Claude Code 2.1.260

在全螢幕模式下寫程式，現在可以輸入 /diff 開啟或關閉側邊的比對面板，即時查看 Claude 編輯時尚未提交的變更。當終端機畫面太滿時，按下 ctrl+l 或 cmd+k 就能清空畫面，想查閱歷史訊息只要往上捲動即可。此外，在全螢幕下按 Ctrl+Z 暫停不會再把外層 shell 畫面錯位畫在暫停介面上，國旗圖示、組合表情符號與重音字母在換行或行末最後兩格時也不會再破版或留下殘影。

透過桌面端 App、Remote Control 或以 -p 參數執行的無頭模式（headless，沒有終端機視窗／視覺介面的執行模式）工作階段，現在能直接使用純文字形式的顧問指令。你可以輸入 /advisor、/advisor <model>、/advisor off 三種文字形式。同時，無頭環境也新增了 /reload-plugins 來重新載入外掛，因此會出現在 Claude Code Desktop 與 SDK 指令清單。如果遠端輸入了不存在的模型名稱，系統會在當下直接報錯拒絕，避免等到下一則訊息才崩潰。非互動模式（-p / SDK）工作階段的閒置時處理器用量改善。

prompt_cache  # status line 欄位
model: fable  # 代理設定
ANTHROPIC_DEFAULT_FABLE_MODEL  # 綁定
/model claude-fable-5-1  # 完整指令

輸入 /cost，或查看 status line（狀態列）的 prompt_cache 欄位時，快取未命中會標出原因，例如 tool definitions、system prompt 變更，或閒置超過 TTL。Claude Fable 5.1 先前沒把工具結果後的脈絡納入提示詞快取，所以每輪 tool-call 曾當未快取輸入重送；現在會納入。中途改 /effort 不再讓快取失效。上列 model: fable 代理不會再忽略綁定上的 [1m] 而默默改用 200K。組織若能用 Fable 5.1，/model 選單會顯示它；先前只能打上列完整指令。Opus 與 Fable 工作階段會在接近 100 萬 token（文字單位）上限前壓縮，極大脈絡時救援壓縮不再 10 分鐘超時。

路徑含括號的 Edit、Write 或 Read 權限規則，以前曾被當成無效丟掉，或被 Bash 沙盒忽略，唯讀資料夾因此可寫。若出現未閉合中括號等無法編譯的正規表達式（regex，用符號比對字串的語法），所有檔案編輯都會報 Invalid regular expression。現在此類拒絕規則會退回保護其字面路徑。若在括號後誤加多餘文字（例如 Bash(ls) x），系統會直接回報設定無效而非默默忽略。Windows 例 Edit(C:\dir\(name)\**) 裡的 \( 會被讀成跳脫括號而非路徑分隔，錯誤訊息會建議明確寫法。此外，Glob 與 Grep 工具現在會先確認權限，才回報路徑是否存在。

閘道支援下列鍵：

oidc.scope_on_refresh  # 範圍鍵
userPluginMarketplacesEnabled  # Desktop 鍵
userPluginUploadsEnabled  # Desktop 鍵
orgPluginSettings  # 清單格式

身分提供者若須再要 openid 才會在更新時給識別令牌，可設定 oidc.scope_on_refresh。更新失敗日誌也會具名失敗步驟。Bedrock 被中斷的請求改用 AWS 免費 CountTokens API 計算輸入 token，需授予 IAM bedrock:CountTokens，而非改送只含一個 token 的請求。兩把 Desktop 鍵寫在 desktop 政策區塊。orgPluginSettings 改以清單格式傳送，供 Claude Desktop 1.15200.0 起讀取；舊版忽略。若 desktop 政策在 MCP（外部工具協定）的 managedMcpServers 或 orgPluginSettings 巢狀物件拼錯欄位，啟動時會拒絕並指出該欄位。

當子代理透過 SendMessage 喚醒另一個代理時，對方完成後的通知現在會正確送回給發起喚醒的子代理，不再誤送到主工作階段。代理團隊在 API 長時間重試等待時（例如處於 CLAUDE_CODE_RETRY_WATCHDOG），對話紀錄不會再被重試通知擠掉而洗白。執行工作流時，如果子代理正在進行長時間的脈絡壓縮，不會再被誤判為停滯而遭粗暴重啟。結構化輸出方面，Workflow 的 agent({schema}) 若遇到永遠無法滿足的 JSON Schema，會事先拒絕；達到重試上限時也會附上最後一次驗證錯誤。子代理啟動的背景指令取消了 1 小時限制，會執行到自己結束或被停止為止，並對齊主工作階段。

移至背景的工作階段不會再於 ListAgents 出現同名的 interactive 幽靈分身，檢視器也不再收到它的 SendMessage。多個工作階段共用同一個專案目錄時偶發的輸出交換錯誤也已修復。刪除帶有未推送提交的工作樹（worktree，Git 讓同一專案平行開多份副本的功能）背景工作階段時，第一次會報出分支名稱與提交數量，再執行一次刪除才會丟掉該工作樹。以 -p --resume 或 --continue 接續若遺失 Git 詮釋資料，失敗一次後會跳過工作樹繼續執行。自託管 runner 若設定了 --kill-session-after-min，遇到純粹等待使用者輸入的暫停狀態時會和平釋放，不會強制終止並報錯。

IDE 選取的行數在觸發技能（skill）或斜線指令時不會再遺失，Claude 能收到 N lines selected 這句脈絡。受控設定的 skillOverrides 若以內建技能別名為鍵（例如用 checkup 覆寫 /doctor）已能正常生效，Skill(name) 拒絕規則也完整涵蓋了 <目錄>:名稱 形式的巢狀技能。內建的 claude-api 技能範例已更新 Go、Java 與 C# 程式碼至最新世代模型識別碼，並說明較便宜的 worker 或子代理模型也應採用新世代模型。雲端審查指令 /ultrareview 與 claude ultrareview 的等候時間上限從 30 分鐘延長至 45 分鐘。

VS Code 擴充套件在介面顯示與工作階段管理上變得更加俐落。底部的模型標籤現在會直接顯示目前選擇的努力程度，切換模型後數值不再停留在舊狀態，且標籤尺寸已恢復為原本的緊湊大小。工作階段清單的狀態篩選選單中，新增了 Open 與 Closed 兩個選項，方便你快速分類與尋找專案。此外，Claude Code 對 dropped response 重試後，畫面不再留下半截訊息、空白工具卡或多餘的 Thought for 行。

在 VS Code 中操作多個分頁與遠端控制時的體驗更加順暢。在歷史工作階段選擇器中點選已在其他分頁開啟的工作階段時，編輯器會直接切換到該分頁，不再重複載入同一份對話。在分頁畫面重新整理期間執行重新命名（Rename）指令，系統現在會確實套用新名稱，不再無聲失效。新開啟工作階段時若自動啟動 Remote Control，歡迎畫面不會再莫名消失。此外，若在分頁剛啟動的當下切換「為所有工作階段啟用 Remote Control」，該設定現在能正確套用到正在初始化的分頁中。

When coding in fullscreen mode, you can now enter /diff to toggle a side-by-side diff panel and inspect uncommitted changes in real time as Claude edits. When the terminal screen gets crowded, pressing ctrl+l or cmd+k clears the transcript view, and you can scroll up to view earlier messages. In addition, pressing Ctrl+Z in fullscreen to pause no longer misrenders the outer shell screen onto the paused interface, and flag icons, joined emoji, or accented letters will no longer break layout or leave stale artifacts when wrapping or falling in the terminal's last two columns.

Sessions running in headless mode via the desktop app, Remote Control, or the -p flag can now directly use text-based advisor commands. You can enter three text forms: /advisor, /advisor <model>, and /advisor off. At the same time, headless environments add /reload-plugins to reload plugins, making it available in the Claude Code Desktop and SDK command lists. If an invalid model name is entered remotely, the system now refuses it with an error immediately instead of crashing on the next message. Idle CPU usage has also improved for non-interactive (-p / SDK) sessions.

prompt_cache  # status line field
model: fable  # agent setting
ANTHROPIC_DEFAULT_FABLE_MODEL  # pin
/model claude-fable-5-1  # full command

When running /cost or checking the prompt_cache field in the status line, cache misses now indicate the likely cause, such as changes to tool definitions, system prompt modifications, or idling past the TTL. Claude Fable 5.1 previously did not include the context attached after tool results in prompt caching, causing it to be re-sent as uncached input on every tool-call turn; it is now included. Changing /effort mid-session no longer invalidates the prompt cache. The model: fable agents listed above will no longer ignore the [1m] tag on pins and silently fall back to 200K. If an organization has access to Fable 5.1, the /model picker now displays it; previously it had to be typed as the full command above. Opus and Fable sessions now compact shortly before reaching the 1M token limit, and recovery compaction on very large contexts no longer times out at 10 minutes.

Permission rules for Edit, Write, or Read whose paths contain parentheses were previously dropped as invalid or ignored by the Bash sandbox, leaving read-only folders writable. If an uncompilable regex pattern appears, such as an unclosed bracket, every file edit fails with Invalid regular expression. Deny rules of this kind now fall back to guarding their literal paths. If extra text is mistakenly appended after the closing parenthesis, such as Bash(ls) x, the system now directly reports an invalid setting instead of silently ignoring it. For Windows examples like Edit(C:\dir\(name)\**) where \( is read as an escaped parenthesis rather than a path separator, the error message now suggests an unambiguous spelling. In addition, the Glob and Grep tools now check permissions before reporting whether a path exists.

The gateway now supports the following keys:

oidc.scope_on_refresh  # scope key
userPluginMarketplacesEnabled  # Desktop key
userPluginUploadsEnabled  # Desktop key
orgPluginSettings  # list format

For identity providers that only return an ID token on refresh when requested for openid again, oidc.scope_on_refresh can be configured. Refresh failure logs also name the step that failed. Aborted requests on Bedrock now use AWS's free CountTokens API to count input tokens—requiring IAM bedrock:CountTokens—instead of sending a request containing only a single token. The two Desktop keys are placed in the desktop policy block. orgPluginSettings is now sent in list format for Claude Desktop 1.15200.0 and later to read; older versions ignore it. If a desktop policy misspells a field in a nested object under managedMcpServers or orgPluginSettings, startup is refused and the exact field is identified.

When a subagent wakes another agent via SendMessage, the completion notification is now correctly sent back to the subagent that initiated the wake-up, rather than mistakenly delivered to the main session. During long API retry waits (such as under CLAUDE_CODE_RETRY_WATCHDOG), agent team transcripts are no longer evicted or wiped blank by retry notices. When running workflows, if a subagent is undergoing long context compaction, it is no longer misidentified as stalled and abruptly restarted. For structured output, Workflow's agent({schema}) now rejects unsatisfiable JSON Schemas up front, and errors hitting the retry limit will include the last validation failure. The 1-hour time limit on background commands started by subagents has been removed; they now run until they exit or are stopped, matching the main session.

Sessions moved to the background no longer appear in ListAgents as a phantom interactive twin with the same name, and the viewer will no longer receive its SendMessage deliveries. Intermittent output swap errors when multiple sessions share the same project directory have also been fixed. When deleting a background session whose worktree has unpushed commits, the first attempt reports the branch name and commit count, and deleting again discards the worktree. Resuming with -p --resume or --continue when Git metadata is lost will fail once, then proceed without the worktree. If a self-hosted runner is configured with --kill-session-after-min, encountering a paused state that is purely waiting for user input will release the session peacefully instead of killing it and reporting a failure.

IDE line selections are no longer lost when triggering a skill or slash command, allowing Claude to receive the N lines selected context. Managed skillOverrides entries keyed on a bundled skill's alias—such as overriding /doctor with checkup—now apply properly, and Skill(name) deny rules fully cover nested skills written in the <dir>:name format. The bundled claude-api skill examples have updated their Go, Java, and C# code to current-generation model IDs, clarifying that cheaper worker or subagent models should also use current-generation models. The maximum wait time for cloud review commands /ultrareview and claude ultrareview has been extended from 30 minutes to 45 minutes.

The VS Code extension brings cleaner interface rendering and session management. The model pill in the footer now directly displays the currently selected effort level, no longer retaining a stale value after switching models, and pill dimensions have returned to their earlier compact size. In the session list's status filter menu, Open and Closed options have been added for faster project categorization and searching. In addition, after Claude Code retries a dropped response, the screen no longer leaves behind half-finished messages, empty tool cards, or an extra Thought for line.

Working across multiple tabs and Remote Control in VS Code is now smoother. Selecting a session from the session history picker that is already open in another tab directly switches to that tab instead of loading the same conversation a second time. Executing the Rename command while a tab view is reloading now reliably applies the new name instead of silently doing nothing. When opening a new session where Remote Control starts automatically, the welcome screen no longer inexplicably disappears. Furthermore, toggling "Enable Remote Control for all sessions" right as a tab is launching now properly applies to the initializing tab.