發布日期：2026-05-27
你的 Codex 面板現在不會因為幾次中斷的對話就整塊變成空白了。以前如果你連續跑了幾個極短的對話（例如剛跑就按 Ctrl-C 取消），系統往前找資料時會被這幾個「空包彈」洗版，找不到額度資料就直接罷工顯示「--」；現在它會往回翻找更久的紀錄，確實把你該有的配額顯示出來。此外，以前專案資料夾名稱裡如果有連字號（例如 my-awesome-project），系統會自作聰明把它切碎，讓你認不出自己的專案；現在它學會了正確拼湊連字號，你的專案終於能用原本的名字見人了。就像你去超商查點數，以前店員只要看到你最近五筆消費都是 0 元，就會直接告訴你查不到資料，而且還會把你的名字念錯；現在他會往前多翻幾頁找出你真正的餘額，而且終於把你的名字念對了。
現在你用終端機打開用量表時，它的語言會跟選單列完全一致。以前終端機有自己的小脾氣，只認得繁體中文跟英文，如果你用日文或韓文介面，它還是會硬塞英文給你；現在它們終於說同一種語言了。同時，我們解決了常駐程式越跑越吃資源的問題。以前系統會把你看過的所有歷史紀錄都永遠塞在記憶體裡，對話越多，它佔用的記憶體就越肥；現在我們幫它裝了個「只能記 512 筆」的腦袋，塞滿了就會自動忘掉最舊的。就像圖書館以前把所有借回來的書都堆在櫃檯不還回去，導致櫃檯被塞爆；現在櫃檯規定只能放 512 本書，有新的書進來，就會把最舊的書乖乖放回書架上。
Your Codex panel will no longer mysteriously go blank just because you had a few brief sessions. Previously, if you ran several quick commands that were interrupted (like pressing Ctrl-C), the system would only check the last 5 logs. Since those short "dud" sessions lacked quota data, the system gave up and displayed "--"; it now looks much further back to find your actual rate limits and model info. Additionally, if your project folder name contained hyphens (like my-awesome-project), the system used to mangle it into an unrecognizable path; it now correctly decodes hyphens, displaying your project's true name. It's like checking your loyalty points: the cashier used to see five $0 transactions and claim you had no account while mispronouncing your name; now they flip back further to find your actual point balance and finally call you by the right name.
When you open your usage tables in the terminal, the language now perfectly matches your menu bar. Previously, the terminal interface was a bit stubborn—it only understood Traditional Chinese and English, and would force English on you even if you had Japanese or Korean selected in the main app. Now, both interfaces speak the exact same language. At the same time, we've fixed a major resource leak in the background app. The system used to hoard every single chat history log you ever loaded in its memory forever, causing its memory footprint to endlessly balloon over time. We've now installed a strict "512-item limit" to its brain, automatically clearing out the oldest data when full. It's like a library front desk that used to pile up every returned book forever until the desk collapsed under the weight; now, the desk holds a strict limit of 512 books, automatically shelving the oldest ones when a new one arrives.
**Codex usage panel no longer falls back to `--` after a burst of short sessions**: `codex_loader.load_rate_limits()` only scanned the 5 most recent jsonl files via `_recent_jsonl_files()` to find rate_limits. Codex CLI (observed on 0.134.0) writes `payload.rate_limits == null` for short or interrupted sessions (a quick `codex exec` run, Ctrl-C, etc.); when the latest 5 sessions all fall into that bucket, the genuinely-valid prior session gets evicted from the lookup window and the entire Codex block in the popover / TUI renders as `--`. The scan window is widened from 5 to 30 (covers a typical 1–2 day usage range); the first non-null result still early-returns, and the `primary.used_percent` / `secondary.used_percent` parsing path is unchanged. The new Codex CLI 0.134.0 schema fields (`limit_id`, `limit_name`, `credits`, `plan_type`, `rate_limit_reached_type`) are deliberately not parsed — UI doesn't use them. Three new tests cover the "5 null then 6th valid", "all 30 null returns None", and "pick most recent valid" scenarios.
**Dashed Claude Code project names now decode correctly**: `history_loader._project_from_path` previously replaced every `-` in the encoded directory name with `/`, so `Desktop-claude-tutorial-video` would become `/Desktop/claude/tutorial/video` — a non-existent path. `resolve_project_name`'s fallback then took the last segment, mis-labeling the project as `"video"` instead of `"claude-tutorial-video"`. The decoder now tries the all-slash candidate first; on miss, it DFS-walks the segments, joining adjacent ones with `-` and preferring whichever variant actually exists on disk. When nothing matches, the encoded name (minus the leading `-`) is kept as-is so dashes round-trip (`plain-project` stays `plain-project`). For most users, the JSONL `cwd` field already overrides the project name, so this primarily fixes older entries that lack `cwd`.**TUI language detection routed through `usage_lang.detect_lang`**: `tui.py` had its own detector that only returned `zh-TW` or `en` (treating simplified Chinese, Japanese, and Korean as English), and ignored `USAGE_LANG` / `TT_LANG` / `LANG` entirely. The menubar already used `usage_lang.detect_lang()`, so the same machine could show Japanese in the menubar and English in the TUI. The TUI now shares the same detector — all five languages render consistently.
**LRU cap on history / codex loader caches**: `_file_cache` and `_jsonl_cache` were unbounded module-level dicts. As `~/.claude/projects/` and `~/.codex/sessions/` accumulated more jsonl files over time, the menubar's resident memory grew without bound — parsed `UsageEntry` lists never got released. Both caches are now `OrderedDict`s with a 512-entry ceiling: cache hits `move_to_end` to mark MRU, inserts on a full cache `popitem(last=False)` the oldest. The mtime/size invalidation logic and codex_loader's `entry.model` rebind on cache hit are unchanged.