發布日期：2026-09-29
當你暫停工作後想接續先前任務，常因快取（暫存計算結果的機制）過期而重送大量資料。開啟進度管家後，只要以 /resume、--resume 或 --continue 接回過期且脈絡達 5 萬 token（AI 計費的文字單位）的對話，Claude Code 會在輸入首句前顯示一行提示，列出閒置時間、預估重送的 token 數並建議先執行 /compact。此提示僅自己可見且不計入脈絡，需 2.1.251 以上版本且下次啟動 usage 自動套用。狀態列在倒數結束後也不再留白，到期當下會即時重畫，顯示如 cold·re-cache 180k 的提示，表示下一則訊息約有 18 萬 token 要重新快取。
埋頭趕工或專注思考時，如果沒有時常緊盯著用量面板，額度（特定週期內可使用的調用配額上限）見底的瞬間往往讓人猝不及防。現在系統調整了通知音效，當額度快用完或已經用完時，會改播 usage 自製的兩聲「uh-oh」簡短警示音，讓你光聽聲音就能分辨用量告急。至於額度恢復時的通知，則繼續維持播放系統預設的提示音，避免干擾既有的使用習慣。
習慣呼叫多個背景助手平行處理繁重任務的使用者，過去看到的用量統計可能嚴重失真。先前在串流（答案一邊生成一邊逐字傳輸的輸出方式）處理時，Claude Code 會把同一請求多次寫入日誌，導致解析器只抓到第一行，使子代理的輸出用量少算約 86%。例如在某台機器上的 17 個子代理日誌中，原本僅記錄到 23,327 個輸出 token（AI 處理與計費的文字單位），修正後改取每筆請求的最大值，實際數值應為 165,395。這項問題只影響子代理，主對話日誌完全不受影響；更新後歷史快取與年度快取會自動重算一次，但日誌已遭刪除的日期則維持封存數字無法補正。
遇到網路斷線或離線環境工作時，最怕計費工具因連不上伺服器而無法正確計算新型號的花費。現在離線備援（主要機制失效時接手的後備清單）價目表正式補上了新模型資料，完整納入 claude-sonnet-5-5 的費率。以前當下載的線上價目表過期時，這款模型的用量會被誤算為 0 美元；現在即使離線也能依正確標準計價，終端機表格中也會顯示為清晰的「Sonnet 5.5」，不再出現被截斷的模型識別碼。
When resuming previous tasks after taking a break, users often resend large amounts of data due to expired caches. With Progress Manager enabled, whenever you resume an expired conversation with at least 50k tokens of context using /resume, --resume, or --continue, Claude Code displays a single-line notice before you enter your first prompt. This notice details the idle duration and the estimated number of tokens to resend, while suggesting that you run /compact first. Visible only to you, this prompt is excluded from context, requires version 2.1.251 or later, and applies automatically the next time usage launches. The status line also no longer goes blank when the countdown ends. Instead, it redraws immediately upon expiration to show prompts like cold·re-cache 180k, indicating that the next message will re-cache approximately 180k tokens.
When rushing through work or deep in thought, running out of quota without watching the usage popover can easily catch you off guard. The system has now updated its notification sounds. Whenever quota is running low or fully exhausted, usage plays a custom two-tone "uh-oh" short alert, letting you recognize the urgent status by sound alone. Notifications for quota recovery continue to play the default system chime, preserving your established workflow habits.
Users accustomed to spawning multiple background assistants for parallel tasks may have previously seen severely skewed usage statistics. During streaming responses, Claude Code wrote the same request to logs multiple times. This caused the parser to capture only the first line, undercounting subagent output usage by about 86%. For example, across 17 subagent logs on one machine, only 23,327 output tokens were initially recorded; after switching to taking the maximum value per request, the actual count was 165,395. This issue affected only subagents, leaving main session logs completely untouched. After the update, historical and annual caches are automatically recalculated once, though dates whose logs were deleted remain archived and cannot be retroactively fixed.
When working offline or during network outages, cost tracking tools often fail to price new models accurately without server access. The offline fallback pricing table now officially incorporates new model data, fully including rates for claude-sonnet-5-5. Previously, when the downloaded online pricing expired, usage for this model was miscalculated as $0. Now it is priced correctly even offline, and terminal tables clearly display "Sonnet 5.5" rather than a truncated model identifier.
**Resuming a stale conversation now tells you what it will cost.** When Progress Concierge is on and you `/resume` (or `--resume` / `--continue`) a conversation whose prompt cache has expired and whose context is at least 50k tokens, Claude Code shows a one-line warning before your first message: how long it sat idle, roughly how many tokens the next message re-sends, and a suggestion to `/compact` first. Only you see it; it is never added to Claude's context. Needs Claude Code 2.1.251 or newer. Existing installs pick up the new trigger automatically the next time usage starts.
**The status line now says what a cold cache will cost.** Once the prompt cache countdown runs out, the segment switches to how many tokens the next message will re-cache (for example `cold·re-cache 180k`) instead of going blank. Claude Code redraws the status line the moment the cache expires, so no refresh interval is needed.Quota warning and quota-depleted notifications now play a short two-note "uh-oh" alert sound made for usage. The quota-restored notification keeps the system default sound.**Claude subagent output was undercounted by about 86%.** Claude Code writes each subagent request to its transcript several times while it streams, and `output_tokens` only reaches its final value on the last line. Both Claude parsers kept the first line, so on one machine 17 subagent transcripts counted 23,327 output tokens instead of 165,395. They now keep the largest value for each request. Main-conversation transcripts were not affected. The history and yearly caches are rebuilt once; days whose transcripts were already deleted keep their archived totals and cannot be corrected.Add Sonnet 5.5 (`claude-sonnet-5-5`) to the offline fallback price table, so its usage is no longer priced at $0 when the downloaded price table is stale, and show it as "Sonnet 5.5" in tables instead of a truncated model ID.