Claude Code 2.1.287：Claude Mods 深度模組架構上線，新增貼身提醒副手

繁體中文

新功能

Claude Mods 深度模組架構上線，新增貼身提醒副手

當你專注趕工而擔心漏看關鍵細節時，全新的模組擴充功能能派上用場。Claude Mods 讓外掛能修改更深層的系統運作方式。官方隨附名為 You should know 的內建模組，執行時會有獨立的副手代理人在旁協助把關，主動標記雙方可能遺漏的細節。在開啟遙測的官方工作階段中，於 Claude 的互動介面輸入 /plugin enable cc-plugin-you-should-know@builtin，就能啟用。

新功能

代理人清單支援名稱篩選，折疊區塊也能直接開啟

當你背景開了數十個工作階段而找不到特定任務時，代理人清單現在能快速搜尋。在 agents 檢視畫面中輸入 n:<關鍵字>，就能同時比對工作階段名稱與任務內容。即使符合條件的項目藏在折疊區塊內，篩選器也會完整列出，按下 Enter 鍵還能直接開啟第一筆搜尋結果。

新功能

遙測事件新增提示詞欄位，相容巢狀鍵監控後台

管理雲端監控與追蹤紀錄時，特定後台可能無法順利解析帶有句點的欄位名稱。在 OpenTelemetry（一套收集日誌與效能數據的開放標準）的 user_prompt 事件中，新增了名為 prompt_text 的欄位，內容是 prompt 的完整副本。當你的日誌後端會把點號鍵名拆成一層層巢狀結構時，讀取這個新欄位就不會出錯。只要你在遮蔽或過濾敏感提示詞，記得將這個新欄位一併處理。

新功能

擴充工具支援網址互動引導，登入授權更順暢

當外接工具需要引導你前往特定網頁完成登入時，現在能直接跳出網址提示。此版本支援了 2025-11-25 協定版本的 MCP（Model Context Protocol，讓 AI 連接外部工具的協定）網址引導功能。

更新後若有特定伺服器無法連線，只要在該 MCP 伺服器的設定區塊加入 {"bareElicitationCapability": true} 即可恢復相容。

新功能

Windows 命令工具加速，管線重導向與工具停用提示更清楚

在 Windows 環境下執行指令或串接管線時，現在執行更順暢且錯誤提示更清晰。當拒絕授權 Bash 工具而連帶關閉 PowerShell 時，啟動時會主動發出警告，避免出現完全沒有可用終端機工具的窘境。互動模式在遇到管線輸入或重導向時不再意外卡死或崩潰，會清楚說明原因後結束；若使用管線輸入，可改用 -p 模式。系統還移除了指令執行前多餘的子行程（由主程式分支出來執行的獨立作業），提升命令執行速度。

新功能

自託管環境內建 GitHub 介面指令，未裝命令列工具也能通

當你在團隊自建的執行主機上跑自動化流程，需要向 GitHub 存取資料時，省去了手動補齊工具的麻煩。於 macOS 與 Linux 環境中，使用官方託管 Git 的工作階段已內建 gh api 功能。即使主機未事先安裝相關指令套件，系統也能直接透過 REST 介面存取各項遠端資源。

問題修復

遠端操控連線更穩定，修復休眠中斷與重試停滯

當你透過遠端操控或在代理人環境執行任務時，連線現在更加穩定可靠。在組織允許但無專屬使用者帳號的遠端工作階段中，快速模式現在能正常保持開啟。遠端操控連線遇上無回應狀況時，會在三十秒後自動放棄並重新嘗試，不再枯等數分鐘。透過 HTTP 代理伺服器執行 claude remote-control 現在可完成註冊；先前註冊失敗時，卻會顯示誤導性的組織權限錯誤。在 Mac 閒置休眠時，執行中的工作也不會中斷。

問題修復

事件掛鉤計數精確，修復反覆喚醒與雲端同步

當你在專案中設定自動檢查或在雲端同步外掛時，事件掛鉤的執行行為更加乾淨準確。設定非同步喚醒的掛鉤（特定事件發生時被觸發的小程式）若缺少腳本檔，只會回報一次錯誤，不再反覆喚醒並發出通知。外掛同步的啟動掛鉤現在能在新雲端工作階段順利執行，執行紀錄也不再把內部回呼算進去，設了一個就如實顯示一個。

問題修復

SDK 工具心跳與串流修復，即時訊息不再中斷背景搜尋

當你透過 SDK（供程式串接 Claude 的軟體套件）或指令模式自動化執行時，資料串流與背景運作更穩定。模型回覆串流停滯時，工具的心跳訊號依然會按時送達。在回覆途中切換模型後，claude -p 與 SDK 工作階段不會在後續每則訊息重複觸發備援機制。使用 --output-format stream-json 或 SDK，並以 /<skill> 執行 context: fork 技能時，技能各回合現在也會串流輸出。插隊發送的緊急訊息不會中斷背景正在進行的網頁查詢。

問題修復

企業雲端模型檢查修復，護欄阻擋與標頭驗證更精確

企業透過雲端通道租用模型時，啟動檢查與安全防護回應更清晰穩定。啟動時的模型清單檢查不再遺漏受控設定，選單能完整呈現所有可用型號。思考回覆過程中若觸發安全護欄阻擋，現在會清楚顯示護欄的說明文字而非泛用系統錯誤。關閉實驗功能時，工作階段標題與提示掛鉤請求會徹底移除雲端網關拒收的格式。在 CLAUDE_CODE_SKIP_*_AUTH 檢查中，若 ANTHROPIC_CUSTOM_HEADERS 重複指定 Authorization，模型可用性檢查送出的授權標頭現在會與實際請求一致。Claude apps 閘道在 Amazon Bedrock 拒收模型 ID 時，開發者會在畫面上看到哪個模型不可用，閘道紀錄也會寫出實際送出的 ID。

問題修復

模型切換保留思考深度，雲端預設開啟百萬上下文

在不同模型間切換或需要顧問協助時，各項設定與紀錄能妥善保留。以 claude.ai 登入並在 /model 選擇 Fable 時，儲存的預設模型會跟隨最新版；在 Opus 5.5 與 Sonnet 5.5 間切換，不再改寫工具紀錄而弄丟先前的深度思考。Opus 4.7 以上與 Fable 在 Bedrock、Vertex、Foundry 和 Claude apps gateway 預設使用 1M 上下文空間且不需 [1m] 後綴，設定 CLAUDE_CODE_DISABLE_1M_CONTEXT=1 可維持 200K。/advisor 現支援 Sonnet 5.5 擔任 Opus 4.7 與 4.8 的顧問，並會事先指出 API 不接受的配對。因安全標記切換模型時，也會維持你原先指定的思考力度。

問題修復

終端機指令安全防護升級，敏感檔案寫入嚴格攔截

當你放手讓 AI 批量清理目錄或執行各類系統腳本時，潛在的危險操作現在受到更嚴謹的把關。對 / 或家目錄執行危險 rm，且同一指令把輸出重導向到 ~ 或萬用字元路徑時，不再漏掉確認提示；權限選單也改以白話說明取代內部代碼。Linux 沙盒中的 Bash 指令不再繼承 Claude Code 執行檔已開啟的控制代碼。透過已提交至儲存庫的符號連結（指向另一個位置的捷徑檔）寫入敏感位置或工作目錄外時，會指出最終落點並等待人工同意。即使設定了整個 Bash 工具的允許規則或允許執行的 hooks，對 Claude Code 檔案工具直接拒絕寫入的檔案（Anthropic 個人資料儲存處、主機憑證檔），shell 寫入也只會跳出確認、不會直接執行。

問題修復

對話壓縮不再重複載入設定，動態加入專案即時生效

當對話內容過長觸發自動精簡或恢復先前進度時，各項專案設定能準確維持。恢復工作階段或執行對話壓縮（自動精簡長對話以節省空間的機制）後，資料夾的 CLAUDE.md 不會被重複載入兩次，壓縮後提交署名提醒也不再出現在工具結果裡。雲端階段在對話壓縮途中重啟不會弄丟前半段對話，雲端與 SDK 工作階段中途新增的儲存庫，現在會載入其技能與外掛，CLAUDE.md 也不會在切換目錄後才延遲載入。

問題修復

背景代理人管理更順暢，支援指令排隊與授權提示

當你派發多個背景代理人平行處理任務時，清單檢視與指令互動更加順手。背景任務結束並清理工作樹（Git 讓同一專案平行開多份副本的功能）後，工作階段依然能從管理清單中順利重新開啟。背景中正在等待審核的提示會清楚列在清單中。代理人回覆改為循序排隊送達，回合進行中輸入的斜線指令（除了 /stop 會立刻中斷）都會在當前回合完成後依序執行。

效能優化

終端機滾動不再崩潰，按鍵貼上與授權提示體驗改善

在長對話中捲動畫面或審核授權時，終端機的視覺回饋與操作細節更加穩定。系統負載高時長按捲動鍵不再導致全螢幕介面意外崩潰；使用 --include-partial-messages 時，被截短的回覆不再過晚或漏送 message_stop，應用程式不會再把它誤顯示為仍在進行。開啟減少動態效果時動畫會確實靜止；在 /rewind 確認畫面輸入備註時，「多久以前」的時間不再跳動。淺色主題增強了輸入框外框與歷史箭頭的對比度。Windows、Linux 的右鍵貼上及 Linux 的中鍵貼上改在放開按鈕時生效，放開前移開指標可取消；多個待審核提示則由舊到新依序排列，不再突兀覆蓋目前畫面，但有倒數計時的提示仍會顯示在最上層。

新功能

MCP 工具權限把關更嚴密，協定切換與提示排版最佳化

當你串接外部擴充工具或在背景自動化執行時，工具權限與提示排版更加一致。組織針對 MCP（Model Context Protocol，讓 AI 連接外部工具的協定）工具設定的權限上限，先前會對名稱為 __proto__ 的工具失效，現在不再被忽略。MCP 伺服器更換所支援的協定版本時，連接器的工具呼叫不再偶爾執行兩次，也不會一直失敗到重啟為止；無頭模式（沒有視覺介面的執行模式）的工作階段不會因單次拒絕而誤判需要重新認證。各類工具呼叫與保留訊息的提示畫面統一加上虛線外框，視覺排版更整齊。

新功能

MCP 大量資料處理降載，支援工具延遲載入與獨立重試

當外接工具回傳龐大資料或連線多組伺服器時，資源消耗與載入過程更穩定。MCP 結果存成單行 JSON（以文字儲存資料的格式）時，Claude 不再被錯誤指示使用 Read 的 offset、limit 分頁讀取，因為這種方式拆不開單一長行。處理超額資料時省下記憶體與工作階段儲存空間，也不再為了計算 token（AI 處理與計費的文字單位）而額外上傳。無介面模式下，多伺服器啟動時若遠端伺服器首次連線暫時失敗，會個別重試，無須等待最慢的連線。伺服器設定為 alwaysLoad: false 時，旗下工具會延後至工具搜尋時載入。

問題修復

螢幕閱讀器搜尋與選單修復，按鍵指引更精確（上）

依賴報讀軟體操作終端機時，搜尋框游標位置與選單按鍵指引現在更加準確。搜尋框與登入驗證碼欄位中的游標不再偏離輸入字元。回溯對話時未輸入額外備註也能直接按 Enter 確認，介面也不再顯示無效的 Tab 或方向鍵提示；空選單或搜尋框取得按鍵時，也不再誤顯示「Select with numbers」。當前一個畫面高度超出視窗時，後續的審核提示、設定變更清單，以及遭拒計畫那一行的頂部內容都能完整讀出。

問題修復

螢幕閱讀器檔案差異完整報讀，降低動畫干擾（下）

透過報讀軟體審查程式碼變更與背景進度時，報讀干擾明顯減少且內容更完整。檔案編輯審核提示與其他差異畫面，現在能完整報讀新增、刪除或修改的文字行。傳送進度與表單檢查不再隨著動畫每格重複觸發語音，代理人清單的時間更新頻率也調降以避免語音狂跳。輸出文字時移除了行首多餘停頓，讓報讀節奏更加流暢；設定 CLAUDE_AX_PREPARK_MS=50 可恢復原本的停頓。

問題修復

深度審查上傳引導修正，不再給出錯誤設定建議

當你使用 /ultrareview 指令上傳程式碼進行深度審查時，錯誤引導更加精確。若上傳因無法解析特定編碼的 .gitattributes 檔而中斷，系統不再給出修改其他路徑的無關建議。上傳被拒時，也不會再錯誤引導你將專案設定變數複製進個人全域設定。

問題修復

登入憑證失效提示明確化，不再顯示模糊代碼

當你的 claude.ai 登入憑證被撤銷時，系統現在會直接點出登入憑證已失效。以前登入狀態被撤銷時，終端機只會顯示模糊的 API Error: 401，讓人摸不著頭緒。現在系統會清楚標示「OAuth 權杖已被撤銷（OAuth token revoked）」。在 -p 指令模式下，錯誤訊息也會以「身分驗證失敗（Failed to authenticate）」開頭。以前你會困惑是不是網路或伺服器故障，現在一眼就能看出需要重新登入。

新功能

內建指令操作更順手，設定選單支援左右翻頁與鍵盤切換

在終端機中調整設定或回報問題時，各項內建指令的互動與錯誤回報更加清晰。/config 設定清單支援左右方向鍵雙向切換循環數值，狹窄視窗會自動折行，並支援快捷鍵快速翻頁；/memory 也能直接用左右鍵切換自動記憶等開關。訊息中途打入 /skill 名稱時能被正確識別，包括 disable-model-invocation 類型；問題回報指令不再將近期錯誤直接填入公開範本，而是在確認畫面完整列出。Claude in Chrome 無法連上 Chrome 時，不再誤報 JSON 解析錯誤；/desktop 開啟桌面版逾時或輸出過多時，錯誤訊息會指出原因。

新功能

外掛市集支援純 HTTP 來源，相依套件安裝與快取更穩固

當你安裝或更新自訂外掛時，套件管理流程更加平順可靠。透過純 HTTP 協定託管的儲存庫，現在支援使用 claude plugin marketplace add --sparse 與 git-subdir 安裝外掛。啟動時下載外掛不再因同時重新載入而損壞工作階段的快取（把查過的結果暫存起來直接用）檔案。市集異常時會以白話說明原因與應對步驟，清單會標註未安裝的相依套件，並在更新外掛時重新嘗試未完成的安裝。

新功能

遠端檔案傳輸支援磁碟串流，失敗自動重試且等待時間放寬

當 Claude 從雲端或遠端操控工作階段傳送檔案給你時，檔案傳輸變得更加穩定可靠。遠端上傳的等待時間放寬至三十五秒；上傳逾時、網路錯誤，或伺服器回傳 502、503、504 時，會重試一次。發生暫時性問題時會主動提醒過幾分鐘後可再次索取。傳送大檔案時改直接從硬碟串流讀取，不再塞滿記憶體，超過大小限制時也會明確列出伺服器的具體上限。

問題修復

超大圖片自動縮放傳送，多附件傳輸上限修復

當你在對話中夾帶大量附件或高解析度圖片時，檔案能順暢傳送不再漏件。遠端工作階段傳送單邊超過八千像素的 PNG、JPEG、WebP 圖片原本會失敗，現在會改送縮小後的副本。從應用程式一次發送十七到二十個附件時能完整傳遞所有檔案，不再只送達前十六個。伺服器若拒收超額檔案，也會清楚說明具體原因。

新功能

[VSCode] 支援指令轉入背景執行，代理人地圖即時呈現輸出

在 VSCode 編輯器中使用 Claude 執行長時間任務時，你可以隨時將它移到背景繼續其他工作。執行中的命令或子代理人新增了轉入背景執行選項，讓你無須在原地等待。代理人地圖卡片會即時顯示背景指令與監控工具的最新輸出，主對話回合結束時背景執行中的任務也不會被誤標為失敗。

問題修復

[VSCode] 側邊欄雲端階段同步改善，防止多重視窗重複開啟

在 VSCode 中切換本機與雲端工作階段時，視窗管理更加井然有序。重新開啟已同步的工作階段會直接定位到側邊欄，不再開出多餘分頁。側邊欄能列出視窗開啟後新增的雲端任務；載入失敗時會顯示「Remote server is not connected」，不再誤稱沒有工作階段。重新載入後也不會重複啟動處理程序，而是標示已在其他位置開啟。瀏覽器擴充套件的預設連線開關現在也會套用至編輯器工作階段；連上後，執行瀏覽器動作前仍會詢問。

問題修復

[VSCode] 檔案超連結點擊修復，設定儲存與工具輸出載入更穩定

在 VSCode 擴充功能中查看檔案或執行各項指令時，操作回饋更精準。工具呼叫列的檔案連結、工作階段清單連結與兩處提示，不再誤顯為純文字；計畫預覽分頁中的檔案連結，點擊後現在會開啟檔案。在遠端或子系統環境下檢視工具輸入輸出，不再因編輯器分頁稍慢才出現就報錯逾時。選單中點選 /usage 與 /context 也能正常執行。儲存設定時不會因回覆資料較大而誤判超時，外掛管理清單操作失敗時會列出原因。

問題修復

[Cloud sessions] 雲端工作階段 GitHub 抓取與推送失敗已修復

當你剛連上雲端環境準備與遠端儲存庫同步進度時，流程不再會無故中斷。過去在雲端工作階段中，GitHub 偶爾會短暫拒絕剛核發的存取 token（用來確認身分的授權令牌），導致執行抓取或推送程式碼時失敗。

修復後，在雲端工作階段執行抓取與推送不會再因 GitHub 短暫拒絕新令牌而中斷。

新功能

[Claude Tag] Slack 討論串避免多餘通知，頻道支出上限支援私密群組

在 Slack 團隊頻道中透過標籤指派 Claude 協同工作時，通知干擾明顯降低。當背景事件喚醒機器人且無人等待回覆時，不再突兀地在討論串中張貼警告。任務清單在背景更新時不會以新訊息形式重複發送，避免打擾關注討論串的成員。對頻道很多的組織，管理員後台的支出上限清單現在能列出近期建立及私密頻道。

新功能

[Code Review] 程式碼審查意見完整呈現，重推推送自動重啟審查

當你發出合併請求（Pull request）等待 AI 批閱變更，或是依據回饋再次提交更新時，檢視流程的反饋更為清晰可靠。審查報告中的各項建議與標記原因不再中途斷句，每條評語皆以完整句子呈現。若前一提交的審查失敗兩次，推送新程式碼後仍會審查最新提交，不再略過。若遇到討論串鎖定，卡片會明確說明因鎖定而未執行、未發布任何內容，也不會收取任何費用。

English

FEATURE

Claude Mods deep module architecture launches, with a built-in close-watch reminder assistant

When you're heads-down rushing work and worried about missing key details, the new module extension capability comes in handy. Claude Mods lets plugins modify deeper layers of how the system works. It ships with a built-in module called You should know, which runs a dedicated assistant agent alongside you to help keep watch, proactively flagging details either side might have missed. In official sessions with telemetry enabled, type /plugin enable cc-plugin-you-should-know@builtin in Claude's interactive interface to enable it.

FEATURE

Agent list supports name filtering, and collapsed sections can be opened directly

When you have dozens of sessions running in the background and can't find a specific task, the agent list now supports quick search. In the agents view, type n:<keyword> to match both session names and task content at once. Even if matching items are hidden inside collapsed sections, the filter lists them all, and pressing Enter opens the first search result directly.

FEATURE

Telemetry events gain a prompt field, compatible with nested-key monitoring backends

When managing cloud monitoring and trace records, certain backends may fail to parse field names containing dots. In the OpenTelemetry (an open standard for collecting logs and performance data) user_prompt event, a new field named prompt_text has been added, containing a full copy of prompt. When your log backend splits dotted key names into layers of nested structure, reading this new field won't error out. If you mask or filter sensitive prompts, remember to handle this new field as well.

FEATURE

Extension tools support URL interactive elicitation, making login authorization smoother

When an external tool needs to guide you to a specific web page to complete login, it can now pop up a URL prompt directly. This version supports the URL elicitation feature from the 2025-11-25 protocol version of MCP (Model Context Protocol, the protocol that lets AI connect to external tools).

After updating, if a particular server can't connect, just add {"bareElicitationCapability": true} to that MCP server's configuration block to restore compatibility.

FEATURE

Windows command tool speedup, with clearer pipeline redirection and tool-disable notices

Running commands or chaining pipelines on Windows is now smoother, with clearer error messages. When denying authorization for the Bash tool also shuts down PowerShell, a warning is now issued at startup, avoiding the predicament of having no usable terminal tool at all. Interactive mode no longer unexpectedly hangs or crashes when it encounters piped input or redirection—it clearly explains why and then exits; if you use piped input, switch to -p mode instead. The system also removed the redundant subprocess (an independent process branched off from the main program) before command execution, improving command speed.

FEATURE

Self-hosted environments get a built-in GitHub interface command, working even without the CLI installed

When you run automation on your team's self-hosted runners and need to access data from GitHub, you're spared the hassle of manually installing tools. On macOS and Linux, sessions using officially hosted Git now have gh api built in. Even if the host doesn't have the related command package preinstalled, the system can access remote resources directly through the REST interface.

BUGFIX

Remote control connections are more stable, fixing sleep interruptions and retry stalls

When you run tasks via remote control or in agent environments, connections are now more stable and reliable. In remote sessions where the organization allows it but there's no dedicated user account, fast mode now stays on properly. When a remote control connection hits an unresponsive state, it now automatically gives up and retries after thirty seconds, instead of waiting idly for minutes. Running claude remote-control through an HTTP proxy server can now complete registration; previously, when registration failed, it showed a misleading organization permission error. Running tasks also won't be interrupted when a Mac goes idle and sleeps.

BUGFIX

Event hook counts are precise, fixing repeated wake-ups and cloud sync

When you set up automated checks in a project or sync plugins in the cloud, event hook execution is cleaner and more accurate. A hook (a small program triggered when a specific event occurs) configured for asynchronous wake-up that lacks its script file now reports the error only once, instead of repeatedly waking up and sending notifications. The plugin sync startup hook now runs successfully in new cloud sessions, and execution records no longer count internal callbacks—set one and it faithfully shows one.

BUGFIX

SDK tool heartbeat and streaming fixes—urgent messages no longer interrupt background searches

When you automate execution through the SDK (the software package for programmatically connecting to Claude) or command mode, data streaming and background operations are more stable. When the model's reply stream stalls, tool heartbeat signals still arrive on schedule. After switching models mid-reply, claude -p and SDK sessions no longer repeatedly trigger the fallback mechanism on every subsequent message. When using --output-format stream-json or the SDK and running a context: fork skill via /<skill>, each turn of the skill now streams output too. Urgent messages sent out of queue won't interrupt a web query running in the background.

BUGFIX

Enterprise cloud model check fixes, with more precise guardrail blocks and header validation

When enterprises rent models through cloud channels, startup checks and safety guardrail responses are clearer and more stable. The model list check at startup no longer misses managed settings, so the menu fully shows all available models. If a safety guardrail block is triggered during a thinking reply, the guardrail's explanatory text now displays clearly instead of a generic system error. When experimental features are turned off, session title and prompt hook requests thoroughly remove the format the cloud gateway rejects. In CLAUDE_CODE_SKIP_*_AUTH checks, if ANTHROPIC_CUSTOM_HEADERS specifies Authorization more than once, the authorization header sent by the model availability check now matches the actual request. When Amazon Bedrock rejects a model ID, the Claude apps gateway shows developers on screen which model is unavailable, and the gateway log records the ID actually sent.

BUGFIX

Model switching preserves thinking depth; 1M context on by default in the cloud

When switching between models or calling in an advisor, your settings and history are properly preserved. When you sign in with claude.ai and pick Fable in /model, the saved default model follows the latest version; switching between Opus 5.5 and Sonnet 5.5 no longer rewrites tool history and loses your earlier deep thinking. Opus 4.7 and above, plus Fable, default to 1M context on Bedrock, Vertex, Foundry, and the Claude apps gateway with no [1m] suffix needed, and setting CLAUDE_CODE_DISABLE_1M_CONTEXT=1 keeps you at 200K. /advisor now supports Sonnet 5.5 advising Opus 4.7 and 4.8, and flags pairings the API won't accept ahead of time. When a safety flag triggers a model switch, the thinking effort you originally specified is kept too.

BUGFIX

Terminal command safeguards upgraded, sensitive file writes strictly intercepted

When you let the AI loose on batch-cleaning directories or running all kinds of system scripts, potentially dangerous operations are now gated more rigorously. Running a dangerous rm against / or your home directory while the same command redirects output to ~ or a wildcard path no longer slips past the confirmation prompt; the permission menu also now explains things in plain language instead of internal codes. Bash commands inside the Linux sandbox no longer inherit handles already opened by the Claude Code executable. Writing to sensitive locations or outside the working directory through a symlink (a shortcut file pointing to another location) committed to the repository now shows the final destination and waits for human approval. Even with an allow rule covering the entire Bash tool or hooks permitted to run, files that Claude Code's file tools flat-out refuse to write (Anthropic's personal data store, host credential files) will only trigger a confirmation for shell writes, never silent execution.

BUGFIX

Conversation compaction no longer double-loads settings; mid-session project additions take effect immediately

When an over-long conversation triggers automatic compaction or you resume earlier progress, project settings stay accurate. After resuming a session or running conversation compaction (the mechanism that automatically condenses long conversations to save space), a folder's CLAUDE.md is no longer loaded twice, and the post-compaction commit-attribution reminder no longer shows up in tool results. A cloud session restarting mid-compaction no longer loses the first half of the conversation, and repositories added mid-session in cloud and SDK sessions now have their skills and plugins loaded, with CLAUDE.md no longer loading late only after you switch directories.

BUGFIX

Smoother background agent management, with command queuing and permission prompts

When you dispatch multiple background agents to handle tasks in parallel, list viewing and command interaction feel much smoother. After a background task finishes and its worktree (Git's feature for running multiple parallel copies of the same project) is cleaned up, the session can still be reopened from the management list without trouble. Prompts awaiting approval in the background are now clearly listed. Agent replies are now delivered in a sequential queue, and slash commands typed while a turn is in progress (except /stop, which interrupts immediately) run in order once the current turn completes.

PERF

Terminal scrolling no longer crashes; key pasting and permission prompts improved

When scrolling through long conversations or reviewing permissions, the terminal's visual feedback and interaction details are more stable. Holding a scroll key under heavy system load no longer crashes the full-screen interface unexpectedly; with --include-partial-messages, truncated replies no longer send message_stop too late or drop it, so apps no longer misread them as still in progress. Animations truly stop when reduced motion is enabled, and the "how long ago" time no longer jumps around while you type a note on the /rewind confirmation screen. The light theme boosts contrast for the input box border and history arrows. Right-click paste on Windows and Linux, and middle-click paste on Linux, now take effect on button release, so moving the pointer away before releasing cancels; multiple pending prompts are ordered oldest to newest instead of abruptly covering the current screen, though prompts with countdown timers still appear on top.

FEATURE

Tighter MCP tool permission checks, plus protocol switching and prompt layout polish

When you connect external extension tools or run background automation, tool permissions and prompt layout are more consistent. Organization-set permission ceilings for MCP (Model Context Protocol, the protocol that lets AI connect to external tools) tools previously failed for tools named __proto__; they are no longer ignored. When an MCP server changes the protocol version it supports, connector tool calls no longer occasionally run twice, nor keep failing until restart; headless mode (running without a visual interface) sessions no longer misread a single rejection as needing re-authentication. Prompt screens for tool calls and retained messages now share a unified dashed border for tidier visual layout.

FEATURE

Reduced load for large MCP payloads, with deferred tool loading and independent retries

When external tools return huge payloads or you connect multiple servers, resource consumption and the loading process are steadier. When MCP results are stored as single-line JSON (a text-based data format), Claude is no longer wrongly told to page through them with Read's offset and limit, since that approach can't split one long line. Handling oversized data now saves memory and session storage, and no extra uploads happen just to count tokens (the text units AI processes and bills by). In headless mode, if a remote server's first connection temporarily fails during multi-server startup, each one retries individually without waiting on the slowest connection. When a server is configured with alwaysLoad: false, its tools defer loading until tool search.

BUGFIX

Screen reader search and menu fixes, more precise key guidance (part 1)

When operating the terminal with screen-reading software, the search box cursor position and menu key guidance are now more accurate. The cursor in search boxes and login verification-code fields no longer drifts away from the character being typed. When rewinding a conversation you can press Enter to confirm without typing an extra note, and the interface no longer shows invalid Tab or arrow-key hints; empty menus or search boxes receiving keys no longer wrongly display "Select with numbers". When the previous screen is taller than the window, the top content of subsequent review prompts, settings-change lists, and the rejected-plan line all read out in full.

BUGFIX

Screen readers now read file diffs in full, with less animation interference (part 2)

When reviewing code changes and background progress through screen-reading software, spoken interference drops noticeably and content is more complete. File-edit review prompts and other diff screens now fully read out added, deleted, or modified lines. Send progress and form validation no longer retrigger speech with every animation frame, and the agent list's time updates are throttled to keep speech from jumping wildly. Extra pauses at line starts when printing text have been removed for smoother reading rhythm; setting CLAUDE_AX_PREPARK_MS=50 restores the original pause.

BUGFIX

Deep-review upload guidance fixed; no more wrong settings advice

When you use the /ultrareview command to upload code for deep review, error guidance is more precise. If an upload breaks because a .gitattributes file in a particular encoding can't be parsed, the system no longer suggests modifying unrelated paths. When an upload is rejected, it also no longer wrongly directs you to copy project setting variables into your personal global settings.

BUGFIX

Clearer sign-in credential failure messages, no more vague codes

When your claude.ai sign-in credentials are revoked, the system now states directly that your sign-in credentials are no longer valid. Previously, when your sign-in was revoked, the terminal only showed a vague API Error: 401, leaving you puzzled. Now the system clearly marks it as "OAuth token revoked." In -p command mode, error messages also begin with "Failed to authenticate." Where you once wondered whether the network or a server had failed, you can now tell at a glance that you need to sign in again.

FEATURE

Built-in commands smoother to use, settings menu supports left-right paging and keyboard switching

When adjusting settings or reporting issues in the terminal, the built-in commands interact and report errors more clearly. The /config settings list supports cycling values in both directions with the left and right arrow keys, wraps automatically in narrow windows, and supports shortcut keys for quick paging; /memory also lets you toggle switches like auto-memory directly with the arrow keys. Typing a /skill name mid-message is now recognized correctly, including the disable-model-invocation type; the issue-report command no longer fills recent errors directly into the public template, but lists them fully on a confirmation screen. When Claude in Chrome cannot connect to Chrome, it no longer falsely reports a JSON parsing error; when /desktop times out or produces too much output while launching the desktop app, the error message states the cause.

FEATURE

Plugin marketplace supports plain HTTP sources, dependency installs and more reliable caching

When you install or update custom plugins, the package management flow is smoother and more reliable. Repositories hosted over plain HTTP now support installing plugins with claude plugin marketplace add --sparse and git-subdir. Plugins downloaded at startup no longer corrupt the session's cache files (which temporarily store looked-up results for direct reuse) due to a simultaneous reload. When the marketplace misbehaves, it explains the cause and next steps in plain language, the list flags uninstalled dependencies, and updating a plugin retries any unfinished installs.

FEATURE

Remote file transfers support disk streaming, auto-retry on failure with relaxed wait time

When Claude sends you files from the cloud or a remotely controlled session, file transfers become more stable and reliable. The wait time for remote uploads has been relaxed to thirty-five seconds; upload timeouts, network errors, or server responses of 502, 503, or 504 trigger one retry. When a temporary problem occurs, it proactively reminds you that you can request the file again in a few minutes. Large files are now streamed directly from disk instead of filling up memory, and when a size limit is exceeded, the server's specific cap is clearly listed.

BUGFIX

Oversized images auto-resized for sending, multi-attachment transfer limit fixed

When you attach many files or high-resolution images in a conversation, everything sends smoothly with nothing dropped. In remote sessions, sending a PNG, JPEG, or WebP image over eight thousand pixels on a side used to fail; now a scaled-down copy is sent instead. Sending seventeen to twenty attachments at once from the app now delivers every file, no longer stopping at the first sixteen. If the server rejects files over the limit, it also clearly explains the specific reason.

FEATURE

[VSCode] Commands can move to background execution, agent map shows output live

When using Claude in the VSCode editor for long-running tasks, you can move them to the background at any time and continue other work. Running commands or subagents now offer an option to move to background execution, so you don't have to wait in place. Agent map cards show the latest output of background commands and monitoring tools in real time, and tasks still running in the background are no longer mislabeled as failed when the main conversation turn ends.

BUGFIX

[VSCode] Sidebar cloud session sync improved, duplicate windows prevented

When switching between local and cloud sessions in VSCode, window management is more orderly. Reopening an already-synced session jumps straight to the sidebar instead of opening an extra tab. The sidebar can list cloud tasks added after the window opened; when loading fails, it shows "Remote server is not connected" instead of wrongly claiming there are no sessions. After a reload, it no longer launches duplicate processes, but marks the session as already open elsewhere. The browser extension's default connection toggle now also applies to editor sessions; once connected, it still asks before performing browser actions.

BUGFIX

[VSCode] File hyperlink clicks fixed, settings saving and tool output loading more stable

When viewing files or running commands in the VSCode extension, feedback is more accurate. File links in the tool-call row, session list links, and two prompts no longer wrongly render as plain text; file links in the plan preview tab now open the file when clicked. Viewing tool input and output in remote or subsystem environments no longer times out just because the editor tab appears slightly late. Choosing /usage and /context from the menu also runs properly. Saving settings no longer falsely reports a timeout when the reply data is large, and when a plugin management list operation fails, the reason is listed.

BUGFIX

[Cloud sessions] GitHub fetch and push failures in cloud sessions fixed

When you have just connected to a cloud environment and are ready to sync progress with a remote repository, the flow no longer breaks for no reason. In cloud sessions, GitHub would occasionally briefly reject a freshly issued access token (an authorization token used to confirm your identity), causing fetch or push operations on your code to fail.

After the fix, running fetch and push in a cloud session is no longer interrupted by GitHub briefly rejecting a new token.

FEATURE

[Claude Tag] Slack threads avoid unnecessary notifications, channel spending limits support private groups

When assigning Claude to collaborate via tags in your team's Slack channels, notification noise drops noticeably. When a background event wakes the bot and no one is waiting for a reply, it no longer abruptly posts a warning in the thread. Task lists updated in the background are not resent as new messages, avoiding disturbing members following the thread. For organizations with many channels, the spending-limit list in the admin console can now list recently created and private channels.

FEATURE

[Code Review] Code review feedback now shown in full, and pushing again automatically restarts the review

When you open a pull request and wait for the AI to review your changes, or when you submit updates based on feedback, the review process now gives clearer and more reliable responses. Suggestions and the reasons behind flags in the review report no longer get cut off mid-sentence; every comment appears as a complete sentence. If the review of a previous commit failed twice, pushing new code will still trigger a review of the latest commit instead of skipping it. If a discussion thread is locked, the card clearly states that nothing ran or was published because of the lock, and that you were not charged anything.

原始 CHANGELOG

Original changelogClaude Mods 深度模組架構上線，新增貼身提醒副手

Added Claude Mods: plugins may now modify deeper behavior
Added You should know, a built-in mod where a side agent watches your back and flags things you or Claude might miss. Turn it on with `/plugin enable cc-plugin-you-should-know@builtin` (for first-party sessions with telemetry on)

Original changelog代理人清單支援名稱篩選，折疊區塊也能直接開啟

Added an `n:<text>` filter to the agents view that matches session names and tasks; a filter now shows matches in collapsed sections and Enter opens the first match

Original changelog遙測事件新增提示詞欄位，相容巢狀鍵監控後台

Added `prompt_text` to the OpenTelemetry `user_prompt` event, a copy of `prompt` for backends that nest dotted keys; drop or mask it wherever you drop or mask `prompt` (anthropics/claude-code#70763)

Original changelog擴充工具支援網址互動引導，登入授權更順暢

Added URL prompts from MCP servers on the 2025-11-25 protocol, for example to sign in. If a server no longer connects after this update, add "bareElicitationCapability": true to its MCP config entry

Original changelogWindows 命令工具加速，管線重導向與工具停用提示更清楚

Windows: Added a startup warning when denying the Bash tool also turns off the PowerShell tool, so Claude has no shell tool
Windows: Fixed interactive `claude` hanging or crashing with "Raw mode is not supported" when its input is piped or redirected; it now says why and exits (use `-p` for piped input)
Windows: Improved Bash tool speed by removing a subshell that ran before every command

Original changelog自託管環境內建 GitHub 介面指令，未裝命令列工具也能通

Self-hosted runner: Added a built-in `gh api` (REST only) for sessions that use Anthropic-managed git on macOS and Linux machines where the GitHub CLI is not installed

Original changelog遠端操控連線更穩定，修復休眠中斷與重試停滯

Fixed fast mode staying off in remote sessions owned by an agent with no user account, even when the organization allows it
Fixed Remote Control not receiving messages for minutes at a time when a reconnect request got no response; it now gives up after 30 seconds and retries
Fixed `claude remote-control` failing to register behind an HTTP proxy with a misleading "Check your organization permissions" error (anthropics/claude-code#97352)
macOS: Fixed Remote Control sessions started with `claude remote-control` stopping mid-turn when the Mac went to idle sleep

Original changelog事件掛鉤計數精確，修復反覆喚醒與雲端同步

Fixed hooks configured with `asyncRewake` waking Claude over and over with "found issues" notifications when the hook's script file is missing; the broken hook is now reported once
Fixed SessionStart hooks from synced plugins not running in new cloud sessions
Fixed the transcript's "N hooks ran" summary and the verbose debug log's matched-hooks count including Claude Code's internal callbacks, so one configured hook no longer shows as two

Original changelogSDK 工具心跳與串流修復，即時訊息不再中斷背景搜尋

Fixed tool heartbeats not reaching SDK hosts while the model's response stream was stalled with no data arriving
Fixed `claude -p` and SDK sessions repeating a model fallback on every later message after the model was switched while a reply was running
Fixed `--output-format stream-json` and the SDK not streaming the turns of a `context: fork` skill run by typing `/<skill>` as the prompt, as they do for the Skill tool's fork
Improved SDK sessions so a message sent with priority "now" no longer cancels a running web fetch or web search; it keeps loading in the background

Original changelog企業雲端模型檢查修復，護欄阻擋與標頭驗證更精確

Fixed Bedrock and Vertex startup model checks ignoring an enforced `availableModels` list, which could collapse `/model` to one Opus row
Fixed Amazon Bedrock Guardrails blocks that arrive mid-response ending the turn with an API error instead of the guardrail's message when the reply began with thinking
Fixed `CLAUDE_CODE_DISABLE_EXPERIMENTAL_BETAS` not removing the structured-output format from session-title and prompt-hook requests, which Bedrock-backed gateways reject
Bedrock, Vertex, Mantle: Fixed model availability checks under `CLAUDE_CODE_SKIP_*_AUTH` sending a different `Authorization` header than real requests when `ANTHROPIC_CUSTOM_HEADERS` repeats it
Improved the Claude apps gateway's error when Amazon Bedrock rejects a model ID: developers now see which model is unavailable, and the gateway log names the ID that was sent

Original changelog模型切換保留思考深度，雲端預設開啟百萬上下文

Fixed picking Fable in `/model` on a claude.ai login saving the current version's id, so your saved default now follows the newest Fable like Opus and Sonnet do
Fixed switching between Opus 5.5 and Sonnet 5.5 (`/model`, `opusplan`) rewriting earlier MCP tool announcements, which could drop earlier extended thinking
Fixed `/advisor` pairing checks: Sonnet 5.5 can now advise Opus 4.7 and 4.8, and advisors the API would refuse are flagged up front instead of being silently dropped
Changed Opus 4.7+ and Fable to use a 1M context window by default on Bedrock, Vertex, Foundry and the Claude apps gateway, with no `[1m]` suffix (`CLAUDE_CODE_DISABLE_1M_CONTEXT=1` keeps 200K)
Changed automatic model switches after a flagged message to keep your current effort level instead of the new model's default

Original changelog終端機指令安全防護升級，敏感檔案寫入嚴格攔截

Fixed a dangerous `rm` (such as one on `/` or the home directory) losing its always-ask safeguard when the same command also redirected output to a `~` or wildcard path
Fixed Bash permission prompts showing internal parser names such as "Contains simple_expansion" instead of a plain explanation
Fixed sandboxed Bash commands on Linux inheriting an open handle on the Claude Code executable
Changed a shell write through a repo-committed symlink onto a sensitive file or out of the working tree to name where it lands and wait for a person, on lines with a `~` target too
Changed whole-tool `Bash` allow rules and allowing hooks to prompt for, not run, shell writes to files Claude Code's file tools refuse outright (the Anthropic profile store, the host credentials file)

Original changelog對話壓縮不再重複載入設定，動態加入專案即時生效

Fixed a folder's CLAUDE.md being attached a second time after resuming a session or after a compaction
Fixed the commit attribution reminder being delivered inside a tool result after a compaction
Fixed cloud sessions sometimes losing the earlier conversation when the session restarted while it was being compacted
Fixed repositories added mid-session in cloud and SDK sessions not loading their skills and plugins, and loading CLAUDE.md late, after Claude changed directory

Original changelog背景代理人管理更順暢，支援指令排隊與授權提示

Fixed background sessions that could not be reopened from `claude agents` after the agent exited and removed the worktree the session was started in
Fixed `claude agents` sometimes not showing the permission prompt a background session is waiting on
Changed replies from `claude agents` to arrive as queued messages; slash commands other than `/stop` sent while a turn is running now run when it ends

Original changelog終端機滾動不再崩潰，按鍵貼上與授權提示體驗改善

Fixed a cause of fullscreen sessions on slow or busy machines exiting with "Claude Code exited after an unrecoverable interface error" while a scroll key was held in a long conversation
Fixed `--include-partial-messages` sending a cut-short reply's `message_stop` late or never, so apps could show the reply as still in progress
Fixed the running-tool dot and three spinners still moving with the "Reduce motion" setting on, and /rewind's confirm screen updating its "ago" time while you type a note
Improved the contrast of the prompt input border in light themes and of the ❯ before your earlier messages
Changed right-click paste on Windows and Linux, and middle-click paste on Linux, to happen when the button is released; moving the pointer away before releasing cancels it
Changed waiting permission prompts to show oldest first, so a new prompt no longer covers the one you're reading (prompts with a countdown still open on top)

Original changelogMCP 工具權限把關更嚴密，協定切換與提示排版最佳化

Fixed organization per-tool permission ceilings being silently dropped for an MCP tool named `__proto__`
Fixed an MCP connector tool call occasionally running twice, or the connector's calls failing until restart, when its server changed which MCP protocol version it supports
Fixed headless sessions reporting an MCP server as needing authentication after one refused call, even though later calls succeed
Improved the prompt for a held message from another session to show the message between dashed lines, matching other permission prompts
Improved MCP and other tool permission prompts to show the tool call between dashed lines, matching file edit prompts

Original changelogMCP 大量資料處理降載，支援工具延遲載入與獨立重試

Fixed Claude being told to page large MCP results saved as JSON with Read's offset and limit, which cannot split one long line
Improved MCP startup in headless mode: a remote server whose first connect fails transiently is now retried without waiting for the slowest server to finish connecting
Improved handling of large MCP tool results: less memory, smaller session files, and no extra upload to count tokens for results far over the limit
Changed MCP server `alwaysLoad: false` to defer all of that server's tools behind tool search

Original changelog螢幕閱讀器搜尋與選單修復，按鍵指引更精確（上）

Fixed screen reader mode leaving the cursor away from the typed text in search boxes (such as /resume and /permissions) and sign-in code fields
Fixed screen reader mode refusing Enter with nothing typed on /rewind's summarize options, whose added context is optional
Fixed screen reader mode showing a "Tab to amend" hint on approval prompts, where Tab does nothing
Fixed screen reader mode listing arrow keys that do nothing in /permissions and /mcp, and saying "Select with numbers" in empty menus or while a search box has the keys
Fixed screen reader mode leaving out the top lines of a second approval prompt, a changed /config row or the rejected-plan line when the previous screen was taller than the terminal window

Original changelog螢幕閱讀器檔案差異完整報讀，降低動畫干擾（下）

Fixed screen reader mode leaving out the changed lines in file edit approval prompts and other diffs
Fixed screen reader mode sending the `claude --teleport` progress screen, and an MCP form field while it is being checked, to the screen reader again on every spinner frame
Fixed times in `claude agents` changing every second in screen reader mode; they now change at most every 10 seconds
Changed screen reader mode to write new or changed lines without first pausing with the cursor at the start of the line; set `CLAUDE_AX_PREPARK_MS=50` to restore the pause

Original changelog深度審查上傳引導修正，不再給出錯誤設定建議

Fixed `/ultrareview` giving advice about `.git/info/attributes` when the upload stops on a committed `.gitattributes` it cannot read, such as one saved as UTF-16
Fixed `/ultrareview` upload refusals advising you to copy a variable named by a repository's settings file into your own user settings

Original changelog登入憑證失效提示明確化，不再顯示模糊代碼

Fixed a revoked claude.ai login showing a generic `API Error: 401` instead of "OAuth token revoked"; in `-p` mode the error now starts with "Failed to authenticate"

Original changelog內建指令操作更順手，設定選單支援左右翻頁與鍵盤切換

Fixed the Claude in Chrome browser picker showing a JSON parse error when Chrome could not be reached
Fixed /feedback and /bug: the pre-filled GitHub issue no longer includes your recent error messages, and the confirmation screen now lists them as part of the report
Fixed `/desktop` quoting partial output when opening Claude Desktop timed out or printed too much output; the error now names the cause
Improved `/config`: settings that cycle show ‹ › and step both ways with ←/→, narrow terminals stack each value under its label, and PgUp/PgDn page the list
Improved `/memory`: the left and right arrow keys now flip its on/off settings, such as Auto-memory
Improved `/skill` names typed mid-message: Claude is now told they are skills, including `disable-model-invocation` ones

Original changelog外掛市集支援純 HTTP 來源，相依套件安裝與快取更穩固

Fixed `claude plugin marketplace add --sparse` and `git-subdir` plugin installs failing with "transport 'http' not allowed" when the repository is served over plain http
Fixed a plugin reload that overlapped the startup `--plugin-url` download corrupting the session's cached copy of the plugin archive
Improved plugin marketplace errors to say in plain words why a marketplace was ignored or refused, and what to do
Improved plugin listings to note when a plugin's dependencies were not installed, and updating a plugin now retries an install that did not finish

Original changelog遠端檔案傳輸支援磁碟串流，失敗自動重試且等待時間放寬

Fixed files Claude sends from cloud and Remote Control sessions failing when the upload finished just after the 30-second timeout; it now waits 35 seconds
Improved delivery of files Claude sends from cloud sessions and Remote Control: an upload that fails on a timeout, a network error or a 502, 503 or 504 is now retried once
Improved what Claude says when a file cannot be sent for a reason that may be temporary: it now mentions that you can ask for the file again in a few minutes
Improved files Claude sends from a remote session: large files now stream from disk instead of being read into memory, and a file over the size limit is refused with the server's limit named

Original changelog超大圖片自動縮放傳送，多附件傳輸上限修復

Fixed PNG, JPEG and WebP images over 8,000 pixels on a side failing to send from a remote session; Claude now sends a scaled-down copy
Fixed messages sent from the Claude apps with 17 to 20 attached files delivering only the first 16
Improved the explanation Claude gives when the server refuses a file it sends from a Remote Control or cloud session, such as an oversized image

Original changelog[VSCode] 支援指令轉入背景執行，代理人地圖即時呈現輸出

[VSCode] Added "Run in background" to a running command or sub-agent, to move it to the background and keep working
[VSCode] Added the output of background shells and Monitors to their cards in the agent map
[VSCode] Fixed a background agent's still-running command showing as failed once the main turn ended

Original changelog[VSCode] 側邊欄雲端階段同步改善，防止多重視窗重複開啟

[VSCode] Fixed reopening a cloud session that the side bar already brought to this machine opening it again in a new tab; the side bar is shown instead
[VSCode] Fixed the side bar's Web tab not listing cloud sessions started after the window loaded; a failed load now says "Remote server is not connected" instead of "No web sessions yet"
[VSCode] Fixed a tab restored after a reload starting a second Claude process on a conversation the side bar already has open; it now shows the "still open somewhere else" notice
[VSCode] Changed the Claude in Chrome "Enabled by default" switch to also connect the editor's own sessions, which still ask before browser actions

Original changelog[VSCode] 檔案超連結點擊修復，設定儲存與工具輸出載入更穩定

[VSCode] Fixed settings dialogs blaming a timeout when Claude Code's reply was too large to confirm a save
[VSCode] Fixed tool-row file links, session-list links and two hints showing in plain text
[VSCode] Fixed a user's own `/usage` or `/context` command opening the extension's dialog instead of running when picked from the command menu
[VSCode] Fixed file links in the plan preview tab doing nothing when clicked; they now open the file like links in chat replies
[VSCode] Fixed opening a tool's input or output in an editor tab failing with "Timeout waiting after 1000ms" on remote hosts such as WSL when the tab is slow to appear
[VSCode] Improved the Manage plugins dialog: a failed marketplace add, remove or refresh now says what went wrong

Original changelog[Cloud sessions] 雲端工作階段 GitHub 抓取與推送失敗已修復

[Cloud sessions] Fixed occasional failures to fetch from or push to GitHub when GitHub briefly refused a newly issued access token

Original changelog[Claude Tag] Slack 討論串避免多餘通知，頻道支出上限支援私密群組

[Claude Tag] Fixed Claude posting a failure warning, such as a spend limit notice, in a Slack thread when a background event like GitHub activity woke it and nobody was waiting on a reply
[Claude Tag] Fixed Claude Tag's spend limits page in admin settings leaving out recently created and private channels in organizations with many channels
[Claude Tag] Improved Claude's task list in long Slack threads: background work no longer reposts it as a new message on its own, so people following the thread aren't notified

Original changelog[Code Review] 程式碼審查意見完整呈現，重推推送自動重啟審查

[Code Review] Fixed finding comments and their "Why this was flagged" text stopping mid-sentence; they now end on a complete sentence
[Code Review] Fixed Code Review skipping a pull request after a new push when its review had failed twice on the previous commit; it now reviews the latest commit
[Code Review] Improved the failed-review card on a pull request whose conversation is locked: it now says the lock blocked the review and that nothing was posted or charged