發布日期：2026-10-11
平日盯著終端機跟 AI 密切協作時，最怕畫面捲動卡頓或排版突然亂掉。
現在程式正式支援 Rex 終端機與 OSC 7501 控制序列。終端機裡的工具呼叫群組在有新操作加入時會持續維持展開，長對話中按下 Ctrl+E 展開或收折思考過程也更加迅速。如果提示詞或引號裡的超連結因為文字過長而折行，點擊時依然能正確連向完整網址。模型回覆裡的行內標籤不會再破壞版面排版。在工作階段內用來恢復舊工作階段的選單，若設定為不分組，現在會跟歡迎畫面的不分組排版一致。
替專案撰寫自動化擴充時，不同目錄的自訂規則現在能精準到位。
如果你在儲存庫的子資料夾啟動 grok，該子目錄下的專案掛鉤（hook，特定事件發生時被觸發的小程式）現在會正確載入。連往本機 localhost 的 HTTP 掛鉤也能正常運作，若網址填寫錯誤，載入時就會立刻報錯。終端機底部的指令狀態列（statusLine，終端機最底部的即時資訊列）腳本在回合執行期間，也能順利取得當前回合的提示詞識別碼與開始時間。
當你想換個更合適的模型解題時，設定結果現在會老實保存。
以前使用 /model 切換模型若中途失敗，頁尾顯示和儲存的預設模型仍會被改掉，現在切換失敗會維持原狀。切換成功後重新啟動 grok，也不會再倒退回先前的舊模型。如果你使用 grok -p --effort 指定思考力度，程式會直接拒絕無效的力度數值；若目前使用的模型忽略這項參數，也會明確發出警告。
整理對話紀錄或搜尋專案內容時，操作遇到阻礙不再默默失敗。使用匯出或複製功能時，若目標檔案已經存在，程式會主動拒絕覆寫並顯示清楚的提示訊息，避免原有內容被蓋掉。在終端機貼上超過 50 MB 的超大圖片時，畫面會跳出明確訊息拒絕接收，不再無聲把圖片吃掉。當工具執行 grep 搜尋或使用 list_dir 列出資料夾時，若遇到搜尋逾時、執行失敗或目錄無法讀取，現在會明確回報錯誤，不再假裝查無資料而回傳空白。以前遇到這些狀況只能自己瞎猜到底有沒有送出或到底查完了沒，現在每一處阻礙都有清楚說明。
當你決定按下中斷鍵或退出程式，手邊正在跑的任務現在會乾淨收尾。
以前退出 grok 或發送中斷訊號後，部分背景指令仍會私下繼續佔用資源，現在結束時會老實終止所有由它啟動的子行程。在終端機按下取消時，畫面也會清楚印出究竟是哪一道指令被停止。如果你使用 /workflow stop 停止工作流程，程式也能精準鎖定當前真正處於執行中的任務，不會誤停其他進程。
放手讓 AI 在背景處理複雜任務時，需要你點頭授權的關鍵卡片不再跟你玩捉迷藏。
背景子代理（subagent，負責分工處理特定子任務的代理程式）送出的審核卡片，在父回合結束時不會再無故消失。跳出的審核通知現在會明確標示是哪一個工作階段需要你批准。在極簡模式下如果權限提示與隱藏的計畫或取消面板同時出現，畫面會確實呈現該權限提示。每當審核卡片浮現，程式也會自動關閉區塊檢視器，避免卡片內容遭到遮擋。
清理專案的多重副本時，重要程式碼與進行中的作業現在有更嚴密的保護。
使用工作樹（worktree，Git 讓同一專案平行開多份副本的功能）時，執行 grok worktree rm 或 gc 垃圾回收清理，若發現有開啟中的工作階段正使用該副本，指令會直接拒絕刪除。grok worktree rm 也設下了嚴格的安全防線，一律拒絕刪除任何 Git 儲存庫、.git 資料夾或任何非 grok 自行建立的目錄，不再因為手滑誤打指令而釀成災難。
跟 AI 進行長篇對話或調用大量工具時，記憶空間被塞爆的窘境現在有了自動排解機制。
如果工具回合內容過長且包含工具使用後的自訂掛鉤輸出，程式會主動壓縮摘要，不再直接撞上提示詞長度上限。當遇上脈絡溢位（context overflow，對話總長度超出模型單次能處理的容量）錯誤時，程式會自動觸發壓縮，不再讓當前回合直接當掉。為了避免無休止的壓縮消耗資源，自動壓縮只要連續三次填滿脈絡就會主動停下，不再陷入死迴圈。
利用自動化腳本批次餵入提示詞或串接外部服務時，隱形字元造成的離奇錯誤現在已徹底排除。在無頭模式（headless，沒有終端機視覺介面的執行模式）下，如果提示詞檔案開頭帶有微軟檔案常見的 BOM 位元組順序記號，程式會主動濾除，不再把多餘標記送給模型，也不會弄壞 JSON 提示詞檔。當外部工具透過 MCP（Model Context Protocol，讓 AI 連接外部工具的協定）回傳結果時，即便內容夾雜無效字元，也不會再導致工具呼叫逾時卡死。以前這些肉眼看不到的編碼符號常在暗處作怪，讓排程任務莫名中斷或卡在等待畫面，現在這類字元解析問題已經修好。
Working closely with AI in the terminal, nothing is more frustrating than scrolling stutter or suddenly scrambled formatting.
The application now officially supports the Rex terminal and OSC 7501 control sequences. Tool call groups inside the terminal stay expanded when new operations are added, and pressing Ctrl+E during long conversations expands or collapses thinking processes much faster. When hyperlinks inside prompts or quotations wrap across lines due to length, clicking them still reliably opens the complete URL. Inline tags within model responses will no longer disrupt layout alignment. In addition, the in-session menu for resuming earlier sessions, when set to ungrouped, now mirrors the ungrouped layout on the welcome screen.
When writing automated extensions for a project, custom rules across different directories now apply with precision.
If you launch grok from a subdirectory of a repository, project hooks under that subdirectory will now load correctly. HTTP hooks connecting to local localhost also operate properly, immediately raising an error during loading if a URL is incorrect. Furthermore, command statusLine scripts at the bottom of the terminal can smoothly retrieve the prompt ID and start time of the current turn while the turn runs.
When switching to a more suitable model to solve a problem, your configuration is now reliably preserved.
Previously, if switching models via /model failed mid-operation, the footer display and saved default model were still overwritten; now, a failed switch leaves the existing state intact. After a successful switch, restarting grok no longer reverts to the previous model. If you use grok -p --effort to specify reasoning effort, the program directly rejects invalid effort values, and it explicitly issues a warning if the active model ignores this parameter.
When organizing conversation logs or searching project contents, blocked operations no longer fail silently. When using export or copy features, if the destination file already exists, the program proactively refuses to overwrite it and displays a clear prompt to keep existing content intact. Pasting an oversized image exceeding 50 MB into the terminal triggers an explicit rejection message instead of silently swallowing the file. When tools execute a grep search or use list_dir to list folders, any search timeout, execution failure, or unreadable directory now explicitly reports an error instead of returning blank output as if nothing was found. Where you previously had to guess whether an operation was dispatched or finished, every obstacle now comes with clear feedback.
When you hit the interrupt key or exit the program, active tasks now wrap up cleanly.
Previously, after quitting grok or sending an interrupt signal, certain background commands continued consuming resources behind the scenes; now, exiting faithfully terminates all child processes spawned by the session. When you trigger a cancellation in the terminal, the screen explicitly displays which command was stopped. If you use /workflow stop to halt a workflow, the program accurately targets tasks currently running without mistakenly terminating other processes.
When leaving AI to handle complex tasks in the background, the key approval cards requiring your sign-off will no longer play hide-and-seek.
Review cards submitted by background subagents no longer vanish without reason when a parent turn ends. Pop-up authorization notifications now explicitly indicate which session requires your approval. In minimal mode, if a permission prompt coincides with a hidden plan or cancel panel, the screen reliably renders the permission prompt. Whenever a review card appears, the program also automatically closes the block viewer to prevent card content from being obscured.
When cleaning up multiple project copies, essential code and active work receive tighter safeguards.
When managing worktrees, running grok worktree rm or gc garbage collection cleanup will directly refuse deletion if an active session is currently using that copy. grok worktree rm also erects strict safety boundaries, flatly refusing to remove any Git repository, .git directory, or any folder not created by grok itself, preventing mistyped commands from causing disaster.
During lengthy conversations with AI or extensive tool usage, the frustration of exhausted memory space now has an automatic resolution mechanism.
If a tool turn runs overly long and includes custom post-tool hook output, the program proactively summarizes and compacts the content instead of colliding with prompt length limits. When a context overflow error occurs, the program automatically triggers compaction rather than letting the current turn crash. To prevent runaway compaction from consuming resources, automatic compaction pauses after filling the context three consecutive times, preventing infinite loops.
When feeding prompt batches via automation scripts or connecting external services, bizarre errors caused by invisible characters are now eliminated. In headless mode, if a prompt file starts with a BOM byte-order mark common to Microsoft files, the program proactively strips it away, preventing extraneous markers from reaching the model or corrupting JSON prompt files. When external tools return results over MCP, invalid characters in the payload will no longer trigger tool call timeouts or freezes. Where invisible encoding symbols previously caused subtle bugs that abruptly halted scheduled tasks or left processes hanging, character parsing issues of this kind are now resolved.
Added support for OSC 7501 and the Rex terminal
Expanded tool-call groups stay expanded as new calls join them, and `Ctrl+E` expands or collapses thinking faster in long transcripts
Wrapped URLs in prompts and quotes link to the complete address
Inline tags in agent replies no longer break line layout
The in-session resume picker matches the ungrouped layout on the welcome screen when that layout is configuredProject hooks in a repository subfolder load when you launch inside that folder
HTTP hooks to localhost work, and bad hook URLs are reported at load time
Command status-line scripts see the current turn prompt ID and start time while the turn is active`/model` switches that fail no longer change the footer or the saved default model
`grok -p --effort` warns when the model ignores the flag, and refuses invalid effort levels
Restarting after a model switch no longer reverts to the previous modelExport and copy refuse to overwrite an existing file, and show a clear message
Pasted images over 50 MB are rejected with a clear message instead of being dropped silently
`grep` and `list_dir` report a clear error when a search times out or fails, or the folder cannot be read, instead of returning no resultsShell commands no longer keep running after you quit grok, or after grok receives a signal
Cancelled terminal commands show the command that was stopped
`/workflow stop` picks the correct live runApproval cards for background subagents no longer disappear when the parent turn ends
Approval notifications name the session that needs approval
Minimal mode shows the permission prompt instead of a hidden plan or cancel panel when both are present
Approval cards close the block viewer so the card is visible`grok worktree rm` and `gc` no longer delete a worktree that an open session is using
`grok worktree rm` refuses to delete git repositories, `.git` folders, or any directory grok did not createLong tool turns that include PostToolUse hook output compact instead of hitting the prompt length limit
Context overflow errors trigger automatic compaction instead of failing the turn
Auto-compaction stops after three context refills in a row instead of loopingHeadless prompts that start with a byte-order mark no longer send that mark to the model, and no longer break JSON prompt files
MCP tool replies that contain invalid characters no longer cause the call to time out