Claude Code 2.1.222

你在 Team 或 Enterprise 方案替成員申請 /usage-credits，先前那次若被駁回，本來該能再送一次。舊版卻把被駁回的申請誤當成還在處理，跳出「你已送出用量點數申請」並把人擋在門外。現在這扇門會在申請被駁回後重新打開：成員可以提出新的用量點數申請，不必因為上一張失敗的申請一直卡住。這就像申請通道在上一份申請被駁回後重新開放，讓你再送一份新申請。

公司網路走 HTTPS proxy（替你轉送加密網路連線的中介站）時，Claude Code 啟動前的連線檢查曾會卡住，最後才失敗。它現在改用和 API 請求相同、認得 proxy 的連線方式，逾時也會說清楚原因。回覆其實已經完整送達，卻被標成「回覆中途關閉」的誤報也修好了。若你把 ANTHROPIC_BASE_URL 指到自訂 gateway，伺服器明明持續送 keep-alive ping，閒置逾時器以前仍會響；現在會把這些 ping 算進去。像快遞車隔著轉運站仍持續回報位置，收件人不會再被通知包裹半路遺失。

你查 /usage 時，MCP（Model Context Protocol，讓 AI 連接外部工具的協定）伺服器的用量不會再被算得太多：只有實際消耗到它回傳工具結果的請求，才會歸到它名下。claude.ai connector 的工作階段 token（登入驗證用的令牌）失效時，畫面不會假稱還要授權，而會提示你用 /login 登入。若你剛移除一台 MCP 伺服器，原本看不到的工具錯誤也會顯示出來；SendMessage 遇到太長的摘要則會截短後送出，不會撞上字數上限直接失敗。把它想成客服中心的案件流程：只有案件確實採用了某櫃台提供的資料，該櫃台才計入用量；通行證失效就附上 /login 指示；撤掉櫃台後，舊案件的錯誤紀錄仍可查看；交接摘要太長，先裁到收件欄上限再送出。

組織限制某個模型家族時，子代理或隊友指定 model: opus 這類別名，先前可能錯退回父代理的模型。現在它會在同一家族裡往下改選組織允許的最新模型，選擇和限制終於對得上。子代理逐字稿的 spinner（畫面上轉圈的等待動畫）標籤改了。它會顯示子代理自己的 effort:（思考力道／投入程度的設定），不再顯示整個工作階段的等級。若主機用 CLAUDE_CODE_PROVIDER_MANAGED_BY_HOST 管理模型，主機的選模鍵會壓過磁碟裡過期的 managed-settings.json。另外，UltraPlan 功能已移除。像派工板終於標出每個人的實際班別，總部的最新規則也不再被檔案櫃裡的舊紙條蓋過。

你正在輸入問題又按了退格鍵，使用 --ax-screen-reader 模式的螢幕閱讀器以前會把整行輸入內容重新念一遍；現在刪除行尾文字時只回應剛刪掉的字元，耳朵不用反覆聽同一句。另一個少見但惱人的中斷也補好了：檔案監看器碰到檔案系統錯誤，或正在結束監看時，Claude Code 可能會當掉。現在這兩種時刻不再引發罕見崩潰。像校對員擦掉句尾一個字，只念那個被擦掉的字；門口的值班員遇到小狀況，也不會整間辦公室關燈離開。

On a Team or Enterprise plan, a member whose earlier /usage-credits request was dismissed should be able to submit another one. Instead, Claude Code treated that dismissed request as if it were still pending, showed “you've already sent a usage credit request,” and blocked the new request. That record now stops blocking the member after dismissal. It is like an application channel reopening after a dismissal, so you can send a new request.

If your company network uses an HTTPS proxy, Claude Code's startup check used to hang and then fail. It now uses the same proxy-aware transport as API requests and gives a clear timeout message. The false “Connection closed mid-response” report is also gone when a response had already finished. And when ANTHROPIC_BASE_URL points to a custom gateway, the idle timer now recognizes keep-alive pings arriving from the server instead of timing out anyway. Think of a parcel still reporting its location through a transfer station: the recipient should not be told it vanished halfway there.

When you open /usage, an MCP server is no longer charged for every turn after someone called it. Its share now includes only requests that actually consumed its tool results. If a claude.ai connector has an invalid session token, it shows a /login hint rather than wrongly asking for authorization. Tool errors also remain visible after the local MCP server that supplied a tool has been removed. Finally, SendMessage truncates an overly long summary instead of failing on its character limit. Think of a customer-service desk's ticket workflow: only a ticket that actually used a desk's supplied information counts usage for that desk; an invalid pass gets a /login instruction; an old ticket's error record stays viewable after its desk closes; and an overlong handoff summary is trimmed to fit the receiving field before it is sent.

When an organization restricts a model family, a subagent or teammate using an alias such as model: opus could previously fall back to the parent agent's model. It now steps down within that family to the newest model the organization allows. A subagent transcript's spinner effort label now shows that subagent's own effort: setting, not the session-wide level. When the host manages models through CLAUDE_CODE_PROVIDER_MANAGED_BY_HOST, its model-selection keys now beat stale on-disk managed-settings.json values. UltraPlan has also been removed. It is like a staffing board finally showing each person's actual shift, while the current rule from headquarters no longer loses to an old note in a filing cabinet.

While you are typing and hit Backspace, a screen reader in --ax-screen-reader mode used to read the entire input line again. Deleting at the end of a line now echoes only the characters just deleted, so you do not have to hear the same sentence repeatedly. Another rare interruption is fixed too: Claude Code could crash when its file watcher met a filesystem error or while that watcher was shutting down. Those moments no longer cause the rare crash. It is like a proofreader who erases one character at the end and reads only that character aloud, while the person watching the door does not switch off the whole office over a small incident.