Google Cloud credentials for Vertex AI are expired or missing, Claude Code used to retry for minutes before failing; it now fails within seconds. You no longer have to stare at a lit elevator button without knowing whether anything is moving. Compaction, which condenses an overlong conversation to make room, also shows more useful signals: its retry countdown and a stall hint now appear while it runs instead of leaving only a progress bar. One change surfaces an error sooner; the other tells a waiting person whether it is retrying or has stopped.
原始 CHANGELOG
Original changelog畫面不再突然靜止，分頁標題也少抖一下
Fixed interactive sessions that could stop redrawing entirely, while the process kept running, after a rare internal layout error
Updated terminal title busy-spinner glyphs to reduce tab-bar jitter on some terminalsOriginal changelogWindows 從上層資料夾啟動，也找得到 Git
Fixed `git` / Git Bash not being found on Windows when Claude Code is launched from a parent folder of the git installationOriginal changelog模型選擇不倒退，檔案寫入規則也講清楚
Fixed `/tui` reverting the session to an earlier model when `/model` had been changed since the last response
Changed the Write tool so newer models can overwrite an existing file they haven't read this session, matching the Edit tool's rules; older models still require the read first
Removed the outdated note about auto mode sessions costing slightly more from the first-use notice for Pro, Max, and Team plansOriginal changelog跨工作階段傳話看得見，遠端續聊不洩題
Fixed cross-session messaging sometimes starting without an inbox in the first session after install or upgrade
Fixed Remote Control `/resume` while connected leaking the resumed conversation's title or history into the connected session
Improved cross-session messages: the sender and body now display inline instead of a collapsed line, and messages to Remote Control sessions on other machines show your Remote Control session name as the senderOriginal changelog自架 runner 不再無端失敗或提早收工
Fixed `claude self-hosted-runner` sessions failing on every fresh runner when the `checkout` hook fails for a repository the session doesn't push to; that repository is now skipped with a warning
Fixed self-hosted runners ending sessions in the gap between a background task finishing and the follow-up turn startingOriginal changelog清理時保住專案記憶與開發中的外掛
Fixed session cleanup deleting contents inside a project's memory folder
Fixed background plugin-cache cleanup deleting a plugin's cache when its only version is a symlinked development checkoutOriginal changelog市集設定各自算帳，不再串到別層標頭
Fixed a settings-merge issue where a marketplace entry redefined in a higher-precedence settings tier could inherit another tier's custom headers; marketplace entries now merge as whole entriesOriginal changelog技能提示不重複，本機指令也不會被雲端蓋掉
Fixed the deferred-tools reminder occasionally being sent to the model twice after a skill invocation
Hardened skills synced from claude.ai: they no longer shadow local commands or MCP prompts, their descriptions are sanitized and labeled, and on your machine their bodies don't run `!` commands or expand `@` filesOriginal changelog憑證失效早點說，整理對話也看得到進度
Improved Vertex AI credential handling: expired or missing Google Cloud credentials now fail within seconds instead of retrying for minutes
Improved compaction progress: the retry countdown and stall hint now appear during compaction instead of only a progress bar