Usage 0.29.17: Cost estimates stay accurate when you're offline

Release notes

FEATURE

Cost estimates stay accurate when you're offline

When you are away from the internet and checking what your usage cost, the number on screen needs to be dependable. It was not: Opus 4.6 and 4.7 were priced at three times their rate, while Opus 5 appeared to cost nothing. If LiteLLM's price table could not be downloaded, or its cache (saved lookup results kept for reuse) had expired, the app used a hand-maintained fallback table that had drifted from the current one. It is now reconciled with LiteLLM's cached table: 4.6 and 4.7 move from $15/$75 to $5/$25 per million billing units, Haiku 4.5 regains its missing twenty percent, and Opus 5 no longer returns $0. Think of an offline checkout pulling out an old price sheet; it now has the right sheet when you need it most.

BUGFIX

Traditional Chinese screens had 52 Simplified Chinese strings mixed in, now fixed

If you chose Traditional Chinese and opened the text interface (TUI, a screen operated in a terminal with text), you could meet Simplified Chinese such as 检测到 and 时间—like opening a book only half proofread. Fifty-two strings had been copied from the Simplified Chinese zh-CN locale without conversion. A hand-written list of Simplified characters had caught only some of them, missing even 来 and 项. Every value now went through OpenCC's s2twp conversion and each difference was reviewed. Nine suggested replacements, including 發送 to 傳送, were left alone because the existing text was already correct. Only two changed strings in the HTML test snapshot were patched, preserving its fixed timestamp and version. Japanese and Korean each still have sixty-three empty strings; those are missing translations, not this bug, so they remain for later work.

FEATURE

Menu-bar refresh work moves house without changing behavior

This is not a new button: your menu-bar refresh still looks the same, but its 203-line back-room job has moved. menubar.py had been split apart twice and then grown again, leaving only four lines below its size limit. Instead of raising that limit, the data work of the background refresh moved to menubar_refresh.py, divided into load_sources() and build_result(). The three performSelectorOnMainThread calls and every selector (a name used to locate a method to run) stay in the original file, because AppDelegate methods are bound by name and moving them would break that connection. The new module describes the App it needs through typing.Protocol and imports no PyObjC. It is like moving prep work from an overcrowded kitchen into a side room: the service window stays put and meals follow the same path. Six timing stages, fifteen degraded-path state assignments, and the guard against a failed refresh getting stuck were compared line by line.

Original changelog

Original changelogCost estimates stay accurate when you're offline

**The offline price table overcharged Opus 4.6 and 4.7 threefold, and valued Opus 5 at nothing.** `_fallback_pricing()` is what the cost estimate falls back on when the LiteLLM table cannot be downloaded or its cache has expired. It is written by hand and nothing ever reconciled it against the live table, so it drifted in silence: `claude-opus-4-6` and `claude-opus-4-7` still carried the Opus 4.1-era $15/$75 per Mtok long after Anthropic dropped Opus to $5/$25, `claude-haiku-4-5-20251001` was twenty percent low, and `claude-opus-5` was missing outright — and a missing key makes `calculate_cost()` return `0.0`, so Opus 5 usage was reported as costing nothing. The corrected figures come from reconciling against the cached LiteLLM table rather than from the changelog: Claude Code 2.1.219 announced Opus 5 at `$10/$50 per Mtok`, but that is the fast-mode price, not the standard one. What makes this class of bug worth naming is when it surfaces — only offline, which is exactly the moment nobody thinks to doubt the number on screen.

Original changelogTraditional Chinese screens had 52 Simplified Chinese strings mixed in, now fixed

**Fifty-two keys in the Traditional Chinese locale contained Simplified Chinese.** The TUI and report strings had been copied across from `zh-CN` without conversion, so a Traditional Chinese user opening the TUI read `检测到`, `时间`, `来源`, `项目` and `消息`. A hand-written list of Simplified characters found only some of them — it missed `来` and `项` — so every value was instead run through OpenCC's `s2twp` and the differences reviewed one by one. Nine of those differences were the converter overreaching on text that was already correct (`發送`→`傳送`, `吹回`→`吹迴`, `連接詞`→`連線詞`) and were left alone. The golden HTML snapshot was patched for the two strings that actually changed rather than regenerated, because regenerating it would also replace the frozen timestamp and version the test asserts on. The same region of `ja` and `ko` holds sixty-three empty strings each; those are missing translations rather than wrong ones and are left for a separate change.

Original changelogMenu-bar refresh work moves house without changing behavior

**`_refresh_in_background` moved into `menubar_refresh.py`.** At 203 lines it was the largest method in `menubar.py`, which had four lines left under its ceiling — and a ceiling exists on that file precisely because it has been split apart and grown back twice. Only three lines of the method touch PyObjC, the three `performSelectorOnMainThread` calls, and those stay behind along with every selector: an AppDelegate method is bound by name, so relocating one breaks the binding. What moved is the data work between them, split into `load_sources()` and `build_result()`, with the app passed in behind a `typing.Protocol` so the new module imports no PyObjC at all — the same rule that keeps `menubar_state.py` clean. `menubar.py` goes from 2116 lines to 1934 and the ceiling was lowered to match rather than raised, since raising it to get CI green would dismantle the only thing stopping the file from growing back a third time. Behaviour is unchanged, and that was verified rather than assumed: the six `refresh_timing` stages, the fifteen state assignments on the degraded path, and the `submitted`/`finally` guard that keeps a failed refresh from wedging the app were each compared line by line against the original.