Claude Code 2.1.267

想讓 Claude 深入推理又擔心 token（AI 處理與計費的文字單位）消耗過量，現在能透過 maxEffortLevel 設定為思考深度加上緊箍咒。

以前在 Bedrock、Vertex 與 Foundry 等不同雲端平台上，很難全面鎖定努力程度上限。現在無論在最外層或 modelSettings 針對個別模型指定，都能統一設限，且日常仍可自由選擇更低的努力等級：

{
  "maxEffortLevel": "medium",
  "modelSettings": {
    "claude-opus-4-7": {
      "maxEffortLevel": "low"
    }
  }
}

此外，自訂指令、技能或子代理在 YAML 前言中指定的 effort: 數值，先前在 Opus 4.7、Opus 4.8 與 Fable 5 等固定預設等級的模型上會失效，如今已修復並能正確生效。

正在反覆微調提示詞（prompt）卻苦於每次改動都得開新對話，新的 --system-prompt-snapshot off 旗標正是為此而生。當你在打磨自訂提示詞內容時，系統預設會沿用工作階段初次錄製的提示詞快照，導致你在外側修改了檔案文字，當前的對話依然讀取舊版內容。現在只要在啟動 Claude 時帶上這個選項，每一次發出提問都會重新讀取並渲染最新的系統提示詞，讓文案調校的測試過程變得更加靈活迅速。

重新打開先前未完成的工作階段時，最怕背景暫存全部失效導致推論延遲。以前從對話分岔出的背景工作單元若把 EnterWorktree 這個工作樹（Git 讓同一專案平行開多份副本的功能）工具加進對話的工具清單，便會破壞快取重用機制。現在恢復工作階段時，不會再因 MCP（讓 AI 連接外部工具的協定）重連時間差而改寫工具清單或丟失思考，首回合執行工具時也能直接重播既有描述而非重新渲染。此外，從 -p 印出模式切換至互動模式時系統提示詞前綴保持一致。透過 --system-prompt 或 --append-system-prompt 啟動的子代理與工作階段，也會一次記錄提示詞與工具定義，不再每次重新渲染。

發布工件（artifact，在獨立視窗展示的成果文件）若中途網路斷線或遇到格式錯誤，現在不再只回傳不明原因的失敗代碼。在過去，如果上傳檔案時連線中斷，Claude Code 往往只能回報未知結果；如今只要判定傳輸未完成，就會自動重試 1 次以確保順利送達。當頁面檔案並非有效的 UTF-8 編碼，或是包含代表損毀的替換字元（U+FFFD）時，錯誤訊息也不再模糊帶過，而是會具體標示出需要修正的行數與欄位。若工件發布遭到拒絕，畫面亦會清楚解釋被拒原因以及具體的應對處置方法。

恢復先前對話或載入超過 5 MB 的超大歷史紀錄時，整個還原過程變得更乾淨迅速。以前在包含大量 Bash 工具呼叫的對話中執行 --resume，首次渲染畫面常需等待；現在首次渲染更快。當工作階段的對話紀錄超過 5 MB 時，先前重新載入對話容易漏失平行工具呼叫與掛鉤（特定事件發生時被觸發的小程式）輸出，如今資料都能完整重現。透過 -p --resume 執行 /compact 壓縮或其他斜線指令後，對話中也不會再多生出一筆多餘的接續提示。如果工作階段中因切換不同模型或模式而回傳不同的額度週期，用量上限警示也不會再忽明忽暗地閃爍。

讓 Claude 在自動模式或沙箱環境執行工作時，工具的引導說明與安全把關變得更親切嚴謹。以往執行終端命令時，Claude 常只是機械式複誦指令；現在更新了 Bash 工具說明指引，引導它用大白話解釋命令的作用。在沙箱內若遇到 pbcopy 等剪貼簿指令執行失敗，Claude 會主動建議改用內建的 /copy 指令。工作流程中帶有龐大輸出結構的 agent() 呼叫，在自動模式下不再直接被拒，而是交由安全分類器審查。在 macOS 與 Linux 上，外掛市集路徑帶有反斜線可繞過邊界檢查的漏洞已堵死；claude agents 的 @ 選單也能即時列出工作階段開啟後新建立的儲存庫。

透過網頁版 Claude Code 操作 GitHub 企業伺服器時，逾期的 token（授權令牌，用於身分認證）現在會自動換新。

先前在 GitHub Enterprise Server 的工作階段中，只要個人認證憑證到期，介面就會誤標為帳號已斷線；如今在進行拉取請求（PR）或討論議題操作時，系統會自動在背景刷新授權並保持連線。另外，在尚未安裝 Claude GitHub App 的組織架構下，執行 gh 命令或呼叫 API 原先會直接報錯，現在會直接改用你已綁定的 GitHub 帳號發送請求，若尚未登入也會明確提示。

負責管理 Claude Tag 團隊連線的管理者，現在不必刪除重來就能將預設通道切換至自訂連接器。在後台設定預設連線表單時，新加入的「Use a custom connector」連結允許你中途直接轉向自訂通道。當團隊組織的用量額度耗盡時，Claude 過去只會冰冷回報 API 拒絕了無效請求，現在會具體說明額度已用完並指引儲值方法。在頻道頂層發布訊息後，成員在討論串內要求編輯或刪除該則貼文時，指令不再被誤判為對話糾錯，而是確實轉交給當初發文的工作階段處理。此外，管理審查介面的工具存取授權請求點擊「Connect」時，之前常跳出授權失敗或誤報授權包已刪除的問題也已修正。

If you want Claude to reason deeply without burning through excessive tokens, you can now cap its thinking depth with the maxEffortLevel setting.

Previously across cloud providers like Bedrock, Vertex, and Foundry, it was difficult to uniformly lock effort level caps. Now, whether configured at the top level or per model under modelSettings, limits apply consistently while still allowing you to pick a lower effort level during daily use:

{
  "maxEffortLevel": "medium",
  "modelSettings": {
    "claude-opus-4-7": {
      "maxEffortLevel": "low"
    }
  }
}

Additionally, effort: values specified in YAML frontmatter for custom commands, skills, or subagents previously failed on models with pinned default effort levels (such as Opus 4.7, Opus 4.8, and Fable 5); this has now been fixed to take effect properly.

If you are iterating on prompt text and tired of starting a new conversation for every tweak, the new --system-prompt-snapshot off flag is built for you. When polishing custom prompt content, the system by default reuses the prompt snapshot recorded at session start, causing active conversations to keep reading stale content even after you edit external files. Now, by passing this flag when launching Claude, every request will reload and render the latest system prompt fresh, making prompt experimentation much faster and more flexible.

When reopening unfinished sessions, cache invalidation causing inference delays is a major pain point. Previously, when a background worker forked from a conversation added EnterWorktree to the conversation's tool block, it broke prompt-cache reuse. Now when resuming a session, tool lists are no longer rewritten and thinking is no longer dropped due to MCP reconnection timing differences; turns executing tools on the first step also replay recorded descriptions directly instead of re-rendering them. Additionally, the system prompt prefix remains consistent when resuming a print-mode (-p) conversation interactively. Subagents and sessions launched with --system-prompt or --append-system-prompt now record prompt and tool definitions once instead of re-rendering them.

When publishing artifacts, network drops or formatting errors no longer return vague failure codes. Previously, if a connection was cut off mid-upload, Claude Code often reported an unknown outcome; now, whenever it detects that an upload never completed, it automatically retries once to ensure delivery. When a page file is not valid UTF-8 or contains replacement characters (U+FFFD) indicating corruption, error messages no longer gloss over the issue, instead explicitly naming the line and column to fix. If an artifact publish is refused, the screen also clearly explains the reason along with actionable remediation steps.

Resuming previous conversations and loading massive transcripts over 5 MB is now cleaner and faster. Previously, running --resume on sessions with many Bash tool calls caused noticeable delays on first render; first-render time is now shorter. When a session transcript exceeds 5 MB, reloading used to drop parallel tool calls and their hook output; all of it now comes back intact. Running /compact or other slash commands via -p --resume no longer inserts an extra spurious continuation prompt. If switching models or modes within a session returns varying quota windows, usage-limit warnings no longer flicker on and off.

When running Claude in auto mode or sandboxed environments, tool descriptions and security checks are now clearer and stricter. Previously when executing terminal commands, Claude often mechanically echoed the command; updated Bash tool guidance now steers it to explain what commands do in plain words. Inside the sandbox, if clipboard commands like pbcopy fail, Claude proactively suggests using the built-in /copy command. In workflows, agent() calls with large output schemas are no longer flatly refused in auto mode, routing to the safety classifier for review instead. On macOS and Linux, a case where a marketplace entry path containing a backslash could bypass the containment check has been closed, and the claude agents @ menu now lists repositories created after session start.

When working with GitHub Enterprise Server via Claude Code on the web, expired tokens now refresh automatically.

Previously in GitHub Enterprise Server sessions, expired authentication credentials caused the interface to report the GitHub account as disconnected; PR and issue operations now automatically refresh tokens in the background to stay connected. Additionally, in organizations without the Claude GitHub App installed, running gh commands or making GitHub API calls previously threw errors, but now routes through your connected GitHub account and explicitly notes when none is connected.

Administrators managing Claude Tag team connections can now switch preset connections to custom connectors without starting over. When configuring preset connection forms in admin settings, a newly added "Use a custom connector" link lets you pivot directly to custom channels. When an organization runs out of usage credits, Claude previously replied with a generic message stating the API rejected an invalid request; it now explicitly explains credits are exhausted and details how to add more. After Claude posts a message at a channel's top level, thread requests to edit or delete that post are no longer treated as conversational corrections, routing properly to the session that originally posted it. Additionally, clicking **Connect** on Tool access requests under Admin settings > Review requests previously failed with "Authorization failed" or showed the access bundle as deleted; this has now been resolved.