Claude Code 2.1.236

當工作階段（session）切換進入的資料夾被外部刪除後，各項背景與周邊功能不再整批壞掉。自 2.1.229 版本以來，一旦所在目錄消失，剪貼簿複製、背景日常維護、背景工作階段與本機 MCP（讓 AI 連接外部工具的協定）日誌都會損毀；SDK 與 VS Code 工作階段在技能變更時也會反覆報錯。現在系統已妥善處理目錄遺失的狀況，所有功能均能穩定維持運作。

全螢幕渲染器現在具備自我修復與備援（fallback）機制。過去如果全螢幕渲染器在啟動時遭遇單一次失敗，整個設定便會陷入永久損毀狀態，導致後續每次啟動 Claude Code 時都會直接崩潰退出。現在系統遇到全螢幕渲染啟動失敗時，會主動退回至穩定的傳統渲染器，確保終端機依然能順利開啟並供你正常操作。

輸入 /model 挑選模型時，選單的高度現在會自動符合終端機視窗大小。當清單內容超出視窗高度時，系統只會顯示容得下的項目，其餘模型可透過滾動查看；同時，介面改為僅醒目標示最新釋出的模型名稱，讓視覺焦點明確落在最新版本上。

跨工作階段（session）傳送訊息時，格式解析變得更具容錯能力。以往在使用 SendMessage 工具時，如果訊息內容中含有格式錯誤的閉合標籤，系統解析時會不小心把整段訊息文字誤塞進摘要（summary）欄位中，進而導致整筆發送請求被後端直接拒絕。現在這項解析邏輯已修正完畢。

當外部子行程（subprocess）無法啟動時，系統現在能優雅地捕捉錯誤。在 2.1.234 版本中，如果在關閉 Windows 互通性的 WSL（Windows 子系統）環境中嘗試啟動 powershell.exe，系統會因為未妥善處理非同步拒絕而直接崩潰；現在這類錯誤會被妥善捕捉與處理。

全螢幕模式下的畫面重繪與文字顯示迎來多項細節修正。過去在縮放終端機視窗大小後，新送出的訊息有時會卡住，直到下一次畫面更新才突然冒出來；清除多行提示詞時，輸入框上方也常留下一道空白帶；且將視窗縮小後再放大時，各面板畫面更會出現無法重繪的缺塊。現在這些介面重繪問題已全數修復。

系統修復了受管設定（managed settings，由企業或管理者統一推送的設定檔）在啟動時的嚴重安全風險。過去在特定狀況下，啟動時的設定核准視窗未能正常顯示在畫面上，卻依然偷偷將使用者的第一次按鍵當成同意審查，造成誤按核准；現在對話框必定會確實顯示後才接收輸入。

在 tmux（特別是搭配 iTerm 的 tmux 整合模式）中執行時，分頁標題不再持續抖動。過去系統每隔 960 毫秒就會進行一次動畫式的標題重寫，導致終端機分頁上的文字不斷閃爍跳動，干擾視覺專注；現在系統改為僅在標題文字真正發生改變時才執行寫入操作，保持分頁列的穩定與清爽。

當你存取雲端環境時，遇到清單異常不再摸不著頭緒。以往如果從遠端取回的雲端環境清單為空白，或者回傳的資料結構格式損毀，系統只會跳出語意模糊、難以排查的錯誤代碼；現在錯誤訊息已大幅改善，會清楚指明清單為空或格式不符的具體原因，方便你即時排查網路或遠端設定問題。

使用遠端遙控（Remote Control）時，額度提示不再逾時亂選。針對 Fable 5 的首次使用額度確認對話框，過去若等了 60 秒仍未收到使用者回應，系統會自動套用備援（fallback）模型；現在已移除這項自動選擇，靜候你的決定。

在等待 Claude 運算時，旋轉指示器旁的小提示（spinner tips）現在能穩定顯示。以前如果個人目錄設定檔 ~/.claude.json 中快取（cache，把算過或查過的結果暫存起來直接用）的訪客通行證獎勵資料格式損毀，系統會在背景不斷陷入重複報錯的迴圈，導致轉圈等待時的小提示完全無法跳出來。現在系統會妥善防護並略過格式錯誤的快取資料。

自架執行器（self-hosted runner）的工作階段調度更穩固了。工作階段可能因閒置、除役（retire）或啟動逾時而釋放。過去在此類情境下，後置掛鉤（post-session hook，特定事件發生時被觸發的小程式）可能還在收尾，系統卻偶爾已在另一個執行器上恢復該工作階段；現在必定會等待掛鉤完全結束才允許接續。

在 iTerm2 終端機中看見的 Clawd 吉祥物變得更加工整可愛。過去在某些特定的字型與字級大小設定下，終端機渲染文字圖形時會產生微小字元寬度偏差，導致 Clawd 的眼睛左右不對稱、雙腳高度參差不齊；這次更新微調了字元輸出對齊方式，確保在各種字型大小下都能呈現端正勻稱的吉祥物樣貌。

工作階段（session）的摘要功能不再無限膨脹。過去無論是系統自動產生的摘要，還是手動輸入 /recap 產生的回顧，偶爾會失控生成落落長的冗長篇幅；現在摘要文字被嚴格限制在最多 400 個字元內，並在單字邊界截斷，不從單字中間切開。

跨工作階段（session）傳送訊息時，資料收發的可靠度大幅提升。過去如果短時間內密集送出大量突發訊息，導致數量超出接收端收件匣所能承載的上限，系統會表面回報發送成功、背地裡卻把超出的訊息悄悄丟棄；現在 SendMessage 會在最前線直接拒絕超量的發送請求並回傳錯誤，讓你能即時掌握傳輸狀況。

Background tasks and utilities no longer break when a session's directory is deleted from underneath it. Since 2.1.229, removing the active working directory broke clipboard copying, background housekeeping, background sessions, and local MCP (Model Context Protocol) logs, while triggering errors on every skill change in SDK and VS Code sessions. The runtime now handles missing working directories smoothly without failing.

The fullscreen renderer now recovers gracefully from startup failures. Previously, if the fullscreen renderer failed to initialize a single time, the application became permanently broken and crashed on every subsequent launch. It now falls back to the classic terminal renderer instead of exiting, keeping your session operational.

The /model picker now fits cleanly inside your terminal window. If the list of available models is taller than the terminal, it renders only what fits and lets you scroll through the rest. Furthermore, the highlight style is now reserved exclusively for the newest model release rather than arbitrary items.

Cross-session messaging now parses malformed payload tags reliably. Previously, when calling SendMessage, a broken closing tag could accidentally displace the message body into the summary parameter, causing the request to be rejected. The parser now isolates the text correctly regardless of imperfect tag formatting.

Subprocess startup failures no longer trigger unhandled promise rejections. In version 2.1.234, trying to launch a command like powershell.exe on WSL with Windows interop disabled caused unhandled runtime crashes. The process manager now catches and handles these startup errors cleanly.

Fullscreen mode received several screen-drawing and layout fixes. Resizing the terminal window sometimes caused newly submitted messages to remain invisible until the subsequent UI tick. In addition, clearing multi-line prompts could leave an empty blank band above the input area, and resizing panes away and back left views unpainted. These visual rendering issues have all been resolved.

A dangerous security issue with managed settings prompts has been fixed. Previously, the startup prompt asking for approval of managed configurations could fail to display on screen while still recording your very first keystroke as consent. The prompt now reliably renders before processing user input.

Terminal tab titles inside tmux (especially via iTerm tmux integration) no longer jump jittery across updates. The title string used to be rewritten continuously on a 960ms animation loop, creating visible flicker on the tab bar. It is now written only when the title text itself changes.

Fetching cloud environments now returns descriptive diagnostics if the response is empty or damaged. Previously, when the cloud environment endpoint returned an empty list or malformed data, the CLI presented an opaque, unhelpful error message. It now explains the failure clearly so you can diagnose configuration issues quickly.

When using Remote Control, the initial usage-credits confirmation prompt for Fable 5 no longer automatically picks a fallback model after 60 seconds of inactivity. It now remains open until you explicitly respond.

Spinner tips now appear reliably while waiting for responses. Previously, if the cached guest-pass reward in ~/.claude.json was malformed, the system entered a repeated background error loop and failed to render loading tips completely. The cache reader now handles corrupted data gracefully.

Self-hosted runner session lifecycle handling is now race-free. When a session is released due to idle, retire, or startup timeout, it used to occasionally resume on a different runner before its post-session hook finished executing. The coordinator now strictly awaits hook completion before migrating the session.

The Clawd mascot in iTerm2 now displays with consistent proportions. At certain terminal font sizes, character cell measurement differences caused Clawd's ASCII eyes and feet to render misaligned and lopsided. Spacing has been adjusted so the mascot appears cleanly balanced across different font configurations.

Session recaps will no longer run away with endless text. Both automatic recaps and manual /recap outputs are now capped at a maximum of 400 characters, neatly truncated at a clean word boundary so that sentences do not cut off mid-word.

Cross-session messaging now guarantees honest feedback when handling sudden bursts. Previously, when rapid bursts exceeded what a destination session's inbox could accommodate, SendMessage reported them as delivered even though messages were dropped. It now rejects excessive messages up front with an error so the caller can back off or retry.