發布日期：2026-08-18
在終端機裡打提示詞時，現在可以開啟即時拼字檢查。只要在個人設定檔中加入 "spellcheck": true，系統會使用你電腦裡安裝的 aspell、hunspell 或 ispell 等拼字檢查工具，在你邊輸入的同時即時標記拼錯的單字。若要啟用，可於 ~/.claude/settings.json 中設定：
{"spellcheck": true}如此一來，在送出指令之前就能即時發現打錯的指令關鍵字或單字，不必等送出後才發現出錯。
寫程式時若是語言伺服器在對話中斷線又重連，你不再需要心疼遺失的快取（cache，把算過或查過的結果暫存起來直接用）與額外消耗的費用。過去在同一個工作階段（session）中，語言伺服器只要短暫斷開連線或重新接上，整個提示詞快取就會被徹底清空失效，迫使後續對話重新計算所有標記。現在這項問題已修正，連線中斷或重連時會妥善保留既有快取。
在多行提示詞中輸入斜線指令或關鍵字時，高亮標記不再往前或往後歪掉一兩個字元。以前在輸入框打入多行文字時，斜線指令、關鍵字以及標註（@mention，用小老鼠符號提及檔案或符號的語法）的高亮色彩有時會出現字元位移，看起來十分錯亂；同時，若在 Claude 正在輸出回覆的當下執行斜線指令，終端機也不會再把字元印成像 & 這樣的 HTML 實體編碼。
在終端機介面閱讀多層 Markdown 清單時，第 3 層或更深的項目現在會正確對齊，過長項目的換行也會使用懸掛縮排（hanging indent，讓換行後的第二行文字對齊上一行開頭而非標記點的排版方式）。過去，這些深層項目的對齊位置常常歪斜錯位，文字自動換行後也不會往內縮排。
在權限授權對話框的留言欄位裡按 Shift+Tab，現在只會單純關閉該欄位，絕不會手滑把整個工作階段（session）的編輯權限直接送出去。以前在權限提示的備註框按下 Shift+Tab，會誤觸核准並賦予整個工作階段範圍的編輯權限；現在這個危險行為已被修正。此外，授權對話框的說明文字與「不再詢問」選項，現在保證與實際授權涵蓋的範圍完全一致；若對話框無法完整顯示內容，系統會主動隱藏「不再詢問」選項以策安全。
筆記本（notebook）儲存格無法讀取時，刪除或替換前的確認視窗不再一聲不吭地留下一片空白。以往當系統無法讀取 Jupyter 筆記本或特定儲存格的原始內容時，核准對話框會直接略過既有內容，讓使用者誤以為該儲存格原本就是空的；現在對話框會明確標註無法讀取的原因，幫助你在確認修改前了解真實狀況。
在各類選單對話框裡快速連按方向鍵與 Enter，你選中的項目就是游標最後停下的那一項。以前如果你的手速非常快，按下方向鍵切換選項後立刻按下 Enter 確認，介面偶爾會跟不上你的按鍵節奏，導致送出的是前一個高亮停頓的舊選項，容易造成手滑選錯；現在對話框的鍵盤事件處理更加敏捷，能精準捕捉你瞬間移動到的目標。
當你在工作階段（session）中呼叫 Agent 工具卻省略指定型態時，現在會直接收到可用代理的清楚清單。以前在某些不支援全能預設代理的特殊環境中，系統仍會聲稱有預設代理可用，導致執行時碰壁；現在若省略了 subagent_type 參數，它會直接回報明確的錯誤訊息，並條列出目前環境中實際可用的代理清單。
跨工作階段（session）傳送訊息時，文字量如果超出傳輸上限，SendMessage 現在會當場拒絕收件。以前在不同工作階段之間傳遞超大型訊息時，系統會因為資料量過大而默默把訊息丟棄，呼叫端完全收不到任何警告或報錯，造成資料在不知不覺中遺失；現在系統會在最前端直接攔截並主動報錯，明確告知訊息過大無法送達，讓你能提早拆分或縮短內容。
當你在背景執行 /ultrareview 或 /autofix-pr 等雲端工作階段（session）時，記憶體與 CPU 使用量已改善。以往背景任務只要收到任何進度更新，系統就會把累積的事件串流從頭重新掃描並重新渲染一次；現在不再這樣做。
在 macOS 與 Linux 原生版本中使用內建搜尋，遇到病態比對樣式時現在會立即中止，不再把記憶體吃到乾涸。內建的 grep 搜尋工具迎來兩項重要改進：首先是遇到病態比對樣式時，系統會迅速判定失敗並中斷，避免耗盡系統記憶體；其次是當你搭配 -m N 限制匹配次數，並使用 -A 或 -C 抓取周圍上下文時，輸出的前後行數現在完全正確。
當對話長度觸及上下文上限，而你關閉了自動壓縮功能時，現在會看見明確的提示。以前如果使用者關閉了對話自動壓縮（auto-compact），一旦上下文長度達到上限，系統只會拋出冷冰冰的上下文上限錯誤，讓人摸不著頭緒；現在錯誤訊息會主動說明當前已停用自動壓縮，並明確指引你輸入 /config 指令前往設定介面將它重新開啟。
習慣使用 Vim 模式編輯的開發者，按下 ctrl+o 查看詳細對話紀錄或關閉側邊面板後，NORMAL 模式與游標位置都會原封不動保留下來。在過去的版本中，只要你在終端機中切換完整逐字稿檢視（按 ctrl+o 開啟或關閉），或者關閉任何彈出面板，系統就會重設 Vim 的狀態，導致游標跑位或跳出 NORMAL 模式，打斷編輯節奏；現在切換檢視後狀態會完整保留，讓你流暢地繼續輸入。
重新連回仍有未結任務的工作階段（session）時，按 ctrl+t 展開的任務清單不會再強制收合。在以往的版本中，即使當前工作階段還有未完成的任務，每當你重新連線或重新啟動進入該工作階段時，展開的任務清單總是會預設縮回摺疊狀態；現在系統會記住展開狀態，讓你一開工就能立即掌握所有進行中的任務進度。
背景自動更新完成後，提示詞底部的狀態列（status line，顯示在終端機最底下的狀態列）現在會老老實實秀出提示，提醒你重新啟動程式以套用新版。過去當 Claude Code 在背景默默下載並安裝好新版本時，最底部的提示頁尾往往漏掉顯示「Update installed」（更新已安裝）的重啟通知，導致使用者完全不知道背後已經升級、依然繼續跑著舊版；現在通知會如實出現，讓你能即時重啟換上最新版。
透過 claude rc 啟動遠端控制時，現在會像一般互動式啟動一樣，在第一時間確實檢查企業閘道器是否可用。在企業資安環境中，互動式啟動 Claude Code 時原本就會嚴格檢驗閘道器連線狀態；然而過去使用遠端控制指令 claude rc 啟動時卻遺漏了這層檢查，容易引發連線異常。現在兩者的啟動檢驗標準完全拉齊，讓遠端連線同樣具備企業級的穩定度。
在 VS Code 裡開啟多個 Claude 面板時，重新載入或還原視窗，焦點不再於各分頁之間胡亂跳動。以往在同一個視窗中若開了多個 Claude 面板分頁，當視窗重新載入（reload）或從背景還原時，編輯器焦點會在各個分頁之間自己亂跳，打斷工作思路；現在分頁焦點在視窗重載後會穩定留在原處，避免視窗切換帶來的困擾。
You can now turn on real-time spelling feedback while entering a terminal prompt. When you enable the optional spellcheck setting, it uses your locally installed aspell, hunspell, or ispell tools to underline misspelled words as you type. To turn it on, add it to your ~/.claude/settings.json:
{"spellcheck": true}This catches misspellings and prompt typos before you send the message.
When language servers disconnect or reconnect in the middle of a session, your prompt cache will no longer be wiped clean. Previously, any brief hiccup or reconnection with a language server invalidated the entire prompt cache, forcing subsequent turns to re-process tokens and waste credits. The cache is now safely preserved across mid-session language server reconnections.
Multi-line prompt input highlighting and slash command characters now render cleanly. Previously, syntax highlights for slash commands, keywords, and @ mentions could appear shifted by one or more characters across multi-line inputs. Additionally, running a slash command while Claude was actively streaming a response showed raw HTML entities instead of real characters. Both visual bugs are now resolved.
When you read deeply nested Markdown lists in the terminal UI, third-level and deeper items now align properly, and wrapped lines use a hanging indent. Previously, lists nested at depth 3 or deeper suffered from misalignment, and list items that wrapped across multiple lines lacked clean indentation.
Permission prompts are now much safer and more accurate about what you grant. Previously, pressing Shift+Tab inside the permission comment field accidentally approved the edit and granted session-wide edit permissions instead of just closing the comment box. In addition, permission dialog text and "don't ask again" options now strictly match the true scope of the grant, and the "don't ask again" option is withheld whenever the full contents cannot be displayed.
Notebook cell delete and replace approval dialogs now state why content is missing when a read fails. Previously, if Claude could not read a notebook file or an individual cell, the approval dialog silently omitted the existing content, making it look as though you were deleting an empty cell. The dialog now explicitly explains why the existing content could not be read so you can make an informed decision before approving the change.
Rapidly navigating dialogs with arrow keys and Enter now picks the correct option. Previously, if you tapped an arrow key and immediately pressed Enter in quick succession, the dialog sometimes lagged behind and selected the previously highlighted choice instead of the one you just moved to. Dialog keyboard handling is now properly synchronized to catch fast keystrokes.
When you omit subagent_type while calling the Agent tool in a session without a general-purpose default, you now get a clear error listing the agents available to you. Previously, in sessions where the standard general-purpose agent was not supported, omitting the subagent_type parameter falsely advertised that a default was present.
SendMessage now explicitly rejects oversized messages before delivery rather than dropping them into the void. Previously, when sending data across different sessions, messages exceeding the cross-session payload limit were silently discarded without any warning or error, leaving developers wondering where the message went. The system now validates size constraints up front and immediately returns a refusal error so you can trim or chunk the content.
Running cloud sessions like /ultrareview or /autofix-pr in the background now uses memory and CPU more efficiently. Previously, whenever a background cloud session updated, its entire event stream was re-scanned and re-rendered from scratch. That no longer happens on every update.
The embedded grep tool in native macOS and Linux builds is now much safer and more accurate. When encountering pathological search patterns, grep now fails fast instead of exhausting system memory. Furthermore, combining -m N (maximum match count) with -A (after context) or -C (surrounding context) flags now prints the exact required context lines without dropping or duplicating them.
The context-limit error message now gives actionable advice when auto-compact is disabled. Previously, when conversations ran out of context window space with auto-compact turned off, the error gave no hint as to why compacting didn't happen. The system now explicitly warns you that auto-compact is off and directs you to run /config to turn it back on.
Vim mode now reliably preserves NORMAL mode and your exact cursor position across UI toggles. Previously, pressing ctrl+o to toggle the detailed conversation transcript or closing an open panel would reset your Vim mode state and lose your cursor placement. The editor now remembers your active NORMAL mode state and cursor coordinates throughout these view transitions.
The expanded task list (ctrl+t) no longer collapses on session restart if you still have active tasks. Previously, whenever you resumed or relaunched a session with unfinished tasks, the task list UI always defaulted to a collapsed state. It now properly starts expanded so you immediately see the status of all pending items upon resumption.
The prompt footer now reliably shows the "Update installed" restart banner after background auto-updates. Previously, when an update finished installing quietly in the background, the footer notice failed to pop up, leaving you in the dark that a restart was needed to activate the new version. The notification now appears as expected so you know exactly when to restart and enjoy the latest release.
Launching Remote Control via claude rc now performs the exact same enterprise-gateway availability check as interactive startup. In corporate environments, standard interactive sessions verify gateway accessibility right at startup, but claude rc previously skipped this step, leading to inconsistent behavior. The check is now unified across both entry points to ensure reliable connectivity under enterprise policies.
VS Code users with multiple Claude panels open will no longer experience phantom focus jumping when reloading or restoring windows. Previously, restoring a workspace containing several Claude tabs caused the cursor focus to spontaneously hop between tabs on its own. Focus is now cleanly preserved upon window reloads.
Added an optional `spellcheck` setting that underlines misspelled words in the prompt input as you type, using your installed `aspell`, `hunspell`, or `ispell`Fixed whole-prompt-cache invalidation when a language server disconnected or reconnected mid-sessionFixed prompt input highlights (slash commands, keywords, mentions) appearing shifted by one or more characters in some multi-line prompts
Fixed slash commands run while Claude is responding showing HTML entities instead of the actual charactersFixed nested markdown list items misaligning at depth 3+ and added a hanging indent to wrapped list items in the terminal UIFixed Shift+Tab inside the permission prompt's comment field approving the edit and granting session-wide edit permission instead of closing the field
Improved permission dialogs: display text and "don't ask again" options now always match what a grant would cover, and "don't ask again" is withheld when contents cannot be fully displayedFixed notebook cell delete/replace approval dialogs silently omitting the existing cell content when the notebook or cell could not be read; the dialog now says whyDialogs: arrow keys and Enter pressed in quick succession now select the option you navigated to instead of the previously highlighted oneFixed the Agent tool advertising a general-purpose default in sessions where that agent is unavailable: an omitted `subagent_type` there now gets a clear error listing the available agents`SendMessage` now refuses messages too large for cross-session delivery up front instead of silently dropping themImproved memory and CPU usage while cloud sessions such as `/ultrareview` or `/autofix-pr` run in the background — their event streams are no longer re-scanned and re-rendered on every updateImproved the embedded `grep` in native macOS/Linux builds: pathological patterns now fail fast instead of exhausting memory, and `-m N` with `-A/-C` prints correct contextImproved the context-limit error to say when auto-compact is off and point to `/config` to re-enable itVim mode: NORMAL mode and cursor position are now preserved when toggling the detailed transcript (ctrl+o) or closing a panelFixed the expanded task list (`ctrl+t`) always starting collapsed when resuming or relaunching into a session that still has open tasksFixed the prompt footer not showing the "Update installed" restart notice after a background auto-updateRemote Control: `claude rc` now applies the same enterprise-gateway availability check as interactive startup[VSCode] Fixed focus jumping between open Claude tabs on its own when a window with several Claude panels is restored or reloaded