RELEASE NOTES

Every release, a little better.

From a number in the menu bar to visible usage trends — macOS stable releases and Windows betas are documented here.

70 releases · Original release notes preserved

v1.39.3Latest macOS stable

v1.39.3 — Reset alerts: 24-hour window and smarter routes panel

v1.39.3 — Reset alerts: 24-hour window and smarter routes panel

macOS 1.39.3 (build 86). Windows remains its separate v1.38.6-beta.1 prerelease.

  • The reset-alert panel now shows only the last 24 hours: the unread count, pinned signal and clock history share that window. Older announcements stay archived locally for 30 days for revision deduplication, but never appear in the UI.
  • Each message gains a Dismiss action: it disappears from the panel immediately and the same revision cannot return on the next poll; a newer revision is still delivered normally.
  • The right-click panel's OpenAI routes section now starts collapsed: it auto-expands only when OpenAI connectivity is unreachable and collapses again once reachable. "Follow running apps" now aligns only the connections section, so a manually expanded routes section is no longer folded back.
  • When "Restore on input" lacks Accessibility permission, the Dim controls show an orange hint with an Allow button that opens the Accessibility pane in System Settings; pointer activity keeps restoring without that permission.

No display values, provider credentials or input contents are uploaded. The activity listener only triggers brightness restoration and does not store keys or pointer data. Reset alerts keep reading public feeds only and never send Codex credentials.

Install the PKG to replace the app in Applications. Requires macOS 14+ on Apple Silicon. The app is ad-hoc signed; the installer is unsigned and not Apple-notarized.

View this release on GitHub ↗

v1.39.2 — Dim restores on local input by default

v1.39.2 — Dim restores on local input by default

macOS 1.39.2 (build 85). Windows remains its separate v1.38.6-beta.1 prerelease and does not yet include per-display Dim.

  • Renames the compact Dim option to Restore on input and enables it by default when no prior choice exists.
  • Built-in trackpad/keyboard and Bluetooth mouse/keyboard movement, clicks, scrolling and key presses restore the saved brightness. Modifier keys are included.
  • Keyboard detection requires macOS Accessibility permission. Pointer activity still restores without that permission.
  • Turning the option off keeps the existing manual paths: click Restore or hold the menu-bar icon's right button.
  • An explicitly saved enabled or disabled choice continues to win after an upgrade.

No display values, provider credentials or input contents are uploaded. The activity listener only triggers restoration and does not store keys or pointer data.

Install the PKG to replace the app in Applications. Requires macOS 14+ on Apple Silicon. The app is ad-hoc signed; the installer is unsigned and not Apple-notarized.

View this release on GitHub ↗

v1.39.1 · Quota rings and menu account masking

v1.39.1 · Keep quota rings visible during alerts

macOS 1.39.1 (build 84). Windows remains the separate v1.38.6-beta.1 prerelease.

  • Unread Codex reset updates preserve all quota rings. A persistent yellow dot replaces C, and only the Codex unit alternates between yellow and normal colors every 0.75 seconds. Quota values continue updating; Reduce Motion holds the yellow frame. Dim remains visible. Task waves pause until all alerts are acknowledged. If hidden by display selection, Codex is temporarily appended.
  • Adds optional Settings → General → Account display → Mask accounts in menus, off by default. Both menus mask names and emails (abcdefg@example.com → a****g@example.com), including tooltips, delete prompts and external-switch details. Renaming pauses while masked and action summaries omit identities. Settings pages are outside the scope; domains and some name characters remain visible.
  • Masking changes presentation only: account identity, switching, collection and sync retain their original data. Mobile Dashboard masking stays independent. These two features are macOS-only.
  • Adds explicit --test-reset-alert launch argument for an in-memory, clearly labeled local preview. It uses the real alert path and disappears when acknowledged or on restart, without changing real announcements.
  • The bilingual website now demonstrates yellow-dot ring alerts and documents menu masking, platform limits and current downloads.

Install the PKG to replace the app in Applications. Requires macOS 14+ on Apple Silicon. The app uses ad-hoc signing; the installer is unsigned and not Apple-notarized. Public reset announcements are third-party reports, not confirmation that your personal quota reset. System notification presentation depends on macOS settings.

View this release on GitHub ↗

v1.39.0 — Codex reset inbox and visible Dim state

v1.39.0 — Codex reset inbox and visible Dim state

For macOS 14+ on Apple Silicon. Windows remains the separate v1.38.6-beta.1 prerelease; these Mac features are not included in its existing ZIP.

Public reset announcements with explicit acknowledgement

  • The left-click menu follows codex-reset.com and codex-resets.com, preserving source links and distinguishing clues, previews, reported resets and Banked messages. A third-party announcement does not prove your individual account has reset.
  • Unread messages replace the quota indicator with an orange bell, alternating brightness every 0.75 seconds or staying steady with Reduce Motion. New messages attempt a system notification; reminders repeat every five minutes until explicitly acknowledged.
  • Opening a source or dismissing a notification does not mark messages read. Acknowledge one or all messages without closing the panel. Source links close the panel before opening a browser. An old notification cannot acknowledge later message revisions.
  • History is always accessible and covers 30 days. Older unread entries remain until acknowledged, and the latest explicit preview, reported reset or Banked signal stays visible after acknowledgement without restarting reminders. State survives app restarts.
  • Three feeds are polled independently. A failed source does not clear unread entries; meaningful announcement updates become unread again. System notifications depend on permission and Focus settings; menu-bar alerts continue independently. Test-alert controls are absent from the release.

Visible Dim state during remote access

Step Away Dim adds a steady screen-and-moon badge to the menu bar, including during quota animations and reset alerts. Restoring the displays removes the badge. This is a state indicator without a quota progress ring.

Menu and website updates

The menu material, rounded outline and window shadow stay aligned as content resizes in the same panel. Bilingual website and user guides now explain recent brightness controls, account sources, credential mirroring and Windows differences. Windows coverage includes installation and upgrades, Codex CLI fallback, Watch, Codex task protection, mobile monitoring and actual limitations. The changelog labels the independent Windows beta while Mac stable and Windows versions remain separate.

Repository rules require assessing website impact before pushing, synchronized bilingual content and version checks, and production verification after deployment. See the release checklist for actual test, packaging, checksum and publication results.

Distribution retains ad-hoc app signing, without Developer ID installer signing or Apple notarization.

View this release on GitHub ↗
v1.38.6-beta.1Windows beta

v1.38.6-beta.1 — Windows tray visibility, settings layout and Codex Watch

v1.38.6-beta.1 — Windows tray visibility, settings layout and Codex Watch

Windows-only prerelease for x64 Windows 10+. Includes the merged macOS v1.38.5 source; the macOS stable release remains v1.38.5.

  • Both tray modes now have a dark backing and light border. Detailed mode displays provider and quota on two lines, with full details in the tooltip.
  • Settings rows size to their content and wrap long captions. Settings, Routes and Connections have an explicit 96 DPI layout baseline; the control center rechecks the primary monitor DPI. Full multi-monitor and scaling acceptance remains outstanding.
  • Settings → Providers now includes Codex account sharing and watching: select local accounts to share, add/remove watched emails, reject invalid or duplicate addresses, refresh, and inspect status, all quota windows, publication time and read counts.
  • Watch changes save immediately and request synchronization without waiting for the regular ten-minute poll. Cancel does not undo Watch edits. Existing sharing and watched-email settings are preserved.
  • Sharing stays off by default. Anyone who knows a shared email can read its published quota. Provider credentials are never uploaded. Pool accounts without fetched local quota must be signed in and refreshed before publishing.

Download AIQuotaBar-win-x64.zip, extract it, and run AIQuotaBar/AIQuotaBar.exe. The package includes .NET 8; no installer or code signing is provided. Quit the running app before replacing its program files. Keep your %APPDATA%\AIQuotaBar settings and vault.

Verification: 543 Windows tests and 124 website tests passed; website syntax checks passed. macOS tests were not rerun on Windows. Account login management, provider data-source selection and the macOS credential mirror are not ported in this release.

The Windows release is marked prerelease and does not replace the macOS stable releases/latest. The website's stable changelog intentionally excludes prereleases; its Windows download cards link directly to this release.

View this release on GitHub ↗

v1.38.5 — Credential guard and background performance

macOS feature release, build 81.

  • Every Codex account seen on this Mac keeps a copy of its credentials outside ~/.codex: ~/Library/Application Support/com.techfanseric.aiquotabar/codex-auth-mirror/<account digest>.json, directory 0700, files 0600, atomic writes. The directory is created on the first write.
  • The mirror follows ~/.codex/auth.json and every backup in the account pool. It only follows: switching still happens by atomic rename between auth.json and the pool, and mirrored files are never read back, moved, or rewritten.
  • The app watches both the ~/.codex directory and the auth.json file itself, and syncs on any change; a 10-second background poll backs it up. The moment a gateway (magpie and friends) swaps the sign-in, that account's credentials are already safe.
  • When an external tool rewrites the sign-in, the accounts section says so and names who was switched away and who came in. Switches this app performs itself are not counted as external.
  • Also fixes a guard that never worked: URL.resourceValues(forKeys:) caches resource values on the same URL instance — writing one file three times in a row returned the same modificationDate every time. The tear-read protection compared that cached value with itself, so it always passed and never guarded anything. It now uses a fresh stat.
  • Account switching, panel rendering, and cloud sync are unchanged; the watch-account and team channels are independent and unaffected.

See Credential guard mirror.

Lower background work

  • Kimi and MiniMax use file events and incremental history checks, with full-scan recovery and periodic reconciliation.
  • Codex reuses unchanged, sanitized task metadata. Menu-bar rendering skips unchanged state and closed-menu layout; display discovery follows configuration events.
  • Refresh intervals, task detection and animation playback cadence are preserved.
  • In a five-minute active-task measurement on one Mac, average CPU fell from 21.76% (60-second baseline) to 12.12%. This is not an idle benchmark or a guarantee: a later active-task window measured 17.65%. Idle below 1% has not been demonstrated; native menu-bar animation redraws remain a hotspot.
  • GPU-only animation experiments are not shipped because menu-bar replica playback and appearance equivalence are unverified.

See the performance audit for memory measurements, methodology and limitations.

View this release on GitHub ↗

v1.38.4 — 提高低额度预警的辨识度

v1.38.4 — 提高低额度预警的辨识度

macOS 修复版,build 80。

  • 低额度预警由偏暗的黄色调整为饱和橙色:浅色外观使用深橙色,深色外观使用明亮橙色。
  • 预警不再只显示在剩余额度那一小段弧线上:顶部供应商字母、中心有效图形也使用预警色,让接近 0% 时仍能辨识预警。
  • 故障与离线状态的供应商字母同步使用红色,正常状态继续使用系统模板着色;额度和节奏的几何含义保持不变。
  • 77 项相关测试通过,覆盖浅色/深色外观、0%/1%/5%/20% 低额度可见性、故障优先级、警示解除后的模板恢复,以及两种任务动画。
View this release on GitHub ↗

v1.38.3 — 菜单栏图标跟随系统着色

v1.38.3 — 菜单栏图标跟随系统着色

macOS 修复版,build 79。

  • 正常、加载、待配置与中性占位图标改用系统模板着色,跟随菜单栏前景色,修复深色背景下圆环仍保持深色的问题。
  • 低额度、故障、离线和自检继续保留黄色、红色警示。多个供应商合并显示时,只要存在警示就保留彩色图像;警示解除后恢复系统模板着色。
  • 两种任务动画布局使用透明缺口绘制剩余额度段上的光带,避免模板着色后动画与圆环融在一起。
  • 76 项相关测试通过,覆盖浅色/深色外观、警示颜色与恢复、两种任务动画、占位图标和周趋势。
View this release on GitHub ↗

v1.38.2 — 修复亮度控制卡住,缩短快捷键响应

v1.38.2 — 修复亮度控制卡住,缩短快捷键响应

macOS 修复版,build 78。

  • 修复外接屏关机、切换输入源等情况下,后台显示器服务创建卡住并占用亮度控制队列的问题。服务发现移到控制队列之外,复用有效连接,并限制服务创建的等待时间。
  • 后台扫描不再被误显示为仍处于 Dim;Dim 保持任务执行期间,手动恢复按钮仍可操作。
  • 连续亮度键调节复用短期亮度读数,硬件亮度不变时只更新 Gamma,减少重复枚举、无效写入及逐次读回确认,改善按键响应。
  • 修复反复扫描显示器时遗漏释放 IORegistry 根句柄的问题,减少长时间运行后的资源累积。
  • 77 项相关测试通过,包含服务超时隔离、连接复用、连续调节的读写次数,以及保持任务执行中的手动恢复。修复版已在本机安装,用户确认重新授权后亮度键正常。

若更新后亮度键只控制内置屏,请在「系统设置 → 隐私与安全性 → 辅助功能」中,将 AI Quota Bar 关闭后重新开启,让更新后的应用重新接管亮度键。

View this release on GitHub ↗
v1.38.2-beta.1Windows beta

v1.38.2-beta.1 — Windows flyout countdown no longer goes negative

v1.38.2-beta.1 — Windows flyout countdown no longer goes negative

Windows prerelease, build 77, cut from the v1.38.1 code with Windows-side fixes only.

  • Fixes the negative countdown in the tray flyout header: when a refresh is overdue because of slow or failing fetches, the header showed "next in about -17 minutes"; it now shows "refreshing any moment" while overdue, and keeps "next in about N minutes" when on schedule.
  • Packaging fix: the previous AIQuotaBar-win-x64.zip was actually a tar stream with a zip extension that Windows 10 Explorer cannot extract; from this release on it is a real zip (same layout — a top-level AIQuotaBar/ folder; unzip and run AIQuotaBar.exe).

Verification

  • All 529 tests pass on Windows 10 19045.
  • Verified on a 200%-scaled display: the flyout renders without scaling artifacts — title row, chart axes, usage bars and footer links all aligned; the window is 360 DIP wide, hugging the right edge above the taskbar.

Everything else matches v1.38.1; macOS is unaffected, releases/latest keeps pointing at v1.38.1, and this prerelease stays out of the changelog feed.

View this release on GitHub ↗

v1.38.1 — 修复 Dim 后外接屏无法调亮

v1.38.1 — 修复 Dim 后外接屏无法调亮

macOS 修复版,build 76。

  • 修复外接屏的硬件亮度已经升高、画面却仍停留在 Dim 压暗状态的问题。调暗后的 15% 色表不再被当作正常亮度基准,也不会形成调亮上限。
  • 兼容修复 v1.38.0 保存的错误色表基准,保留各颜色通道的相对关系与曲线形状。
  • 修复 Dim 尚在读取或写入亮度时按亮度键的竞态:等待 Dim 完成后,先恢复原亮度,再执行键盘调节。
  • 72 项相关测试通过,包含错误基准的恢复、连续调亮及 Dim 进行中按键的回归测试。两个外接屏的本地 Gamma 输出已验证可恢复到 100%。

亮度键仍控制指针所在屏幕,Dim 目标和智能还原偏好保持不变。若更新后亮度键未生效,请在「系统设置 → 隐私与安全性 → 辅助功能」中关闭后重新开启 AI Quota Bar 的授权。

View this release on GitHub ↗

v1.38.0 — 亮度键跟随指针,Dim 支持分屏设置

v1.38.0 — 亮度键跟随指针,Dim 支持分屏设置

macOS 稳定版,build 75。本次发布包含 2026-10-19 的屏幕亮度相关更新。Windows 仍使用独立的 v1.38.0-beta.1 测试版。

亮度键控制指针所在屏幕

  • 键盘的屏幕亮度键只调节鼠标指针所在的显示器,支持长按连续调节。
  • 内置屏使用系统亮度;支持 DDC/CI 的外接屏优先调节硬件亮度,降到硬件最低后用 Gamma 继续压暗。组合范围与 MonitorControl 默认方式一致,最低保留 15% 色表输出。
  • Option + Shift + 亮度键可精细调节。普通亮度键不会改写保存的 Dim 目标。
  • 在 Dim 或设置预览期间按亮度键,先恢复原亮度,再调节所选屏幕,避免两种控制相互覆盖。
  • 「设置 → 通用」新增亮度键开关与辅助功能授权入口。使用此功能时请退出 MonitorControl,避免两个应用同时处理亮度键和 Gamma。

Dim 按屏幕设置与即时预览

  • 每块屏幕可独立保存目标亮度,并显示当前硬件亮度读数。
  • 拖动滑块即时预览所选屏幕;松开后移动鼠标或按任意键恢复,离开设置页也会恢复。不再等待固定的预览倒计时。
  • Dim 按钮旁的「智能还原」默认关闭:未勾选时,再次点击按钮或长按菜单栏图标右键才恢复;勾选后支持输入活动自动恢复。
  • 不创建软件遮罩,不修改显示器的对比度或 RGB 增益。恢复时使用调暗前保存的硬件亮度与原色表,并支持异常退出后的恢复。

UU 远程连接时保持本地亮度

  • 修复 UU 连入或退出时 Gamma 被重置、按钮仍显示“恢复”但本地屏幕变亮的问题。Dim 期间会检查并补回被重置的 Gamma,显示器重新连接后按稳定标识重新匹配。
  • 普通亮度键使用的 Gamma 调暗同样支持重置后的保持;从这种亮度进入 Dim 时按原始色表计算,避免重复压暗。
  • 本机短测确认:补回本地 Gamma 时,UU 远程画面保持正常。未修改 UU 配置。

验证

68 项相关 Debug 测试通过;内置屏及两块外接屏分别做过小幅调节与恢复验证。发布配置的测试编译受已有菜单栏测试调用 Debug 专用方法的问题影响;发布版应用构建正常。

View this release on GitHub ↗
v1.38.0-beta.1Windows beta

v1.38.0-beta.1 — Windows public beta, and the Codex fetch chain fixed

v1.38.0-beta.1 — Windows public beta, and the Codex fetch chain fixed

Windows build 74, a public beta (prerelease). Two things in one release: the Windows edition is downloadable for the first time — the macOS menu bar app rewritten module by module in C# / .NET 8 as a native tray app — and the Codex fetch chain on Windows is brought up to macOS behavior: the system proxy is honored, chatgpt_base_url in ~/.codex/config.toml overrides the base URL, and tokens refresh proactively before expiry. The Mac stable release remains v1.37.0 and is untouched by this one.

The Windows edition, in public beta

Left-click the tray icon for the quota panel (ring gauges, model rows and usage bars follow the Mac design language); right-click for the native command menu. Features were migrated item by item against the specs in docs/win-migration/analysis/:

  • Codex: reads the OAuth / PAT sign-in and the multi-account pool from ~/.codex, with 5h / Weekly dual-window quotas; hides itself when not signed in
  • GLM: the API key is the only required field, with cURL import as an optional advanced path; MiniMax: paste a token in settings and it is on
  • Kimi: desktop / web sessions and CLI /status parsing
  • Clash / Mihomo: local controller auto-discovery, group switching, latency tests, a connection overview and auto-recovery
  • Task protection: PowerCreateRequest keeps the machine awake while a task runs, with a battery safety net
  • Mobile dashboard: read-only HTTP + SSE on port 18765, reusing the original PWA assets
  • Teams: Cloudflare Worker client for create / join / usage reporting / quota upload

Credentials are encrypted with DPAPI (CurrentUser) into vault.bin; silent launch at login, a single-instance guard, light / dark / follow-system themes, bilingual text and file logging are all in. Local verification (Windows 10 19045): all 529 unit / integration tests pass, the Clash integration tests execute against a real controller, and the dashboard and task→keep-awake chains were exercised on the machine.

The Codex fetch chain, aligned with macOS

Three fixes, all of the shape "the Windows port copied the letter of the implementation and dropped the behavior macOS had been leaning on":

  • The system proxy is honored. Codex quota requests and Clash probes now go through the system proxy — networks that only egress via a proxy are no longer bypassed into direct connections by .NET defaults
  • chatgpt_base_url override. A top-level chatgpt_base_url in ~/.codex/config.toml (before the first [section]) is read and used as the API base, the same rule as the macOS fetcher; self-hosted gateway users no longer need extra configuration
  • Proactive token refresh. When the access token's exp is within five minutes, it is refreshed before the request instead of waiting for a 401; the old refresh-once-and-retry on 401 stays. Tokens without an exp claim fall back to an eight-day age cap

The flyout empty state: real errors, not a false "not configured"

When a configured provider fails to fetch, the panel now shows "fetch failed: <reason>", matching the macOS menu's <Provider>: error warning row; the "not signed in / not configured" guidance appears only when credentials are genuinely missing. Before, a network hiccup read as "you have not logged in yet" — and the debugging started in the wrong direction.

Known boundaries (beta)

  • The Windows artifact is a self-contained zip; unzip and run. No installer yet
  • The auto-update channel (/v1/app-update) still points at the macOS artifacts; the Windows side only checks and reminds for now
  • The macOS stable release remains v1.37.0 and is untouched by this one; releases/latest keeps pointing at v1.37.0
View this release on GitHub ↗

v1.37.0 — the watched row stops disappearing, and the tooltip can finally leave

v1.37.0 — the watched row stops disappearing, and the tooltip can finally leave

macOS build 73. This release fixes one recurring mistake: which value a row shows should not be decided by which side claimed the slot first. The last release made local and cloud merge by "take the newer"; this one pulls the watched channel into the same rule — and fixes the long-standing problem of a tooltip that cannot leave the menu.

The watched row no longer disappears because its slot is taken

You add an address to the watch list because you want to see it. At the same moment it probably has another source too: this Mac is signed in as it (~/.codex/accounts/<address>.json), or another device in your team is reporting it. So the same window carries a Cloud row and a Watch row at once.

Rows were de-duplicated by key, and whichever side found the key occupied was dropped — which was the watched one. On a real account, this cost exactly what it sounds like: the watch fetch had succeeded seven seconds earlier, and that number was thrown away in favour of a Cloud row that could be days old. The eye icon went with it, so the row read as "cloud data, not something you watched" — the opposite of what adding the address was for.

The two now merge into one row: the later sample wins, the same rule local-versus-cloud already used. Both sides' sources are kept, so an account that is both watched and cloud-reported draws the cloud and the eye rather than picking one.

The winner's own source segment is left alone. "Cloud" still decides the stale-snapshot date display and the cloud cycle visibility limits, and merging two rows must not quietly switch those off.

The tooltip: ours, not the system's

The account header, the provider header and the cycle chips all explained themselves through .help(). Inside NSMenu that never actually worked: the menu's modal event-tracking loop swallows mouse-moved and mouse-exited for views inside an item, so the tracking area never learns the pointer left, AppKit keeps believing it is still there, and the tooltip window stays on screen long after the cursor is gone.

The hover chain is ours now. Each anchor reports its id together with its frame in one named coordinate space, and a layer at the root of the menu draws the bubble. Reporting only the id would pin the bubble inside the row, where the ScrollView clips it; reporting only the frame would mean threading a callback up through every section.

  • 350ms before a bubble appears. Account rows and cycle chips sit next to each other, and with no delay a sweep of the pointer is a string of empty bubbles flashing past.
  • It flips above the row when there is no room below, rather than being clipped in half by the panel window — a half bubble reads as "stuck and will not go away" too.
  • The bubble takes no hit-testing, or it would steal the hover from the anchor under it and close itself while the pointer rests there.
  • Hover state is cleared when the menu closes, otherwise a bubble pops into existence the next time you open it.

"How long ago" changes gear

A watched account that stopped reporting two days ago read as "Updated 888m ago". The number is arithmetically correct and practically useless: is that 15 hours or three days? Those two call for different actions, and the reader has to do the division in their head first.

The tiers now follow the scale: just now under a minute, minutes under an hour, hours under a day, and past a day the day and the time it was measured. At that point "how long" stops mattering and "which moment" takes over.

The last tier uses an explicit format string with a pinned locale rather than following the system language — otherwise a Chinese interface would print "10/5/26, 2:32 PM", with a width that moves around per region. This string is the interface's language, not the system's.

Known boundaries

  • The merge needs a sample time on both sides; when either is missing the row that was already there stays. "Cannot compare" is not "the watched one is newer"
  • The bubble is capped at 248pt wide and wraps; its height is estimated from the line count purely to decide whether to flip, and is not measured
  • The 350ms hover delay is fixed and does not follow the system's accessibility setting
  • Watch and team cloud remain two independent channels. An account in both still renders as one row — with both icons on it
View this release on GitHub ↗

v1.36.0 — the freshest sample wins, and the icon says where it came from

v1.36.0 — the freshest sample wins, and the icon says where it came from

macOS build 72. The release is one idea: when an account has two sources, show whichever was measured more recently — and say where that number came from.

The freshest sample wins, local or cloud

The account row used to show the local value whenever one existed, no matter how old it was and no matter what the team cloud had just published. So this happened: someone refreshed on another Mac, and your side still showed this machine's cached number. The fact that someone had just used the account was invisible.

Now, for the same account and the same window, when both sides have data the later sample wins. That is what carrying the sample time was for. Without it the comparison was impossible and "local always" was the only available answer.

When either side has no time — an owner on an older build, or an endpoint that returns no timestamp — it stays local-first. "Cannot compare" is not "the cloud is newer", and presenting a possibly days-old cloud value as current is harder to explain than always-local.

Kimi, GLM and MiniMax carry a sample time too

All three were writing sampledAt: nil. It is the same omission that produced the future timestamp fixed last release, just on rows nobody happened to be watching.

None of their endpoints return a server timestamp, so the honest answer is the moment the response was decoded, taken once per response and shared by the rows and the enclosing UsageData, so the two cannot disagree by milliseconds. Without this, "take the newer" would have been a Codex-only behaviour and the feature would have looked broken everywhere else.

Provenance is no longer a new word — the icon already said it

The row used to read [icon] label · time · percent, where the label could be Local, Cloud, Latest or Mix. This release removes the last two.

They existed only to make provenance visible, and each one cost a reader a lookup before they could act on it. Mix was worse than that: nothing else in the app had ever taught the reader what a "mix" is, so the word had no intuition to rest on.

The icons already carry the same information: no icon means this Mac, a cloud means the team cloud, an eye means watched by address. So the row keeps the icon and the number, and the explanation moves into the tooltip, where it can be a sentence instead of a label: where the number came from, when it was measured, and — when both sources are involved — that the newer of the two is showing. That last one is not optional; without it a number that disagrees with the local one reads as a bug.

The local account still shows its source name (OAuth / Codex CLI). That one is not invented vocabulary — it is which credential this Mac is actually running on, and dropping it would lose information rather than remove noise.

Remaining percent without expanding

The account header now shows the selected cycle's remaining percent, so reading the number no longer requires expanding the account.

Which cycle is a choice is stored per account. Wanting 5h for one account and Weekly for another is the normal case, and a single global value means switching back and forth. With one cycle there is no switcher at all — a button with one option does nothing when pressed, and in the common weekly-only case the header already shows that number. With two cycles, a compact switcher appears under the header.

Before you have chosen, the row falls back to whichever cycle is currently running, so it self-corrects across a reset instead of pointing at a window that just closed.

This selection does not drive the menu bar ring. Those are two separate intents: the ring is a persistent display preference; this is "what did I want to read in this row just now".

The provider header stops spending a slot on a count

The header used to show 2/5 for Codex (how many accounts are still fresh) and 2/2 for everyone else (how many models still have quota). Two different meanings in one position, and neither is a question you ask daily — both are for troubleshooting.

It now shows when that section was last refreshed, plus the remaining percent of the current account. Current rather than aggregate, because that is the number the menu bar ring is already showing: the same app answering differently in two places is worse than showing one less number. With no local Codex account there is no "current" to speak of, so the percentage is omitted rather than picked arbitrarily — an arbitrary pick reads as a deliberate choice. A section that is entirely stale still shows its time: the newest timestamp would itself be old, and each row carries its own time regardless.

A horizontal reference line when hovering a cycle in deficit

The chart already draws a dashed diagonal for "where you would be at this rate if you burned evenly", but reading the overspend off it means eyeballing along a slope. A horizontal line from the hovered point to the right edge turns that into a distance you can read: where it meets the diagonal is how much of the window you have already spent early.

It only appears in deficit. On pace, the same line would suggest something is wrong when nothing is — and onTrack with a non-zero delta is still on pace, so the decision is made on the stage rather than the delta sign. Those two signs are opposite conventions in this codebase (UsagePace.deltaPercent is actual-minus-expected, currentIntervalPaceDeltaPercent is expected-minus-actual), so reading stage is what keeps this from coming out backwards.

The line uses the chart tint rather than a fixed green, which keeps it in the forecast family instead of reading as data, and it stays consistent with the curve when the curve itself has gone orange or red.

Smaller corrections

  • Team member names are often email addresses and often longer than the space they get, so they now truncate at the tail instead of shrinking — a shrunken name makes that row read as if that member were special
  • "Whole team" in the menu bar strip becomes "Whole". The old wording did not fit and pushed the selected member's name out of the visible area entirely, and that name is the one thing the control exists to show. It sits directly after the Team button, which is why the reader knows it means the team. The settings pickers keep the longer wording: they have the width, and they follow a subject-prefixed pattern there ("All devices", "All accounts")
  • Groups with no account name — GLM and MiniMax when there is no local account name — no longer draw an account header at all. It rendered as Local | Unknown account, where the source word was already in the section title and "Unknown account" only restated the absence. Zero information, one line of noise. Kimi drops out the same way when it has no account name
  • MiniMax's subscription date drops the leading zeros: 2027-04-05 → 27/4/05. A renewal date answers "which day", and the right edge does not have the width to spare

Known boundaries

  • The merge needs a sample time on both sides. An owner who has not updated keeps local-first, which is a deliberate conservative choice
  • The horizontal reference line decides on the stage rather than the delta sign, because the two delta conventions in this codebase have opposite signs
  • A local + cloud merged row renders only the winner, but it remembers that both sides took part (mergedSourceKinds), which is what the tooltip reports
View this release on GitHub ↗

v1.35.0 — a placeholder with a choice, a trend in the ring, honest timestamps

v1.35.0 — a placeholder with a choice, a trend in the ring, honest timestamps

macOS build 68. This release also changes how the app is installed.

Installing is now one double-click

The DMG asked for three steps: download, mount, drag. The installer package asks for one, and it does two things a drag-and-drop cannot:

  • The running app is closed before its bundle is replaced, and comes back afterwards. A menu bar app that silently stops being a menu bar app until you happen to notice is a worse outcome than one extra click
  • The installer deletes itself when it is done, so it does not sit in your Downloads folder

Both installer scripts run as root and were verified twice on a real machine before this shipped.

One thing worth knowing: the package is not Developer ID signed and not notarized (the app inside it is ad-hoc runtime signed). Double-clicking it after a browser download makes macOS warn more than it did for the DMG. If it says the developer cannot be verified, Control-click the package and choose Open, or use Privacy & Security → "Open Anyway". After that it behaves normally.

The release still carries a DMG for anyone who prefers it. The two are functionally identical.

The menu bar placeholder is a choice now

The old setting was a single toggle — "show the provider count when everything is hidden". That is a boolean expressing two shapes, so it had no room for a third. It is now a picker:

Option What the menu bar shows
OpenAI connectivity (default) A muted ring with a reachability mark in the centre (check, cross, or probing)
Provider count A muted ring with the number of enabled providers
Brand mark The app icon's letterform

Connectivity is the default because a placeholder has no quota to show, and "can I launch ChatGPT right now" is the one thing still true at that moment that you actually care about. Without it, the placeholder is a decorative dead ring.

The mark is deliberately unlike "Codex is down": that state stacks a red ring, a red pill and a pulse. This keeps the ring muted and completely static, and carries state only in the centre mark, so it stays readable at twenty-odd pixels. Reachable uses the neutral label colour rather than green — this app's semantic colours are amber and red, and green would be a third thing to learn.

The tooltip gains a line explaining the mark, because a glyph that small has no other explanation.

An account with only a weekly window finally gets a shape

An account with a Weekly window and no 5h window had a single pace number spanning days, which says nothing about shape. Now:

  • The ring and the inner channel both resolve to Weekly for such accounts. You may have set the ring to 5h, but there is no 5h to show — Weekly is the quota you were actually asking about
  • The inner channel draws the window's own history as a curve instead of the bidirectional pace sector
  • The curve needs at least two samples to exist (one point draws no line, and would leave a lone dot where the sector should be, which reads as a rendering bug) and is capped at 24 points (the menu bar is ~22px wide; more is just aliasing)
  • The horizontal axis is each sample's position on the window's own timeline, not its index in the array. Ordering by index makes two samples look like a full week's swing, which is a lie

The inner channel is its own setting. If you point it at a different window, the sector keeps that channel — a 10% weekly overspend is the signal most worth seeing, and the trend should not take it away.

Watched accounts were showing the wrong time

The account row's trailing time is the last sample time. It read Watch · 10/11 23:59 for a snapshot taken five minutes earlier.

The cause was two hops: CodexUsageDataMapper dropped the real fetch time (sampledAt: nil), and the publish projection then fell back to endTime — which for Codex is resetsAt, the end of the window, days ahead. The data was current and the time was wrong, which reads worse than having no time at all: you would conclude something was broken.

Fixed at four layers:

  • Source — every Codex row carries the real fetch time
  • Publish — never falls back to a window boundary
  • Read — a sample time later than "now" is not a time, it is a bug report, so it falls back to the publish time
  • Server — future sample times are clamped at publish

The last two exist so that an owner who never updates still produces correct rows for everyone else. Verified in production: the other Mac is still on the old build and keeps sending the future timestamp, the server rewrites it on arrival, and every reader gets the right time. resetsAt is a separate field and is untouched — the "35% over · 10/11 23:59" reset line still shows correctly.

Smaller corrections

  • A long account name truncates instead of shrinking. A shrunken name makes the rows in the same block inconsistent, which from a distance reads as "this one account is special"
  • When width runs out, the plan name yields before the account name does — the plan is decoration, the account is identity
  • Timestamps are atomic and are never cut into a half 10/5 11:1
  • Cross-day timestamps carry the date (9/28 11:50), in the admin console as well as the menu. A bare 11:50 does not say which day's snapshot it was
View this release on GitHub ↗

v1.34.0 — Watching an account by email: no team, no key

v1.34.0 — Watching an account by email: no team, no key

macOS build 64.

Watching an account: knowing the address is enough

The last two releases both made "see the quota if you know the account" harder than it needs to be. v1.32.1 was a LAN peer-to-peer link requiring a typed IP, port and a hand-rolled key; v1.33.0 moved onto team cloud sync, which fixed the transport but still required both Macs to join the same team with an invite code.

This release removes that middle layer too: anyone running AI Quota Bar who knows an email address can read that account's quota.

  • No team, no invite code, no access key, no IP or port
  • The two Macs do not need to be on the same network — the transport goes through the app's own cloud service
  • The watching Mac does not need to sign in to Codex
  • The sharing Mac only has to turn on "Let other Macs watch my accounts" and tick the address

Still under Settings → Providers → Codex → 关注账号.

Data flow

the Mac holding the sign-in                  the Mac that watches
┌────────────────────────────┐             ┌──────────────────────────────┐
│ local Codex credentials    │             │ enter one email under         │
│        ↓                   │             │ "Accounts I watch"            │
│ fetch quota                │  POST       │        ↓                     │
│ project only ticked accts  │ ──────────► │ GET /v1/watch/quota?email=…   │
│ percent / reset / plan     │  app's key  │ the address is the lookup key │
│        ↓                   │             │        ↓                     │
│ /v1/watch/publish          │             │ grouped by account in the menu│
│ replace-all: [] = unshare  │             └──────────────────────────────┘
└────────────────────────────┘

Codex credentials never leave the machine they are signed in on. Only remaining percentages, reset times and the plan name travel.

The boundary of this channel, stated up front

The read path is open to anyone who knows that email address. This is a direct consequence of "no team, no key", not an oversight: without either, the server has no secret it can check. The real authorization is the owner ticking that address one by one — ticking it means "a Mac that knows this address may see my quota", and unticking deletes it from the cloud immediately.

The server stores sha256 of the normalized address, so the table never holds a plaintext email. But be precise about what that buys: the hash is hygiene, not access control — anyone can compute the lookup key for an address they know.

Within that, the following still hold:

  1. Default deny. With the switch on, only individually ticked Codex accounts are published; unticked ones are not sent at all. Quota with no account name is not published either — it cannot be watched by email.
  2. Only percentages cross the wire. Remaining percentage, reset time and plan name. No token, no API key, and no absolute usage — absolute counts would expose how much this person actually burns.
  3. Revocation is immediate. Publishing is replace-all, so unticking deletes the row right away: no waiting for the next refresh, and no separate revoke endpoint that could half-apply.
  4. Stale data does not pose as current. A snapshot older than 3 hours is labelled "these numbers may be old" rather than rendered as a live value. The row is kept — an account you explicitly asked to watch should not vanish.
  5. Reads are visible. The sharing side shows how many times each account has been read and when it was last read. Count and time only, no viewer identity; the owner's lever is "untick", not "blocklist this reader".
  6. No relaying someone else's account. The share list comes from the accounts signed in on this Mac, not from the menu's cloud-merged view — the latter contains teammates' reported accounts, and sourcing from it would make this Mac republish a colleague's quota as its own.
  7. The team system is untouched. /v1/watch/* is its own namespace and cannot read team data, device credentials or the admin console. No team-isolation guarantee was relaxed.
  8. Rate limited. Reads are limited per source IP (120/hour), publishes per publisher (120/hour).
  9. One publisher per address. A second publisher is refused (409); otherwise the original publisher's later "unshare" could not delete the row and the account would stay published forever.

Endpoint authorization uses the app's built-in key (called SYNC_TOKEN server-side), which every install already carries — zero configuration. It is not a secret: it sits in the binary. What it does is keep casual curl out, give reads an attribution bucket and give rate limiting something to key on. The real authorization is the per-account tick in point 1.

Turning the switch on now fetches by itself

What gets published comes from the quota this Mac just fetched. Right after the switch is turned on there is usually no data yet, so the switch now runs a usage refresh first and publishes what it gets — no manual refresh needed.

The misleading empty state is gone too. When the list is empty the app no longer says "No Codex account is signed in on this Mac" (it may simply not have fetched yet, which calls for the opposite action), and instead says the quota has not been fetched, with a retry button.

Three problems fixed

  • A failed fetch deleted the cloud snapshot. On a cycle where the Codex fetch produced nothing, the app published an empty set — and since publishing is replace-all, one network blip silently unshared the account. It now publishes nothing when there is no local data.
  • Expired numbers were shown as current. The status check tested "is there a snapshot" before "is it stale", and the cached snapshot is deliberately kept on a stale response (the row must not disappear), so numbers 2.5 hours old rendered as "Cloud · N entries". Staleness is now checked first.
  • The shareable account list was stale. It was @State, refreshed only when the panel opened and when accounts were added or removed — a quota refresh did not update it. It is now a computed property that always follows the latest fetch.

Differences from the previous two releases

v1.32.1 (build 62): LAN peer-to-peer. The owner machine ran its own plain-HTTP server; the watcher typed an IP, a port and a hand-rolled key; same Wi-Fi only. Removed: the LAN endpoint, seven CodexWatch* types, and the "keep listening so an account can be watched" branch the mobile dashboard grew.

v1.33.0 (build 63): moved onto team cloud sync. This release reverts that version's changes to the team channel — upload filtering and the startup read condition are undone, and team sync behaves exactly as it did before any of this work.

Upgrading from v1.32.1: previously ticked accounts migrate automatically (read from codexWatch.isServing / codexWatch.grants.v1) and do not need re-ticking.

View this release on GitHub ↗

v1.33.0 — Watching a Codex account now takes one email, over the cloud

v1.33.0 — Watching a Codex account now takes one email, over the cloud

macOS build 63.

Watching a Codex account: one email address

The previous release (v1.32.1) built this feature as a LAN peer-to-peer link: the Mac holding the sign-in started its own plain-HTTP server, and the watching Mac had to type an IP address, a port and an access key. That made it heavy to use, it only worked on the same Wi-Fi, and it sent you to Mobile Dashboard to find the key.

This release replaces it. The transport now uses the team cloud channel the app already had; the backend is unchanged and authorization still uses the existing device credentials.

The watching side now takes a single email address — no address, no port, no access key, and no Codex sign-in on the watching Mac.

  • Turn on "Let other Macs watch my accounts" and tick the accounts to share; their quota syncs to your team's cloud alongside Share account quota.
  • On the watching Mac, enter that email under "Accounts I watch" and press Add. The account appears in the menu as its own group row, with a status line reading "Team cloud · N entries · updated …".
  • Codex credentials still never leave the machine they are signed in on; only remaining percentages, reset times and the plan name travel.

The complete flow now

  1. Put both Macs in the same team under Settings → Team — create one, or join with the invite code. This is one-off; the member passphrase is kept in this Mac's Keychain and is not asked again.
  2. On the Mac signed in to the Codex account: turn on "Let other Macs watch my accounts", tick the account, and make sure Share account quota in the Team section is on.
  3. On the Mac that wants to watch it: enter the email under "Accounts I watch" and press Add.

The team is the authorization boundary of this channel, not a usability compromise: the cloud isolates data per team, so without joining one there is nothing to read.

The sharing side shares nothing by default

With the switch on, only individually ticked Codex accounts are uploaded; unticked ones are not sent at all. Quota with no account name is not uploaded either — it cannot be watched by email, so keeping it would only suggest that something is being shared.

With the switch off, no filtering happens at all: every Codex quota upload behaves exactly as "Share account quota" did before, so existing team users see no behavior change on upgrade. The switch also only governs Codex and does not affect the existing Kimi / GLM / MiniMax sharing behavior.

No relaying someone else's account

The share list is taken from the account store of accounts signed in on this Mac, not from the merged view the menu renders. The latter already includes what teammates reported, so sourcing it from there would make this Mac re-publish a colleague's quota as its own. This Mac only looks at that account; it has no standing to share it.

Three different "no data" states, reported separately

  • Not in a team → "Not in a team yet, so no one else's account can be read."
  • Channel fine, but no snapshot for this email in the cloud → "No data yet. Ask them to tick this address on their Mac, and check that Mac is running."
  • Data available → "Team cloud · N entries · updated …"

These need opposite actions from the user, so collapsing them into one "no data" message would send people looking for something to reinstall.

Also fixed

  • A Mac that only wants to watch other accounts, without sharing its own quota, now loads cloud data properly. Previously the startup path was tied to the Share account quota switch, so a watch-only machine showed empty rows until the next refresh.
  • The Add button is now disabled only when the field is empty. Invalid or already-watched input stays clickable and explains itself — a button that does not say why it is greyed out is worse than clicking and being told the format is wrong.

What was removed

v1.32.1's LAN channel is gone in full:

  • The /api/v1/watch/quota HTTP endpoint, and the "keep listening just so an account can be watched" branch the mobile dashboard service grew for it — the dashboard is back to only listening when you enable it.
  • CodexWatchPeerClient / CodexWatchPeerStore / CodexWatchResponseBox / CodexWatchPeerModel / CodexWatchProjector / CodexWatchPayload / CodexWatchGrantStore.
  • The per-peer tokens kept in the Keychain under codexWatch.peer.<uuid> (adding a watched account no longer produces any).

That channel's hand-rolled access key, address allowlist, redirect refusal and header-only token constraints went with it — there is no longer a credential to distribute across processes. Nothing in the team isolation guarantees (docs/team-isolation.md) was relaxed, and no new public endpoint or app-wide shared key was introduced.

Upgrading from v1.32.1: previously ticked accounts are migrated automatically (read from codexWatch.isServing / codexWatch.grants.v1) and do not need re-ticking.

Also

  • Mobile Dashboard is back to "listens only when you enable it": in v1.32.1, enabling account sharing alone also brought the LAN service up.
View this release on GitHub ↗

v1.32.1 — Watch a Codex account from another Mac, per-provider ring windows

v1.32.1 — Watch a Codex account from another Mac, per-provider ring windows

macOS build 62.

Watch another Mac's Codex account

  • A Codex account signed in on one Mac can now be observed from another Mac's menu bar — without signing it in there, and without needing to.
  • The owner machine does the fetching and serves the result to an approved peer on the LAN. Codex credentials never leave the owner machine, and nothing transits a server; only remaining percentages and reset times are sent.
  • Both halves live in Settings → Providers → Codex → 关注账号:
    • "Let other Macs watch my accounts": a master switch plus a per-account checkbox, off by default. An email that is not listed is never sent.
    • "Accounts I watch": add or remove a peer by LAN address, port and access key.
  • The access key is the one from Mobile Dashboard, so no second credential has to be handed out. Sharing quota does not require enabling the mobile dashboard: with only the sharing switch on, the LAN service still runs, but the dashboard and its account-name masking policy are unaffected.
  • Adding a peer handshakes first, so a wrong address, an unshared owner or a bad key fails immediately with a specific message instead of leaving a permanently unreachable row in the menu.
  • The access key has three constraints: it is only ever sent to loopback / private / link-local addresses (pasting a public address is rejected locally), every redirect is refused (otherwise a single 302 could walk the token to another host), and the token only ever travels in the Authorization header, never in a URL. It is kept in the Keychain and genuinely deleted with the peer.
  • Authorization can be tightened at any time and takes effect immediately; when the owner revokes access, the watching side drops its local cache instead of continuing to show quota it is no longer allowed to see.
  • "No new data" for two genuinely different situations — peer unreachable, and quota window elapsed — is reported separately rather than collapsed into one confusing error.

A watched account never disappears on its own

Cloud account data is hidden once it passes a staleness limit. That is right for cloud team data, and wrong for an account you deliberately asked to watch: the account would vanish from the menu whenever the other machine went quiet. Watched accounts take a different path — when the window elapses the display range is stretched to cover now, the row stays and is marked stale; when the peer goes away the last known values are kept rather than dropping to zero.

Separately, an account learned from another Mac is not re-shared by this machine: it only watched that account, and has no business publishing it onward.

Menu bar ring: per-provider data window

  • The outer arc and the reserve centre previously shared one global window choice, so you could not read Codex as Weekly and Kimi as its own current window at the same time.
  • Both windows are now configurable per provider; anything left unset follows the global default.
  • When a provider cannot source the requested window (MiniMax has no weekly window) it falls back to the current window instead of rendering something meaningless.
  • General gains a window picker per provider, and changes reach the menu bar immediately.

Admin console: quota detail is now traceable

  • The console showed only the latest snapshot per account, with no way to see how that number got there — a wrong figure and a genuine spike looked identical.
  • Raw sample history is now served for a single account behind an explicit row budget, and truncation is reported so a capped response can never render as a complete one.
  • Newest rows win the budget: taking the oldest slice made a busy account silently display days-old numbers with nothing to say so.
  • A model filter keeps a wide account from spending its whole budget on siblings the console filters away anyway.
  • Unrecognised query parameters are rejected rather than ignored — a typo used to answer 200 with the default range, which reads as a working filter.
  • The detail view renders history as model columns and loads history lazily.

Also

  • The local-usage pruning helper is renamed: pruneDelivered only dropped cloud-acknowledged rows, but the helper also removes local rows nothing can consume, so the old name made the second caller look like a misuse.
View this release on GitHub ↗

v1.32.0 — Step-away dimming, closed-lid dimming and an adaptive Clash panel

v1.32.0 — Step-away dimming, closed-lid dimming and an adaptive Clash panel

macOS build 61.

Step away: dim every screen with one click

  • The protection popover gains "Step away · dim all screens": every display is instantly covered by a black shade (screen-saver level windows) while the built-in backlight drops to zero and external monitors dim over DDC/CI; moving the mouse or pressing any key restores everything.
  • Press and hold the right-click on the status bar icon for 0.6 seconds to dim directly; a quick right-click still opens the popover.
  • Shade darkness is adjustable in Settings → General (60–100%, 90% translucent by default): below 100% is a pure overlay that never touches the backlight, and only full black performs real dimming.
  • Keyboard restore uses a listen-only event tap behind the Accessibility permission and silently falls back to mouse-only restore when it is missing.
  • On launch the app restores any dim state left behind by a previous crash or force-quit, so screens never stay dark.

Automatic built-in display dimming while working with the lid closed

  • While closed-lid protection is running (a task is keeping the Mac awake), the built-in display is held at minimum brightness without affecting external monitors; system auto-brightness recovery is continuously countered.
  • The original built-in brightness is persisted and restored on the next launch after an abnormal exit.

The Clash panel adapts to the screen height

  • The right-click Clash panel switches from a fixed baseline height to the anchored screen's available height: on small screens it first compresses the account section (internally scrollable), then yields space from the connection list, and finally collapses the accounts to their header row — the connection list is no longer squeezed to death or clipped at the tail.
  • Display changes and re-anchoring recompute the budget automatically.

Detection accuracy

  • Kimi desktop agents now use the conversation status ledger as the authoritative lifecycle source: tasks end immediately (no more lingering up to 120 seconds) and start lighting up instantly, with the heartbeat kept only as a crash fallback. The CLI silence window widens to 10 minutes so long tool calls no longer flicker out of the active set.
  • Kimi Work (daimon) desktop runtime sessions are now detected and labeled "Kimi Work" in the task list.
  • Codex hook turns gain a 10-minute TTL reaper: kill -9 or a lost Stop hook no longer leaves ghost tasks pinning the protection ring to "working" forever.
  • Claude Code model attribution follows the newest assistant message in the transcript: switching a session back to the official model stops counting it toward GLM/MiniMax immediately.
  • Switching or re-signing into a Codex account refreshes the current account in the left-click menu instantly instead of waiting up to 60 seconds.

Admin console

  • Account quota detail is rebuilt as an app-style provider → account → model-column hierarchy, one column per model side by side, with search, expand/collapse-all and reset view.
  • Expanding an account lazily loads up to 168 hours of quota history; timestamps use UTC+8 in the Chinese UI (UTC in English).
  • Charts now draw to the container's actual width with graceful narrow-column degradation; MiniMax and Z.ai provider logos are added.

Local data

  • The local usage database slims itself: cloud-confirmed events are dropped locally after 40 days (the charts read a ~35 day window), and the file is vacuumed when free pages exceed 64 MB; pending, rejected and local-only rows are never touched.
  • Detection polling logs only on active-task count transitions, and the MiniMax detector skips stale background folders, cutting daily log volume and disk scanning noticeably.

Requires macOS 14+; the DMG targets Apple Silicon and keeps the ad-hoc signature without Apple notarization.

View this release on GitHub ↗

v1.31.0 — Team view in the menu usage card

v1.31.0 — Team view in the menu usage card

macOS build 60.

The menu usage card can now show the whole team

  • The Codex usage card in the left-click menu gains a Local / Team switch (shown once a team connection exists): Local keeps reading the on-device scan, while Team reads server-aggregated current-month usage and the rolling last-24-hours timeline.
  • The team view ships a member picker: "Whole team" by default, or drill down to a single member to see their month heatmap and 24-hour distribution.
  • The team view does not require a local Codex sign-in — the card stays usable without one and only reads what members report via "Share local usage"; an empty-state hint appears until reports arrive.
  • Server aggregates move far slower than the menu's once-a-minute re-render, so each fetch is reused for five minutes; reopening the menu or switching member/team refreshes immediately. When a member leaves the team the selection falls back to "Whole team", and disconnecting the team returns to the local view.
  • While the first fetch is in flight the "Last 24h" caption already shows, matching the server's fixed 24-hour hourly window so the layout does not jump.

Requires macOS 14+; the DMG targets Apple Silicon and keeps the ad-hoc signature without Apple notarization.

View this release on GitHub ↗

v1.30.0 — Login-free Codex account switching

v1.30.0 — Login-free Codex account switching

macOS build 59.

Switch Codex accounts without logging in again

  • The right-click panel gains a "Codex accounts" section: the active account and every stashed backup are listed, and clicking Switch performs an atomic rename swap of ~/.codex/auth.json against the account pool (~/.codex/accounts/) — no login prompts, no QR codes.
  • "Log in a new account" stashes the current credentials into the pool, then reopens ChatGPT for you to sign in once; from then on the new account is one click away.
  • Guard rails: the app never quits ChatGPT/Codex on your behalf. While the desktop app or a codex CLI process is running, a switch request parks itself with a banner ("quit ChatGPT — the switch finishes automatically") and completes by itself within seconds of you quitting normally. The running state is re-checked at the last moment before any file moves.
  • The pool is manageable in place: rename a backup, delete one (moves to Trash), or import earlier hand-made auth *.json backups with one click. A same-account re-login replaces the older copy instead of duplicating it, and API-key-mode auth files are never touched automatically.
  • Each account row shows the last quota captured while that account was active: 5-hour and Weekly remaining percentages with their reset times (hover for full timestamps). Quota snapshots are recorded per account digest on every successful refresh.
  • Under the hood, auth-file parsing now lives in the shared local-usage core (same rules as the usage-account observation); the switch itself is two atomic renames that preserve 0600 permissions, and no token content is ever read or stored — only the account digest and email.

Requires macOS 14+; the DMG targets Apple Silicon and keeps the ad-hoc signature without Apple notarization.

View this release on GitHub ↗

v1.29.0 — Task waves end instantly, idle load drops sharply

v1.29.0 — Task waves end instantly, idle load drops sharply

macOS build 58.

GLM/ZCode tasks now end on time

  • A live experiment established that ZCode's turn_usage ledger only INSERTs a row once a turn reaches a terminal state (completed/error/cancelled, including user cancellation) — no running row is observable while a turn executes.
  • Detection is now hybrid: start and running still come from time_updated freshness, and every poll consumes newly-landed terminal rows by rowid cursor, dropping the session from the active set immediately.
  • End/cancel latency falls from up to ~120 seconds to a single poll cycle (2–3 s), and cancellation becomes visible for the first time. A turn continued right after a cancel is protected by the terminal-timestamp gate and is not mistakenly dropped.
  • If the table schema drifts, detection falls back to freshness-only behavior, identical to the previous release.

Idle-footprint fixes (two pinned cores)

  • Kimi CLI: wire transcripts are parsed incrementally, resuming from the last consumed newline. Idle machines previously re-read hundreds of historical transcripts in full on every poll, pinning a core at 100%.
  • MiniMax: each bg_* directory's mtime is prefetched with the listing call, and the probe files are read with a raw stat — FileManager previously raised and caught an ObjC exception per missing file, which across thousands of accumulated task folders saturated a core on its own.

Task wave: chase-queue layout (new default)

  • A second task-wave layout, chase queue, becomes the default: all waves line up single-file along the visible arc with constant-bright head beads, so every wave stays visible and countable at any task count.
  • The previous full-orbit behavior remains selectable in settings as evenly spaced.
  • Ambient animation (task wave, offline pulse) re-renders at 12 fps instead of 30 fps: every tick costs a full CPU-side bitmap render, so the slower cadence more than halves the idle render workload while slow motion stays visually smooth. With Reduce Motion the queue parks in a static pose on the visible arc.

Admin console (website)

  • /admin gains "Open dashboard": the platform admin can enter any team's dashboard without its password — including device-only teams that predate usage_teams.
  • Rotating the invite code of a device-only team now returns 404 instead of silently handing out a code that was never persisted.

Requires macOS 14+; the DMG targets Apple Silicon and keeps the ad-hoc signature without Apple notarization.

View this release on GitHub ↗

v1.28.1 — Reliable task detection, launch no longer freezes

v1.28.1 — Reliable task detection, launch no longer freezes

macOS build 57.

MiniMax task detection fixed

  • The working signal now comes from file activity inside the session directory instead of manifest.json, which measured minutes behind the live transcript — a working task no longer drops out of the count.
  • MiniMax Code desktop background tasks (~/.minimax/background-tasks) are now detected; they were invisible before, which was the main cause of "tasks running but the count stays zero".

Kimi task detection fixed

  • The old signals — session_index.jsonl (no longer maintained by the new runtime) and the process working-directory gate — stopped matching entirely under the new Kimi desktop runtime, so tasks flickered in and out.
  • Detection now reads each conversation's own updatedAt from the desktop agent's conversation-context-usage.json.
  • CLI wire transcripts still count: a turn must be open and the file must be freshly touched, so a killed process decays instead of protecting forever.

Launch no longer freezes behind the Keychain prompt

  • The one-time Keychain authorization after a re-signed install used to freeze the whole app at launch — no status bar, and app quit/launch events went unnoticed.
  • The credential-vault preload is now asynchronous: the status bar appears immediately with every feature working; once the dialog is answered, the configured cache is corrected and one refresh runs.

Requires macOS 14+; the DMG targets Apple Silicon and keeps the ad-hoc signature without Apple notarization.

View this release on GitHub ↗

v1.28.0 — AI task protection covers all four providers

v1.28.0 — AI task protection covers all four providers

macOS build 56.

AI task protection now includes GLM and MiniMax

  • "AI task protection" used to keep the Mac awake only during Codex and Kimi tasks; GLM and MiniMax coding tasks on this Mac now participate too. Task counts, the protection popover, and the mobile dashboard task list merge all four providers.
  • Three new read-only, passive local detectors (no message content is ever read):
    • GLM (ZCode): reads the session database shared by the ZCode desktop app and CLI; an interactive session counts as working while it keeps updating.
    • MiniMax CLI: reads each session's manifest.json update time under ~/.minimax/v2/sessions.
    • Claude Code bridge: GLM / MiniMax tasks run through Claude Code with a custom endpoint are attributed to the matching provider by the responding model name in the transcript (other models, such as the official Anthropic endpoint, never trigger protection).
  • Detection uses one signal — a session updated within the last ~2 minutes — plus the existing 5-minute turn-end grace period. All assertions are still released the moment tasks finish or the app exits.
  • Which providers participate still depends on local presence: GLM needs a Keychain plan credential or a ZCode installation; MiniMax needs a plan token or a MiniMax CLI session store.
  • The menu status line lists the locally detectable providers dynamically (for example, "Ready · waiting for Codex, GLM, Kimi, and MiniMax"); the system power assertion name is now the neutral "an AI task is working".

Note: MiniMax CLI detection is based on session manifest freshness, and its task-count edge cases are still being calibrated.

Requires macOS 14+; the DMG targets Apple Silicon and keeps the ad-hoc signature without Apple notarization.

View this release on GitHub ↗

v1.27.0 — Collapsible menu sections, one Keychain prompt

v1.27.0 — Collapsible menu sections, one Keychain prompt

macOS build 55.

Collapsible provider sections in the left-click menu

  • "Follow running apps" now collapses a provider's section in the left-click menu when its desktop app isn't open, instead of removing it — the header row stays, the data stays in place.
  • Click the header to expand or collapse manually. Automatic switching and manual toggles share one remembered state across menu reopens and app restarts.
  • The menu bar ring still appears and disappears with app presence, as in v1.26.0.

One Keychain authorization at launch

  • Launch no longer asks separately for the "AI Quota Bar" vault and the "kimi-desktop Safe Storage" item: after one interactive authorization, the Kimi desktop key is cached in our own vault, so only one item is ever accessed.
  • Background refreshes never touch another app's Keychain item (macOS partition checks prompt even in non-interactive mode). If Kimi rotates the key, refresh skips silently and the next "Allow Desktop Access" click in Settings re-authorizes.
  • Legacy Keychain items are only probed while no vault exists, and are deleted after a successful migration.

Kimi monthly bar is segmented by day

  • "Total usage" only carries an expiry date; the bar now derives a one-calendar-month display window from it — one divider per day, date labels on hover (thinned to six at most).
  • The pace tip, reserve/deficit hint, and the menu bar ring's surplus/deficit sectors are estimated from the same window.

Requires macOS 14+; the DMG targets Apple Silicon and keeps the ad-hoc signature without Apple notarization.

View this release on GitHub ↗

v1.26.0 — Menus follow your running apps

v1.26.0 — Menus follow your running apps

macOS build 54.

Follow running apps

  • Settings → Display adds a "Follow running apps" toggle (off by default). When enabled, the menu bar and left-click menu only show a provider while its desktop app is open: ChatGPT (Codex), ZCode (GLM), Kimi, or MiniMax Code.
  • Providers appear and disappear automatically as apps launch and quit — no manual toggling. When no provider app is running, the menu bar keeps a placeholder so the status item stays clickable.
  • If the left-click menu ends up fully hidden, a placeholder card explains why and offers a one-click switch back to showing all providers.

Scope

Display-layer only: background quota refresh, data sources, account handling, and the mobile dashboard are unchanged. Provider pause, ordering, and per-item hiding settings keep working as before.

Requires macOS 14+; the DMG targets Apple Silicon and keeps the ad-hoc signature without Apple notarization.

View this release on GitHub ↗

v1.25.0 — Reorderable menu providers, correct monthly cycle label

v1.25.0 — Reorderable menu providers, correct monthly cycle label

macOS build 53.

Left-click menu

  • Settings → Display gains a "Menu provider order" section: providers appear as pills in their current order and can be moved one position at a time with arrow buttons. The default order is Codex, Kimi, GLM, MiniMax; moving back to the default clears the custom setting automatically. Preferences stored by older versions load unchanged.
  • The order only affects display; quota collection, history, alerts and sync are untouched.

Monthly cycle label

  • Kimi's monthly "Total usage" row labeled its history bars "Week cycles"; it now reads "Monthly cycles". The mobile dashboard snapshot carries an explicit cycle kind ("short" / "weekly" / "monthly"), and older snapshots without the field fall back to the previous inference.

Scope

Display-only changes. Kimi, Codex, MiniMax and GLM data sources, refresh gating and account handling are unchanged.

Requires macOS 14+. The Apple Silicon DMG retains ad-hoc signing and is not Apple-notarized.

View this release on GitHub ↗

v1.24.0 — Automatically detect Kimi desktop and web sessions

v1.24.0 — Automatically detect Kimi desktop and web sessions

macOS build 52.

Kimi quota sources

  • Automatic mode discovers valid desktop sessions, saved web sessions, CLI credentials and readable browser sessions without manual source selection. A saved API key retains top priority.
  • Supports the newer desktop client's encrypted session store through read-only access; does not change client files or take over refresh-token rotation.
  • Settings show the most recently successful source. Manual selection, browser import and authorization controls are grouped under advanced options.
  • Displays Code short-window and seven-day quotas, plus membership totals when available, preserving separate reset times and account attribution.
  • Background browser discovery avoids Keychain prompts and caches results. Multiple different accounts require selection; network errors after selecting a valid desktop or web session do not switch accounts.

Availability and validation

Automatic desktop access was verified on a real device. Browser import and overseas accounts have not been verified on-device. Availability depends on session validity and local access permissions. If a desktop session expires, restore a valid login in the official client; this app does not refresh its credentials.

Requires macOS 14+. The Apple Silicon DMG retains ad-hoc signing and is not Apple-notarized. MiniMax continues to use its API key directly and does not depend on its desktop client.

View this release on GitHub ↗

v1.23.1 — Calmer menu expiry text, website team preview aligned

v1.23.1 — Calmer menu expiry text, website team preview aligned

macOS build 50.

Left-click menu

  • The MiniMax subscription line drops the repeated "expires" wording and shows just the date.
  • When the provider-header subscription subtitle runs long, it now shrinks by at most 10% instead of being truncated.

Website

  • The homepage "who used how much" preview card reproduces the released /team dashboard: the same page heading with range tools, unified 32 px controls with 6 px corners, and the released member-bar and hour-chart layout.
  • Removed legacy .activity-* rules from the retired hero demo: their class names collide with the real /team frontend and squashed the embedded hour chart to about 20 px tall.
  • The preview card lives in a column whose width varies continuously with the viewport, so its narrow layout now keys off the card's own width via container queries, with viewport media queries as a fallback for engines without support.

Scope

This release does not change monitoring, upload, retention or signing behavior; the performance and upload fixes from v1.23.0 are unaffected. The on-device performance audit is recorded in docs/performance-audit-2026-09-20.md.

View this release on GitHub ↗

v1.23.0 — Lower background load, visible storage and delivery status

v1.23.0 — Lower background load, visible storage and delivery status

macOS build 49.

Keep monitoring with less repeated work

  • Fix appearance callbacks repeatedly rebuilding every menu-bar animation frame. Rebuild only when visible content, appearance, size or display scale changes; retain 30 fps playback and existing monitoring intervals.
  • Process changed local-usage sessions and their related branches incrementally, persist lightweight indexes, and reuse unchanged statistics. Load only the latest 35 days into the interface while retaining older records on disk.
  • On one 10-core, 16 GiB Mac, average process CPU fell from 95.39% over 61 seconds to 9.07% over a ten-minute run. Here 100% means one CPU core. Sampled physical memory was approximately 121–177 MiB versus 346 MiB after restarting the old version. Results depend on hardware, log volume and active tasks; these are not universal guarantees.

Inspect storage and delivery

  • Usage settings show database and journal bytes, original Codex log size, quota retry-queue bytes, and acknowledged, pending, rejected and local-only record counts.
  • Show the latest upload acknowledgement, oldest pending record and retry time. A read-only cloud check compares record counts and all token totals for the current member, device and a fixed time range.
  • Fix oversized quota snapshots causing HTTP 413 and retry backlogs. New uploads and queued retries omit full utilization histories that the team endpoint does not consume. Failed retries retain their files, and legacy bindings remain isolated.
  • Unbound history is not a cloud backup and is not automatically uploaded. The existing quota queue limit remains; prolonged outages can still evict its oldest snapshots. Local token events use a separate durable outbox.

Retention boundaries

This release reduces in-memory history without deleting or lossily aggregating stored events. Recent detail, older daily summaries and distant monthly summaries remain a future retention proposal. Unacknowledged events should retain their original records. Original logs managed by Codex are not cleaned by this app.

View this release on GitHub ↗

v1.22.0 — Clearer menu-bar icons and accurate GLM status

v1.22.0 — Clearer menu-bar icons and accurate GLM status

All changes since v1.21.0. macOS build 48.

States that remain readable at menu-bar size

  • The center now uses two opposing fans: upper segments mean reserve, lower segments mean deficit, and a lit center dot means on pace. Unknown pace stays muted instead of appearing healthy.
  • Larger center shapes, a 104° top opening, and a slightly raised provider letter improve separation at small sizes. C / K / M / G stay visible at all times.
  • The outer arc continues to show remaining quota for the selected window. Yellow means low quota. Failed or unavailable data and Codex connectivity outages use a red dash, retaining the last known quota arc when available.
  • Staged and continuous pace modes remain available. Self-test demonstrates deficit, on pace, and reserve. Active-task waves follow the open ring without crossing the provider letter; hover retains pace and motion.
  • Final status-button images preserve warning colors instead of losing them to macOS template rendering. Self-test and offline pulses now have complete animation sequences. Icons regenerate for appearance and backing-scale changes and respect Reduce Motion.
  • Task waves now follow a physical circle and are masked through the provider opening instead of jumping between endpoints. Playback uses 30 fps and 16 gradient segments; the quota ring is no longer dimmed during tasks.
  • Every frame includes native 1× and 2× images, avoiding low-resolution upscaling across displays. Pixel-aligned placement and elapsed-time playback preserve sharpness and motion phase across refreshes.
  • Fix GLM ring and pace indicators ignoring the selected weekly window. When the 5h API omits reset metadata, chart actual samples from the last five hours and classify it as a short window without inventing reset times or pace.
  • Show available GLM personal reset allowances and their next expiry when using supported web credentials. This is read-only, excludes expired allowances, never redeems resets, and keeps optional lookup failures separate from quota.

One visual family across the app and website

  • The launch icon uses A and the same fan geometry. Warning panels, mobile dashboard home-screen icons, website branding, favicon, and live homepage demos are updated together.
  • Settings descriptions and both READMEs explain reserve, deficit, warning, and unavailable states.

Team dashboard refinements

  • A simpler header places sign-out in the panel toolbar and adds a homepage link in the footer.
  • Filter and toolbar controls share consistent heights, corner radii, and backgrounds. Hourly-chart spacing and track backgrounds are refined, with transparent idle cells to prevent visual merging.
View this release on GitHub ↗

v1.21.0 — Reset days on the usage calendar, exact time on hover

v1.21.0 — Quota resets on the usage calendar

All changes since v1.20.0. macOS build 47.

Quota resets at a glance

  • The usage month calendar marks Codex quota reset days with a dark green frame. Hovering shows the exact date and time, e.g. "09/12 16:09 reset" (local time zone, latest reset of that day).
  • Reset history comes from the public codex-resets.com feed, fetched quietly in the background and re-checked every 6 hours; when offline or unavailable the usage matrix is unaffected. The data is only displayed, nothing from your Mac is uploaded.

Hover callouts hug their cell

  • Hover callouts in the month calendar and the last-24-hours matrix now follow the hovered cell: above it by default, moved below when the top rows are full, and never covering the "Last 24h" or "5m / cell" captions. The website demo ships the same behavior.

Website: faithful team previews

  • The shared-account section now recreates the app's Settings → Team pane instead of using a screenshot.
  • The "who used how much" section embeds the same frontend chart component the /team dashboard ships, with sample data for member comparison, the calendar and 5-minute matrices.
View this release on GitHub ↗

v1.20.0 — Member usage charts and one-click team access

Member usage charts

  • Compare members with horizontal bars for tokens, records, cache hit rate and estimated cost.
  • Explore the monthly calendar and 288 five-minute cells covering the last 24 hours. Click a calendar day or select a historical date to inspect that day.
  • Combine member, device and account filters. Inspect input/output tokens, cached input, reasoning and pricing coverage. Incomplete pricing is not presented as a complete cost.
  • Available in both native Settings and the team dashboard. The app reuses the left-click menu renderer. Team charts use UTC and reported data only; missing reports do not imply zero consumption.
  • Time-bucket aggregation runs server-side and respects price intervals. Raw logs and conversations are never sent to the browser.

Create, join and open teams

  • Settings defaults to creating a team. Connected Macs can explicitly create or join another team; the previous connection stays until success, and assigned usage remains with its original team.
  • New managers have their management password saved in macOS Keychain. Keep a separate backup and share only the invite code with members. Failed joining retains recovery details and does not create duplicate teams on retry.
  • Every member can open a read-only team dashboard without typing credentials. Managers open management directly; existing managers enter their password once to save it.
  • Team summaries load automatically, with 7/30/90-day ranges, loading/error/empty states and an update timestamp. Cancelled requests cannot overwrite a newer team view.

Permissions and login

  • Handoff tickets expire after 120 seconds and are single-use; only their hashes are stored. Device credentials and management passwords never enter the URL, and the browser immediately removes the ticket fragment.
  • Sessions use Secure, HttpOnly, SameSite=Strict cookies for eight hours. Member sessions verify device authorization on every request and expire when devices leave, are revoked or rotate their token.
  • Management requires the password of the same team. Members cannot rotate invitations, revoke devices or delete quota history. Cross-team administration remains in independently authenticated /admin.
  • Requests from an old browser tab cannot act on another newly signed-in team.

Website and release maintenance

  • Corrected website previews against native calendar, usage, account, quota, cycle, settings and mobile views, including the overlapping hero heatmap cells. Verified 288 cells with no overlap at 390, 768 and 1440 pixels.
  • Added an actual team-dashboard chart preview, updated real SwiftUI screenshots and both READMEs. Included pending team/admin language switching, English feedback, bilingual changelog layout, navigation and historical release translations; moved language scripts out of inline markup to comply with production CSP.
  • Fixed update-manifest version normalization and verified the fallback when GitHub is unavailable.

Validation and deployment

71 backend tests and 386 Swift tests: no failures, five environment-related skips. Two additional real HTTP integration tests passed, including team creation/joining, role-specific handoff, replay rejection, revocation and five-minute charts. Native and browser views were inspected in both languages; mobile chart overflow was checked.

Self-hosters: apply migrations/0008_team_handoff.sql before deploying Pages. This adds a short-lived ticket table without modifying usage history.

macOS 14+, Apple Silicon, build 46. The public app is ad-hoc signed, not notarized.

Full comparison

View this release on GitHub ↗

v1.19.0 — Usage & Team, multi-team isolation

v1.19.0 — Usage & Team, multi-team isolation

This release covers all changes published after v1.18.0.

App: one place to manage one team

  • Usage and Sync are merged into "Usage & Team": create or join a team right in Settings, with one-time invite codes and management credentials stored for you.
  • Team members, devices, accounts and 7/30/90-day summaries in one place; on-device statistics keep working independently.
  • Local usage reporting and account quota sharing have separate switches, with sync status and logs shown in one place.
  • Members join from another Mac with a passphrase; signing out revokes that device's credential and pauses sharing.

Multi-team data & permission boundaries

  • Quota and usage both use team device credentials; the server enforces team scoping so other teams cannot read or delete your data.
  • Team isolation covers database primary keys, device identity, caches, upload queues and in-flight responses; switching teams drops stale request results, and old unowned queues are never uploaded.
  • Prevents taking over an existing member identity with only a device ID; credentialed requests no longer follow redirects automatically.
  • Team admins clean up their team's accounts and revoke member devices at /team; deletions go to the audit log. Regular device credentials cannot delete cloud account data.
  • Cross-team data, D1, legacy history and deletion audits are consolidated behind independently authenticated /admin, removed from regular settings.
  • New team quota and audit tables; old tables stay read-only, with no guessing of data ownership from emails or account names. Team quota history is retained for 90 days by server time.

Website, docs & quality

  • Bilingual README updates for team creation, joining, data boundaries, admin responsibilities and real settings screenshots; new deployment and isolation notes.
  • The website gains an English homepage, language switching and localized demo copy, with sitemap, version and team entry updates; menu bar quota rings, metric spacing, the account picker and stats typography refined.
  • Previous release notes synced to the website.
  • New regression coverage for cross-team read/write/delete, member identity, queues, request & session switching, native create/join/leave and the bilingual UI; fixes an animation lifecycle issue in the settings window test.

Verification

64 backend tests pass; the full Swift regression runs 384 tests with 0 failures and 5 environment-skipped; 2 real-HTTP integration tests pass. Live checks confirmed anonymous access is rejected and admin sign-in, team listing, legacy data and audit endpoints work.

Post-release live verification: fixed a missing version-string dependency in the update-check endpoint and verified version fallback when GitHub is unavailable. Server-side only — the DMG is unchanged.

Upgrade notes

Legacy shared sync credentials are no longer accepted. After upgrading, create or join a team in "Settings → Usage & Team" to keep sharing; existing team device credentials keep working. On-device statistics are unaffected, and old cloud records stay read-only in the admin console — they cannot be auto-assigned to a new team. Creating or joining enables both sharing types, which can be paused separately. Reporting includes account labels (possibly emails) and analytics metadata — no provider credentials or conversation content.

macOS 14+, Apple Silicon. The app is ad-hoc signed and not notarized by Apple; the first launch may require allowing it via the System Settings steps in the README.

English

Usage and Sync are now Usage & Team, with native team creation/joining, independent usage/quota sharing controls, team summaries and sync diagnostics. Server authorization, data keys, caches, queues and in-flight responses are scoped to the team. Owners manage only their team at /team; platform-wide data, legacy history, D1 and deletion audit live behind independent /admin authentication. Legacy shared-token sync stops working: upgrade and create or join a team. Local statistics remain available, old cloud records stay read-only, and team quota history is retained for 90 days.

Full comparison: https://github.com/techfanseric/ai-quota-bar/compare/v1.18.0...v1.19.0

View this release on GitHub ↗

AI Quota Bar v1.18.0 — Feedback, self-serve teams and a fuller usage view

AI Quota Bar v1.18.0 — Feedback, self-serve teams and a fuller usage view

Summarizes all changes since v1.17.1. macOS 14+, Apple Silicon, build 40.

Self-serve team joining

  • Members no longer need credentials issued one Mac at a time: the team lead creates a team on the website's /team page and shares the invite code; members enter the code and their name under "Settings → Codex local usage" to join.
  • Optional member passphrase: joining a second Mac with the same name requires the passphrase; enrolled devices can set or update it anytime.
  • Invite codes ignore case and whitespace; clear errors cover invalid codes, taken names, already-bound devices, full teams and rate limits.
  • Joining matches the original binding exactly: optionally attribute not-yet-assigned local history to this member; history and reporting behave the same.

In-app feedback

  • "Settings → About" gains a feedback form: nickname and contact are optional, and contact info is visible only to the developer — never shown publicly.
  • Submissions appear instantly on the website feedback wall (/feedback), with a direct link below the form; clear errors cover network failures, rate limits and content validation.

Codex subscription & accounts

  • The monthly calendar marks subscription renewal/expiry days: dashed border for auto-renew, solid for expiry without renewal; only web account metadata consistent with sign-in observations is accepted — never persisted, never uploaded, valid for 24 hours.
  • The left-click menu pins the current Codex account to the top of the account group and shows the account name directly; other accounts keep their existing order.

Usage view

  • Trends become "this month + last 24 hours": a full calendar-month matrix (future dates kept as empty slots) and the last 24 hours in 288 five-minute cells; the menu advances every minute while open and rolls over across hours and months automatically.
  • Estimated cost appears only when every usage record has a price, so unpriced data is never misread as zero cost.
  • The local usage card label is shortened to "Local"; settings summary numbers reflect the current month.

Provider icons

  • Left-click menu group names and the settings sidebar now use each provider's real brand mark (Codex, GLM, Kimi, MiniMax), with monochrome variants that follow menu appearance, replacing the generic symbols.

Website & server

  • New public feedback wall at /feedback with POST/GET /v1/feedback: same-origin web and app Bearer entry points, 5 requests per hour per IP, contact info visible only in the admin console.
  • New /team self-serve team page with its APIs; the admin console gains feedback management (full list, hide/restore/delete).
  • Homepage demo and style polish; sitemap includes the feedback page.

Release verification

  • App tests: 342 total, 3 environment-dependent tests skipped, 0 failures.
  • Local-usage core tests run with the app suite; 57 Cloudflare server tests pass; 35 mobile dashboard frontend tests pass (landscape font-size assertions aligned with v1.17.1's "large landscape keeps enlarging type" behavior).
  • Built with CodexBar revision b6e65a83dc471817b7ff7678e68e0204c9dd604f, pinned since v1.17.0.

Install & scope

  • Download the DMG and replace the app in "Applications". The public build remains ad-hoc signed; update notices link to the download and never install automatically.
  • Existing settings, local usage history and cloud bindings are preserved; devices not yet connected to team services are unaffected.
View this release on GitHub ↗

AI Quota Bar v1.17.1 — Trends, tasks and a lighter menu

AI Quota Bar v1.17.1 — Trends, tasks and a lighter menu

Summarizes all changes since v1.17.0. macOS 14+, Apple Silicon, build 39.

Usage & first run

  • Codex local usage is now a fixed last-30-days calendar matrix with today's hourly bars below; the Tokens / records / cache hit rate / estimated cost switch remains.
  • The left-click menu header is shortened to "Local usage"; it still measures only the signed-in account, and per-account history stays in Settings.
  • Hovering keeps the cycle-bar callout style and drops the vertical selection line; unpriced, zero-usage, calendar padding and future hours are each presented distinctly.
  • New devices without a provider see connection guidance and an optional sample preview; no automatic requests to unconfigured Codex. Samples are generated in-view only — never stored, reported or alerted.
  • A missing Codex log directory is treated as a normal empty state; real read failures still surface. When Codex isn't running, the card points you to the app while keeping history access.

Credentials & updates

  • Member reporting credentials move into a shared keychain dictionary, sharing the cache with provider, cloud sync and dashboard credentials; legacy credentials migrate transparently.
  • Missing legacy entries cache the failed lookup, and access-denied or corrupted data no longer triggers repeated background reads; the provider settings "Connection help" offers an explicit retry.
  • The left-click menu gains a low-key new-version hint between Settings and Quit; Settings mirrors update status with download and changelog links.
  • Discovered updates survive restarts; concurrent checks are deduplicated, failed checks can retry after 15 minutes, and a failed notification submission is not counted as "reminded".
  • Downloads still use the DMG for manual replacement — no automatic installation is claimed.

Icon & website

  • The app icon adopts a black-white-gray design of the menu bar quota ring and central pace indicator.
  • The website moves to a neutral gray-white palette while charts keep meaningful status colors.
  • Adds a full versioned changelog with search, plus a build script that syncs release notes from GitHub.
  • Adds self-test / idle / single-task / multiple-task icon demos implemented with the app's actual drawing rules — pausable and respectful of Reduce Motion.
  • The dashboard section now covers phones, tablets, car browsers and other computer browsers that can reach the Mac on the LAN.

Menu & sync diagnostics

  • Removes the provider quick-pause shortcut from the left-click menu; pausing and resuming stay in provider settings.
  • Cloud read errors no longer occupy menu space; Settings → Sync shows the current state with a link to local sync logs.
  • Logs keep the last 200 read/upload results, merging consecutive identical results with counts, and can be cleared; error codes are recorded — never tokens, request URLs or response bodies.
  • Cloud-sync-only new devices keep waiting for cloud history instead of being wrongly asked to add a local provider.

A fuller website & cross-device demo

  • The homepage shows the left- and right-click panels together, introducing task sleep protection, optional lid-closed work, route testing/switching and active connections.
  • Homepage and changelog share the app icon; the ring auto-cycles self-test, idle, single-task and multiple-task at true menu bar size, with manual exploration and pause.
  • The website's local usage demo now mirrors the 30-day matrix and today's hourly bars, dropping the old range switcher and pause entry.
  • The browser dashboard shows landscape light and portrait dark side by side, using the app's actual frontend with separate sample data.
  • Large landscapes stop enlarging fonts further, leaving more room for quota, task and connection data; the phone's adaptive layout remains.
  • The repository About section gains the website link and a feature summary.

Install & scope

  • Download the DMG and replace the app in "Applications". The public build remains ad-hoc signed; update notices link to the download and never install automatically.
  • Self-serve team creation, invited device joining and similar are still in planning; this release does not describe existing member attribution reporting as a full self-serve team system.
View this release on GitHub ↗

AI Quota Bar v1.17.0

AI Quota Bar v1.17.0

This release summarizes everything since v1.16.0: from shared-account quota only to on-device Codex consumption and per-account history, integrated across the menu, settings, cloud services, the admin console and the product website.

Codex local usage & history

  • Incrementally measures tokens, valid usage records, cache hit rate and API-equivalent estimated cost from this Mac's Codex sessions / archived_sessions logs; today by hour, and 7/30-day trends by day.
  • Shows input, output, cache reads/writes, model mix, parse issues and price coverage. SQLite persists the parse cache and pending upload queue so restarts resume instead of rescanning everything.
  • Deduplicates repeat snapshots, archived copies, continuations, fork inheritance and cumulative-value regressions; statistics pause when a parent session is missing rather than guessing usage.
  • Account information from the logs wins; when missing, consecutive local sign-in observations are used. History that can't be attributed stays as "Unknown account" instead of being charged to the current login.
  • Supports importing model prices with effective dates. Unknown prices stay "Unpriced" and never masquerade as zero cost; estimates are not a provider bill, and record counts are not all HTTP requests.

Menu & settings

  • The Codex provider is pinned to the top with local statistics and quota in one group; the left-click menu focuses on the current Codex account — the history account picker moves to Settings.
  • History trends remain, with Tokens / records / cache hit rate / cost switching and hover inspection; interactive controls use a unified neutral palette.
  • A unified "Display" pane combines menu visibility, mobile model selection and chart style into one account / model matrix, reducing duplicated settings.
  • The account master switch preserves individual model choices; mobile keeps its 1–2 item limit; hiding never affects collection, history, alerts or sync, and unavailable models keep their prior selections.

Member & device usage reporting

  • New independent member / device binding and usage reporting switches, with device credentials in the keychain; the admin console issues, rotates and revokes device credentials.
  • Statistics by member, device, account and time range; duplicate uploads of the same event count once, and cross-member attribution conflicts are quarantined.
  • Persistent catch-up uploads, network backoff and rebinding protection; whether history is attributed and back-filled is always the user's explicit choice.
  • Admin-configured device binding is still required. Self-serve team creation, one-time join codes and the invited-join flow are not released yet.

Cloud services & admin console

  • The default cloud endpoint moves to https://ai-quota-bar.pages.dev, keeping existing quota history and usage data; the cloud runs independently of the legacy deployment.
  • New /admin console with separate admin sign-in, anonymous install counts, daily/weekly/monthly actives, 7/30/90-day trends and app / macOS version breakdowns.
  • Anonymous operational analytics default to off — enable, stop and delete local stats in "Settings → General → Privacy"; no Codex emails, conversations or usage costs are collected.
  • Launching or opening the menu/Settings counts as active; background heartbeats don't count toward DAU. Anonymous installs are not real-user counts and exclude older versions and analytics-off devices.

Website & mobile dashboard fixes

  • The website gains a product introduction with interactive menu, usage, display-settings and connection-monitor recreations — all drawn in the frontend, no screenshots.
  • The mobile demo reuses the app's actual frontend components to show tasks, quota, routes, connections and protection with sample data; the demo never reads visitor app data.
  • Core features — cross-cycle trends, per-Mac consumption under shared accounts, task protection and connection diagnostics — are re-organized with better desktop and narrow layouts.
  • Fixes a rendering glitch on the mobile dashboard status bar when item counts changed, preventing quota/protection updates from blanking the route and connection areas below.

Download & upgrade

Download AIQuotaBar.dmg and drag the app into "Applications" to replace the old version. Requires macOS 14+; artifacts are Apple Silicon (arm64), version 1.17.0 (build 35). Existing settings and local history are preserved; the first read of a large Codex history may take a few minutes.

The public DMG stays ad-hoc signed and is not notarized by Apple. If macOS blocks the first launch, allow it via "System Settings → Privacy & Security". Reading existing keychain credentials after upgrading may require re-authorization.

Update checks still use the periodic check / system notification and the manual check in Settings, with downloads linking to the GitHub release; no auto-installer or in-menu update banner was added.

Release verification

  • App tests: 319 total, 3 environment-dependent tests skipped, 0 failures.
  • Local usage core tests: 28 total, 1 skipped without the local HTTP test service, 0 failures.
  • Cloudflare server: 36 pass; mobile dashboard frontend: 35 pass.
  • Built with CodexBar revision b6e65a83dc471817b7ff7678e68e0204c9dd604f, pinned in the source build instructions.
  • With a local Worker running, the Swift client's real-HTTP integration tests pass (identity binding, duplicate-upload dedup, summaries).
  • The mounted DMG verified version, build, binary consistency, deep signing and image integrity.
  • AIQuotaBar.dmg SHA-256: a5707c7c4affb4407cb84f7f82800eda3c55f6397e03cd28e9e0d0a45eeadafc.
View this release on GitHub ↗

AI Quota Bar v1.16.0

AI Quota Bar v1.16.0

This release lets you compose menu bar ring providers freely and completes the quota rings for MiniMax and GLM.

Free ring composition

  • Separates display mode from provider selection: always visible or work-aware, combined with multi-select across Codex, Kimi, MiniMax and GLM.
  • Supports combos such as MiniMax + GLM only or GLM alone, with a select-all action.
  • The previous "All" and "Work-aware" modes migrate to all providers; previous single-provider choices map to the equivalent combos.
  • Disabled providers are excluded from ring options and select-all, and become selectable again once re-enabled.
  • At least one enabled provider must remain; disabled providers no longer count incorrectly toward that check.

Quota & activity display

  • MiniMax and GLM show their latest remaining quota without task detection, updating with the existing quota refresh cycle.
  • GLM prefers the 5-hour quota to avoid picking a weekly window by mistake; an exhausted 0% still keeps an empty ring.
  • Under work-aware, selected MiniMax and GLM stay resident while Codex/Kimi show the active provider — falling back to the provider with the least remaining quota in the selection when idle.
  • When nothing can be shown, the status entry remains instead of reporting missing data as a normal quota.
  • Bilingual settings copy and project docs updated, plus a GLM ring capability research note.

Install

Requires macOS 14 or later; the prebuilt DMG targets Apple Silicon.

The public DMG is ad-hoc signed and not notarized by Apple. If Gatekeeper blocks the first launch, Control-click the app and choose Open, or allow it in System Settings → Privacy & Security. Do not disable Gatekeeper globally.

Release verification

  • Full Swift tests: 311 executed, 3 skipped, 0 failures.
  • Mobile dashboard frontend tests: 34 pass.
  • Cloudflare Worker tests: 10 pass.
  • Artifact version: 1.16.0 (build 33).
  • Architecture: Apple Silicon (arm64).
  • App deep signing and DMG integrity verified.
  • AIQuotaBar.dmg SHA-256: 765f7fbc56f0a4d85b8709b50e4e10bbadd8d1b89e2a49cc08d270bec08f2582.
View this release on GitHub ↗

AI Quota Bar v1.15.0

AI Quota Bar v1.15.0

This release adds GLM Coding Plan quota monitoring and reworks the provider credential editing experience.

GLM Coding Plan

  • GLM joins the provider settings, refreshed and displayed alongside Codex, Kimi and MiniMax.
  • Personal plans can paste a Coding Plan API Key directly; a quota/limit cURL copied from the usage page also works.
  • Supports the new 5-hour and weekly credit quotas, correctly separating cycles, remaining credits, plan tier and reset time.
  • Compatible with legacy tokens and MCP/Search quota responses; empty or invalid payloads are never reported as a successful connection.
  • Credentials stay in the macOS keychain, and historical web credentials can be edited and re-saved losslessly.

Provider credential settings

  • API key inputs are unified as native secure fields, with sizing, spacing and button hierarchy that better match macOS Settings.
  • Adds show/hide credentials, paste from clipboard and undo changes.
  • The save button enables only when content changes; saving restores the hidden state automatically.
  • Test progress and success/error feedback use a stable dedicated layout that no longer squeezes the action buttons.

Install

Requires macOS 14 or later; the prebuilt DMG targets Apple Silicon.

The public DMG is ad-hoc signed and not notarized by Apple. If Gatekeeper blocks the first launch, Control-click the app and choose Open, or allow it in System Settings → Privacy & Security. Do not disable Gatekeeper globally.

Release verification

  • Full Swift tests: 309 executed, 3 skipped, 0 failures.
  • Mobile dashboard frontend tests: 34 pass.
  • Cloudflare Worker tests: 10 pass.
  • Artifact version: 1.15.0 (build 32).
  • Architecture: Apple Silicon (arm64).
  • App deep signing and DMG integrity verified.
  • AIQuotaBar.dmg SHA-256: 82739d36b2b727587a52a3f694c643fba8203d8918d8957eed3b70b1774cd787.
View this release on GitHub ↗

AI Quota Bar v1.14.0

AI Quota Bar v1.14.0

This release adds per-provider pausing and unifies status presentation on the mobile dashboard. Pausing a provider never deletes credentials, and you can resume anytime.

Pause a single provider

  • Provider settings gain a "Use this provider" switch that pauses quota refresh, menu display, threshold alerts and task/lid-closed protection while keeping existing credentials.
  • The menu's provider header gains a quick pause button; the settings list marks paused items with a paused icon.
  • Paused state persists; data refreshes immediately when re-enabled.
  • When a pinned menu bar provider is paused, selection falls back to other enabled providers; if every configured provider is paused, clear recovery guidance appears.
  • Duplicate refresh requests issued mid-refresh run once more after the current request completes; results from providers paused mid-request are not written back to the UI.

Mobile dashboard status consolidation

  • The "Protection" block is renamed "Status", covering both task protection and quota state more accurately.
  • Quota errors become a red status item in the scrolling status bar instead of occupying space above the quota card.
  • Untitled status items hide their empty labels automatically, keeping the layout compact.

Install

Requires macOS 14 or later; the prebuilt DMG targets Apple Silicon.

The public DMG is ad-hoc signed and not notarized by Apple. If Gatekeeper blocks the first launch, Control-click the app and choose Open, or allow it in System Settings → Privacy & Security. Do not disable Gatekeeper globally.

Release verification

  • Full Swift tests: 301 executed, 3 skipped, 0 failures.
  • Mobile dashboard frontend tests: 34 pass.
  • Cloudflare Worker tests: 10 pass.
  • Artifact version: 1.14.0 (build 31).
  • Architecture: Apple Silicon (arm64).
  • App deep signing and DMG integrity verified.
  • AIQuotaBar.dmg SHA-256: 1a77199eb5fc672d68f54f17f2a1483a82b4727f1bc455112c35149a26496d3a.
View this release on GitHub ↗

AI Quota Bar v1.13.0

AI Quota Bar v1.13.0

This release summarizes everything since v1.12.2: cycle-end aligned refreshes, menu panel chart fixes, source-aware mobile dashboard adaptation, cloud read retries and closed-lid observability.

Cycle-end aligned refresh

  • 5-hour and weekly windows rarely end exactly on a 10-minute polling tick: now any cycle end triggers a refresh 15 seconds later, and regular polling restarts from that moment so new-cycle data lands immediately.
  • Applies only to local providers that can actually be fetched (e.g. Codex models drop out of scheduling when Codex isn't running); cloud-synced models don't trigger the mechanism.

Menu panel chart fixes

  • A transient cloud fetch failure keeps the previous samples — curves no longer vanish suddenly.
  • Curves without data no longer keep a fixed-height empty container; the panel shrinks to fit its content.
  • Hover previews of past cycles are tinted by that cycle's own usage; an orange warning from the current cycle's low margin no longer leaks into history previews.

Source-aware mobile dashboard

  • With only Kimi tasks running: the Kimi quota card moves to the front, the OpenAI-only connections block and proxy route block hide, and the title becomes "Kimi activity".
  • Codex-only or mixed tasks keep the full layout; with no tasks everything shows as usual.
  • The snapshot schema bumps to v4 with a new activeProviders field; older snapshots parse compatibly.

Cloud read auto-retry

  • The read path auto-retries transient network errors — timeouts, DNS failures, TLS breaks, lost connections — with exponential backoff, matching the write path; 4xx/5xx server errors are not self-healing and fail immediately.
  • Network error messages are unified into clear Chinese copy (e.g. "Request timed out: the link to the sync server is unreachable — check whether your proxy covers this domain").

Closed-lid observability

  • App and privileged helper gain OSLog logging: lease acquire/release/expire and re-asserting after external tools clear SleepDisabled are all searchable in Console (subsystem com.techfanseric.aiquotabar / .sleep-helper).
  • The helper's lease sweep interval drops from 15s to 5s, shortening the conflict window when external power-management tools rewrite disablesleep.

Install

Requires macOS 14 or later; the prebuilt DMG targets Apple Silicon.

The public DMG is ad-hoc signed and not notarized by Apple. If Gatekeeper blocks the first launch, Control-click the app and choose Open, or allow it in System Settings → Privacy & Security. Do not disable Gatekeeper globally.

Release verification

  • Full Swift tests: 301 executed, 3 skipped, 0 failures.
  • Mobile dashboard frontend tests: 33 pass.
  • Cloudflare Worker tests: 10 pass.
  • Artifact version: 1.13.0 (build 30).
  • Architecture: Apple Silicon (arm64).
  • App deep signing and DMG integrity verified.
  • AIQuotaBar.dmg SHA-256: 5ac5811523f756de59e774968f6d86b3941f5a9dd7ddc95c5e48dd19408f3c67.
View this release on GitHub ↗

AI Quota Bar v1.12.2

AI Quota Bar v1.12.2

This release summarizes everything since v1.11.0: unified chart time-axis rules, local sample caching, past-cycle previews, cloud status diagnostics and hardened sleep protection.

Chart time axes align to natural boundaries

  • Multi-day window dividers align to local midnight (previously cycle start + 24h), so you can tell which day is which at a glance.
  • Single-day windows (e.g. 5 hours) align to local whole hours (previously start + N hours).
  • Hover labels only cover segments fully inside the window: multi-day windows show MM/dd centered on that day's noon, single-day windows show 3PM-style labels at the half hour; partial edge segments are unlabeled.
  • The multi-day progress bar (weekly window) also labels only complete days, centered at noon.

Local sample cache

  • Consumption samples switch to merge-style writes retained for 90 days, so history curves survive reinstalls and restarts.
  • Per-model sample counts are capped with automatic trimming of the oldest data, keeping cache size in check.

Past-cycle hover preview

  • Hovering any past cycle in the cross-cycle utilization bars instantly switches the chart to that cycle's consumption curve; moving away restores the current cycle.

Cloud sync status diagnostics

  • Cloud sync errors are categorized with specific copy for network timeouts, DNS/connection failures and D1 quota exhaustion — no more generic "request timed out".
  • The sync settings page gains a D1 daily usage row (rows read / daily limit) with green/yellow/red thresholds at 50%/80%.
  • The Worker adds a read-only /v1/d1-usage endpoint (Cloudflare GraphQL analytics source, 5-minute edge cache), already deployed on the default cloud; self-hosters must configure CF_API_TOKEN and related switches per cloudflare/README.md — updating only the Mac app does not upgrade a self-hosted backend.

Hardened sleep protection

  • Power assertions and closed-lid leases survive 5 minutes after a task ends, so gaps between turns and activity-detection jitter no longer put the Mac to sleep mid-task.
  • Transient XPC drops no longer release the closed-lid lease immediately; natural lease expiry takes over instead.
  • If the kernel SleepDisabled flag is externally cleared (other tools, power switching), the assertion is re-applied automatically.
  • Heartbeat leases extend on negotiated timeout, and screen-saver user activity refreshes every 45 seconds.

Install

Requires macOS 14 or later; the prebuilt DMG targets Apple Silicon.

The public DMG is ad-hoc signed and not notarized by Apple. If Gatekeeper blocks the first launch, Control-click the app and choose Open, or allow it in System Settings → Privacy & Security. Do not disable Gatekeeper globally.

Release verification

  • Full Swift tests: 293 executed, 3 skipped, 0 failures.
  • Mobile dashboard frontend tests: 32 pass.
  • Cloudflare Worker tests: 10 pass.
  • Artifact version: 1.12.2 (build 29).
  • Architecture: Apple Silicon (arm64).
  • App deep signing and DMG integrity verified.
  • AIQuotaBar.dmg SHA-256: bf1bea33cbaf4d14099f528d4e893714e68d42fec47fd54773caf77235691723.
View this release on GitHub ↗

AI Quota Bar v1.11.0

AI Quota Bar v1.11.0

This release summarizes everything since v1.10.0: per-model chart styles, cloud sync retry fixes and D1 read/write optimizations.

Per-model chart styles

  • In usage settings, choose automatic, area chart or progress bar for each model.
  • Settings are stored locally and survive restarts; switching back to automatic clears that model's manual override.
  • The Mac menu and Mobile Dashboard share the same preference, with bilingual copy.

Cloud sync client

  • Fixes client errors being re-caught by generic exception handling, which caused pointless retries.
  • Clear D1 daily read/write quota errors stop the current retry round; transient network faults and ordinary server errors keep the existing backoff.
  • Account summaries fall back to the latest snapshot only on a 404; empty results, malformed payloads or ordinary server errors no longer trigger extra reads.
  • Normal refresh cadence, upload content and history retention are unchanged.

Matching backend read/write reductions

  • A trigger maintains a small latest-snapshot pointer table so latest-data queries stop scanning all history.
  • Global history queries cap per device first, then merge newest results; expiry cleanup uses existing device/time indexes.
  • Identical snapshot uploads no longer write duplicate samples or device heartbeats.
  • The default cloud has been migrated and released. Self-hosters must run the migration and configure switches per cloudflare/README.md — updating only the Mac app does not upgrade a self-hosted backend.

Verified in production: latest-data SQL read 176 rows, wrote 0 rows, and the API returned 24 samples. A local 200k-row benchmark dropped full-scan steps from 399,998 to 838 — about 99.79% fewer; that is a local query benchmark, not a production-bill reduction.

Existing account-level D1 daily usage is not reset by upgrading; when quota is exhausted, history queries may still be limited until reset or an allowance change. This release does not change your plan.

Install

Requires macOS 14 or later; the prebuilt DMG targets Apple Silicon.

The public DMG is ad-hoc signed and not notarized by Apple. If Gatekeeper blocks the first launch, Control-click the app and choose Open, or allow it in System Settings → Privacy & Security. Do not disable Gatekeeper globally.

Release verification

  • Full Swift tests: 280 executed, 3 skipped, 0 failures.
  • Mobile dashboard frontend tests: 32 pass.
  • Cloudflare Worker tests: 6 pass.
  • Artifact version: 1.11.0 (build 26).
  • Architecture: Apple Silicon (arm64).
  • App deep signing and DMG integrity verified.
  • AIQuotaBar.dmg SHA-256: dcaa6f3210746aea1c86e8d269282a41e9e54a5c227f5f72ba534e002859c026.
View this release on GitHub ↗

AI Quota Bar v1.10.0

AI Quota Bar v1.10.0

v1.10.0 focuses on the compact menu bar's sustained high CPU during task animations, while rounding out multi-provider support, quota forecasting and the Clash route filtering experience.

Compact menu bar performance & visuals

  • Active-task animations hold at 15 FPS.
  • Pre-rendered full multi-provider frames with a dedicated NSStatusBarButton.image path avoid per-frame redraws and whole-view snapshotting.
  • On the release machine, animation CPU dropped from ~54–62% to mostly 8–10%.
  • Retina offscreen drawing uses the correct 2× transform, restoring the rings' original point sizes, spacing and alignment.
  • Hover still shows provider initials, using static pre-rendered frames with no extra continuous drawing.
  • The diagonal pace divider now shares consistent colors with the vertical hour/day state.

Multi-provider & quota judgement

  • Work-aware selection can show separate Codex and Kimi rings at once.
  • Kimi's active tasks now count toward the menu bar work state and sleep protection.
  • Adds quota consumption forecasting and refines reserve/deficit cycle presentation.

Clash route filtering

  • Saved route filters show as a read-only summary by default, preventing accidental edits when opening the panel.
  • Tapping the summary enters edit mode with an explicit regex toggle and a clear finish action.
  • Bilingual copy, edit-state reset after closing the panel and matching tests are all in place.

Install

Requires macOS 14 or later; the prebuilt DMG targets Apple Silicon.

The public DMG is ad-hoc signed and not notarized by Apple. If Gatekeeper blocks the first launch, Control-click the app and choose Open, or allow it in System Settings → Privacy & Security. Do not disable Gatekeeper globally.

Release verification

  • Full tests: 266 executed, 3 skipped, 0 failures.
  • Artifact version: 1.10.0 (build 25).
  • Architecture: Apple Silicon (arm64).
  • DMG and app deep signing verified.
  • AIQuotaBar.dmg SHA-256: d2006caae4684b7ab4f4a688ffff692140038da7a128ea076cd1d9c58238c726.
View this release on GitHub ↗

AI Quota Bar v1.9.0

This release

v1.9.0 adds Kimi Code quota support and left-click menu customization, and fixes an undercount of active tasks with very large Codex logs; the Mobile Dashboard's compact layout and OLED idle screen are also refined.

Kimi Code quota

  • New Kimi Code provider showing 5-hour and weekly quotas, reset countdowns and quota curves.
  • Without an API key configured, it launches the real Kimi Code CLI in a local PTY and sends /status; it never uses kimi -p, so no model conversation is created and no context tokens are spent.
  • Refreshing short-lived OAuth credentials stays with the official CLI, avoiding direct manipulation of rotating access/refresh tokens.
  • First run trusts only an app-specific empty-state query directory — your project directories are never trusted.
  • An optional Kimi Code API key is still supported, stored securely in the macOS keychain.
  • Kimi can be pinned to the menu bar or shown on demand by automatic mode.

Quota menu & curves

  • Left-click menu content is controllable per account and per quota cycle.
  • Hiding only affects the left-click menu — background refresh, history, alerts, cloud sync and the pinned menu bar display keep working.
  • One click restores everything.
  • Improved curve fallback: when provider cycle metadata is incomplete, an existing 5-hour quota is kept and shown first.

Codex active-task reliability

  • Fixes an occasional undercount of active tasks when rollout logs grow very large.
  • The detector now restores log lifecycle state with bounded memory before reading the most recent 16 MB detail tail, so start events outside the tail window are still recognized.
  • Extremely long single lines are skipped without unbounded memory, with a new large-log regression test.

Mobile Dashboard

  • The mobile layout is tightened again, improving space usage on small landscape screens.
  • The Activity, Route and Connections blocks balance available height better in landscape.
  • Confirmed idle switches to a dedicated pure-black OLED screen saver with a single oversized IDLE glyph drifting slowly to reduce burn-in risk.
  • The idle screen, telemetry tracks, event stream and animations keep bounded lifecycles and respect Reduce Motion.

Docs & quality

  • Bilingual README, Kimi setup, privacy boundaries and v1.9.0 notes updated.
  • New tests for Kimi /status parsing, live credentials, left-click menu preferences, oversized Codex logs and 5-hour curve fallback.
  • Version bumped to v1.9.0 (build 24).

Verification

  • Swift: 241 tests, 3 environment-skipped, 0 failures.
  • Mobile Dashboard: 32/32 frontend tests pass.
  • Kimi live check: real CLI /status succeeded, 0 failures.
  • DMG verification, mounting, version, arm64 architecture and deep signing all pass.

Requirements

  • macOS 14 Sonoma or later.
  • The prebuilt DMG targets Apple Silicon.
  • The public build is ad-hoc signed and not notarized; the first launch may require Control-click → Open, or allowing it in System Settings → Privacy & Security.

SHA-256

9e6b025634b6a120adc9e2e59b028e1dc157ac0ca12d0a98d7c1ce6e5a8e3ef1

View this release on GitHub ↗

AI Quota Bar v1.8.1

AI Quota Bar v1.8.1

v1.8.1 fixes the Mobile Dashboard's full-screen IDLE marquee layout so the caption tracks truly cover the viewport.

Fixes

  • The 7 IDLE caption tracks move to a stable full-screen grid, spread evenly from top to bottom.
  • Fixes tracks collapsing and stacking at the top on Chrome versions that don't support calc() division.
  • Keeps varied font sizes, speeds, opacities, staggered animations and Reduce Motion support.
  • Adds a regression test so the incompatible track positioning can't return.

Install

Requires macOS 14 or later; the prebuilt DMG targets Apple Silicon.

The public DMG is ad-hoc signed and not notarized by Apple. If Gatekeeper blocks the first launch, Control-click the app and choose Open, or allow it in System Settings → Privacy & Security.

Release verification

  • Swift tests: 228 executed, 2 environment-skipped, 0 failures.
  • Mobile dashboard frontend tests: 32/32 pass.
  • Release build, hooks, privileged helper and DMG packaging succeeded.
  • DMG integrity and app deep signing verified.
  • Image version: 1.8.1 (23); architecture: arm64.
  • DMG SHA-256: 89b169adae848581d6ee4d807e706ea51eb73eede0f6999776ef03bf9a4830ba.
View this release on GitHub ↗

AI Quota Bar v1.8.0

AI Quota Bar v1.8.0

v1.8.0 adds a safe, installable Mobile Dashboard so a phone or another screen on the same LAN can continuously watch quotas, Codex tasks, routes, connections and sleep protection.

The full run state, from your phone

  • Enable the local dashboard in "Settings → Mobile Dashboard", open it via QR code and install the PWA.
  • Pick up to two quota models, reusing the Mac's quota curves, reset cycles and pace judgement.
  • Codex Activity, Route and Connections split roughly into thirds in both orientations and scale with resolution.
  • The page updates through a live event stream, showing current protection status, active connections, the latest route and the longest connection duration.
  • Automatic, light and dark appearances, OLED shift protection, and an experimental screen-wake media option that's off until you enable it.

A truer, more readable task telemetry barrage

  • One to five active tasks map exactly to one to five marquee lines, each line showing only that task's information.
  • Choose per field: title, status, phase, project, Git branch, source, model, provider, reasoning effort, sandbox, approval mode, tokens, subtasks, time, CLI version, tools, latest event and progress.
  • Each task's field order is randomized independently but stays stable across live updates, so captions never jump.
  • Five speeds at 48 / 40 / 34 / 29 / 25 px/s: newer tasks move faster, longer-running ones slower — distinct yet readable.
  • New captions join the current track queue without resetting scroll position; with a single task the font grows toward the Codex Activity block height.
  • Confirmed IDLE can switch to a pure black-and-white, multi-line full-screen caption screensaver with mixed sizes and speeds.

Local access & privacy

  • A stable local hostname plus explicit IP fallback, with reconnection after network address changes.
  • Optional five-minute manual pairing codes; pairing exchanges for a revocable, Origin-bound install credential that never stays in the saved URL.
  • Account names are hidden by default; task progress text leaves the Mac only after manual pairing is required and sharing is explicitly re-enabled.
  • The local server enforces strict Host/Origin/CORS/CSP validation, request size limits, pairing rate limits and a capped viewer count.
  • Provider credentials, the cloud sync token and the Mobile Dashboard token migrate to a versioned, recoverable keychain vault.

Long runs & compatibility

  • Codex lifecycle reading extends into per-task telemetry, covering representative record formats across 38 CLI version families.
  • Marquee field updates preserve animation phase; task count changes explicitly release old animations and DOM nodes.
  • Track queues, IDLE effects, SSE connections, timers, connection polling and reconnect state all use bounded lifecycles.
  • A browser stress test of 30 rounds of 5 → 1 → 5 held DOM nodes at 772 → 772 and listeners at 32 → 32, with exactly 5 track animations.
  • Respects the system Reduce Motion setting.

Install

Requires macOS 14 or later; the prebuilt DMG targets Apple Silicon.

The public DMG is ad-hoc signed and not notarized by Apple. If Gatekeeper blocks the first launch, Control-click the app and choose Open, or allow it in System Settings → Privacy & Security. Do not disable Gatekeeper globally.

Release verification

  • Swift tests: 228 executed, 2 environment-skipped, 0 failures.
  • Mobile dashboard frontend tests: 32/32 pass.
  • Release build, hooks, privileged helper and DMG packaging succeeded.
  • DMG integrity and app deep signing verified.
  • Image version: 1.8.0 (22); architecture: arm64.
  • DMG SHA-256: 5c839ede27ee124c15c2a5cbc5614e89a7a782c8683200600bf2828a95c81932
View this release on GitHub ↗

AI Quota Bar v1.7.0

AI Quota Bar v1.7.0

v1.7.0 is about the menu bar saying two things more clearly with less: whether a Codex task is still working, and whether OpenAI needs a route change right now — or just more patience.

Know Codex is still working without opening the menu

  • With no active tasks, the remaining-quota ring stays a solid 100% for easy reading.
  • When tasks run, the base ring drops to 60% so the energy waves flowing counterclockwise along the consumed portion stand out.
  • Each active task owns one wave, up to five; beyond five tasks the count stays at five to keep the 22pt icon uncluttered.
  • The wave head is most solid and its clockwise tail fades out — showing direction without turning into a distracting dot.
  • Waves never enter already-consumed quota segments.
  • With macOS Reduce Motion enabled, a static task hint is kept.

The left-click panel returns to true macOS native

  • The left-click panel uses the system NSMenu again, restoring native vibrancy, shadow, corner radius and menu behavior.
  • Available height is computed from the display currently hosting the menu bar icon, reducing misalignment with external displays, stacked screens or different menu bar sizes.
  • Left- and right-click entries dismiss each other before opening, so the two panels never overlap.

A tighter, easier-to-scan right-click control center

  • The right-click panel uses a consistent width, rounded corners and an arrow-free boundary — no more popover arrows whose position varies across system environments.
  • The auto-recovery switch stays pinned right; explanatory text stays left.
  • Routes are single-line, with H2, VL, TLS, SS badges distinguishing protocols at a glance.
  • Route and active-connection lists keep a minimum browsing height and no longer shrink to an invisible sliver with little data.
  • The last three switches show "from → to" on the left, with relative times (59m ago) within an hour and absolute times for older records.
  • With no switch history, the history area collapses instead of wasting space.

Judging OpenAI connection health faster

  • Download, upload and connection totals stay above the chart, so you can confirm data is actually flowing the moment the panel opens.
  • The connection-age legend lives in the chart's title row, using New → Old for the color shift from fresh to long-lived connections.
  • Active connection details start with protocol fields like TCP — no more placeholder "unknown process" text with no diagnostic value.
  • Monitoring still only reads active connections for openai.com and chatgpt.com; it never closes connections or edits Clash configuration.

Interface preview

AI Quota Bar v1.7.0 control center

Install

Requires macOS 14 or later; the prebuilt DMG targets Apple Silicon.

The public DMG is ad-hoc signed and not notarized by Apple. If Gatekeeper blocks the first launch, Control-click the app and choose Open, or allow it in System Settings → Privacy & Security. Do not disable Gatekeeper globally.

Release verification

  • Default suite: 119 executed, 117 passed, 2 environment-skipped, 0 failures.
  • 2 locally disabled tests for live Clash streaming and the panel were additionally enabled and passed.
  • Release build, hooks, privileged helper and DMG packaging succeeded.
  • DMG integrity and app deep signing verified.
  • Image version: 1.7.0 (21); architecture: arm64.
  • DMG SHA-256: ec8e8b9b4a4e6e8a1b05ddbd0c86d60576201a83502eba9d3f43128e7a5abac4
View this release on GitHub ↗

AI Quota Bar v1.6.0

AI Quota Bar v1.6.0

v1.6.0 makes long Codex and ChatGPT sessions worry-free: no more worrying about the Mac sleeping mid-task, and no more switching back to Clash to figure out slow or broken connections.

Let long Codex tasks finish in peace

  • AI Quota Bar shows how many Codex tasks are currently working.
  • Display, screen saver and idle sleep are held only while tasks run; everything restores automatically when they finish.
  • Multiple simultaneous tasks are counted separately — protection doesn't end early just because one finished first.
  • If you occasionally need to close the MacBook while a task continues, closed-lid mode can be enabled separately. First use needs administrator approval but no Apple Developer identity.
  • Closed-lid mode restores your original sleep settings when tasks complete, the app loses contact, battery runs low, the machine overheats, you switch it off — or the 12-hour cap is reached.

Diagnosing slow OpenAI faster

  • The right-click panel shows only openai.com and chatgpt.com active connections — other traffic never muddies the picture.
  • Total download, total upload and the current connection count are directly visible.
  • The last 60 minutes chart makes peaks, idle gaps and long-lived connections easy to spot.
  • Each connection is one colored block: green for new, shifting toward orange as it ages.
  • When data is available, the list shows domain, app, route, duration and speed.
  • All connection monitoring is read-only — no closing connections or editing Clash configuration.

Fewer steps when OpenAI drops

  • Routes and active connections now stack vertically — no tab switching.
  • The right-click panel is narrower while keeping enough browsing space for routes and connections.
  • The last three switches show briefly; with no history, no height is wasted.
  • Country filters like 🇯🇵, 日本, JP keep working, plus regex matching for precise route names.
  • Choose routes manually, or on a real outage let the app pick the lowest-latency route from your filter results.

Auto-recovery stays conservative: opening the panel, testing speed or toggling the switch never changes routes by itself. Recovery starts only when openai.com and chatgpt.com are unreachable together for two consecutive rounds, stops immediately on success, and sends a system notification.

Steadier daily use

  • Less waiting on the system keychain at launch, so the menu appears faster.
  • Upgrades now stop the old version before fully replacing the bundle and relaunching, reducing corruption from overwriting a running app.
  • Automated tests added for task detection, sleep protection, connection history, route history and layout.
  • The README is now fully bilingual, clearly explaining installation, privacy boundaries and closed-lid caveats.

Install

Requires macOS 14 or later; the prebuilt DMG targets Apple Silicon.

The public DMG is ad-hoc signed and not notarized by Apple. If Gatekeeper blocks the first launch, Control-click the app and choose Open, or allow it in System Settings → Privacy & Security. Do not disable Gatekeeper globally.

Interface preview

AI Quota Bar v1.6.0 control center

Release verification

  • 101 automated tests pass with zero failures.
  • 2 locally disabled tests for live Clash streaming and the panel were additionally enabled and passed.
  • Release build, hooks, privileged helper and DMG packaging succeeded.
  • DMG integrity and app deep signing verified.
  • Image version: 1.6.0 (20); architecture: arm64.
  • DMG SHA-256: b518eebfcb4c5bb9fb6ed2c052e96d6f3d647594f77c825ba0b87628a4116632
View this release on GitHub ↗

AI Quota Bar v1.5.0

Headline: OpenAI route auto-recovery

Right-click the menu bar icon to manage OpenAI routes directly — no more switching back to Clash:

  • Auto-discovers the local controllers of Clash Verge Rev, Mihomo or Clash.
  • Auto-locates the switchable strategy group carrying the openai.com / chatgpt.com rules.
  • Tests speeds automatically when the panel opens, sorting routes by latency.
  • Country alias filters like 🇯🇵, 日本, JP; regex mode supports anchors, start/end and OR matching.
  • Any route can be switched manually from the list.

OpenAI route panel

Conservative auto-recovery

Automatic switching triggers only when openai.com and chatgpt.com are unreachable together for two consecutive rounds. It then speed-tests the filtered results, switches starting from the lowest-latency candidate, and re-checks connectivity after each switch; success sends a macOS notification.

Opening the panel, clicking "Test again" or enabling auto-recovery never switches routes by itself, avoiding flapping. Auto-recovery requires a non-empty, valid filter.

Compatibility & safety

  • Supports the standard external controller APIs of Clash Verge Rev, Mihomo and Clash.
  • Only controllers bound to 127.0.0.1, localhost or ::1 are accepted; remote controller addresses are refused.
  • Requires macOS 14 or later.

Verification

  • Built from the clean commit matching the tag.
  • 57 automated tests pass.
  • DMG integrity, app code signing and binary consistency inside/outside the image verified.
  • DMG SHA-256: 251a0e3ccac9751e74e9d9e2aacbacaf6b8226f4ff8d1e56e65c5ba0d8970977
View this release on GitHub ↗

v1.4.10

What's new

  • Rebuilt the Codex compact ring: the outer ring shows weekly remaining while the split center distinguishes deficit, on pace and reserve.
  • Deviation is presented on a two-day scale: half the side fills at one day, full at two, with the active border appearing only beyond two days.
  • When both OpenAI domains are unreachable, a persistently pulsing prohibition icon appears and the outer ring shows exhausted.
  • A full status self-test runs at launch and on manual refresh; Settings can choose continuous percentage or staged pace levels.
  • When Codex lacks a usable short cycle, Weekly is automatically promoted to an area chart with day ticks.
  • README and menu bar status previews updated, with new rendering, curve-fallback and time-tick tests.

Verification: all 45 tests pass.

View this release on GitHub ↗

v1.4.9

Changes

  • Add the compact Codex quota ring: Weekly consumption on the outer ring and deficit/on-pace/reserve in the split center.
  • Add menu bar provider, appearance, and pace-detail preferences, including continuous percentage and staged levels.
  • Run a full icon self-test on launch and user-triggered refreshes while keeping background refreshes visually quiet.
  • Check chatgpt.com and openai.com reachability with jittered probes and morph the center into a no-entry state when both are unreachable.
  • Prevent inactive Codex Weekly 0% placeholders from replacing a still-active, higher trusted sample before its real reset.
  • Bump the app to version 1.4.9 (build 17).

Validation

  • 42 tests pass.
  • Release app signature and DMG checksum verified.
  • Installed and launched from /Applications/AIQuotaBar.app.
View this release on GitHub ↗

v1.4.8

Changes

  • Prioritize account identity in Codex multi-account headers while keeping Mix/Cloud on the same row.
  • Keep On pace labels neutral; only visible deficit labels are red.
  • Bump app version to 1.4.8 (16).
View this release on GitHub ↗

v1.4.7

Changes

  • Show weekly pace status instead of leaving the metadata row blank.
  • Show small on-track weekly deltas as reserve/deficit when they round to a visible percentage.
  • Color deficit pace labels red, including small on-track deficits.
  • Bump app version to 1.4.7 (15).
View this release on GitHub ↗

AI Quota Bar 1.4.6

Changes

  • Tightened quota history rows by removing duplicate left text under charts and moving cycle labels into the metadata row.
  • Fixed Codex weekly cycle hover text to show date ranges and current remaining quota correctly.
  • Normalized Codex utilization cycle hover ranges so 5h cycles use the actual 5h window instead of adjacent reset guesses.
  • Ignored idle Codex utilization samples so rolling unused 5h windows no longer appear as many fake 100% left cycles.

Verification

  • swift build
  • swift test --filter ModelUtilizationCycleMergerTests

SHA-256: 9d020a31d855f51b70a0211af8868974e3d06977ce871bc5006e0f0980acb086

View this release on GitHub ↗

AI Quota Bar 1.4.5

Fixes

  • Preserve MiniMax short-window chart points in Cloud sync by uploading percent-based quota samples as total=100, remaining=<percent>, and valueSuffix=%.
  • Keep restored Cloud chart data in percent mode when percent fields are available.
  • Change the default refresh interval to 600 seconds for new installs.

Release

  • App version: 1.4.5 (build 13)
  • DMG SHA-256: 0e8c90e4f57d0d9fc899cfa63497da581288f5e389c71536eb19cdd9c661a6c4

Verification

  • swift build
  • swift test --filter CloudSyncPayloadTests
  • swift test --filter ModelQuotaSampleStoreTests
  • node --check cloudflare/src/worker.js
  • git diff --check
  • make package

Note: Cloudflare API deploy attempts timed out from this machine, but this MiniMax fix is client-compatible with the already deployed Worker because it uses the existing quota_samples fields.

View this release on GitHub ↗

AI Quota Bar 1.4.4

Fixes

  • Persist short-window chart samples locally so curve points survive app restarts.
  • Restore chart curves from Cloud sync history when cloud data is enabled.
  • Add a Cloud history query mode for quota samples (history=1) while keeping the existing latest-per-model API behavior.
  • Update the data report to inspect Cloud refresh-history samples.

Release

  • App version: 1.4.4 (build 12)
  • DMG SHA-256: 9521b14b422e143015a113a6e91e0d830434f6fe1281a6bc34f8c76579c7cd90
  • Cloudflare Worker deployed: 76cfe627-4c71-4945-9718-c86cf557a232

Verification

  • swift build
  • swift test --filter ModelQuotaSampleStoreTests
  • node --check cloudflare/src/worker.js
  • git diff --check
  • make package

Note: full swift test still has the existing MenuBarBrandIconTests failure because MenuBarBrandIcon.image(for:) is currently a stub returning nil; this release did not change that behavior.

View this release on GitHub ↗

AI Quota Bar 1.4.3

Fixes

  • Restored the MiniMax provider setup flow for fresh installs.
  • Removed GLM from the active provider list so the app currently supports MiniMax and Codex only.
  • Added a Save Changes action for provider credentials.
  • Improved the menu bar text when only one provider is displayed by using the first line for the provider name.

Verification

  • swift build
  • make package
View this release on GitHub ↗

AI Quota Bar 1.4.2

Highlights

  • Added remote account management features and update checking improvements.
  • Added cloud data visibility limits and cloud data retention management.
  • Expanded language support across the new usage, sync, and cloud data settings.
  • Introduced new usage and sync settings tabs.

Verification

  • npm --prefix cloudflare run check
  • make package
  • swift test (fails in MenuBarBrandIconTests because test runtime cannot load provider SVG resources; other 19 tests passed)
View this release on GitHub ↗

AI Quota Bar 1.4.1

Highlights

  • Reworked Codex menu grouping to use provider -> account -> quota hierarchy.
  • Added built-in Cloud/Mix account presentation, cloud-only history fallback, and data management improvements.
  • Kept Codex Spark quotas independent from regular Codex 5h/Weekly history and cycles.
  • Skips Codex OAuth/CLI/Web refreshes unless the Codex app is running.
  • Hides empty Spark rows, collapses unused full-quota models by default, and shows Codex account health counts in the provider header.
  • Updates the Cloudflare Worker quota sample query to return the latest sample per model instead of only the most recent rows.

Verification

  • swift test --filter CodexUsageDataMapperTests
  • npm run check (cloudflare)
  • make package
View this release on GitHub ↗

AI Quota Bar 1.4.0

Highlights

  • Added optional Cloudflare Worker + D1 cloud backup for quota snapshots.
  • Added a Settings shortcut to view remote stored quota data as a local HTML report.
  • Kept provider credentials local in Keychain while storing only compact quota history remotely.
  • Added launch-at-login preferences and improved model grouping for exhausted/full quota rows.

Verification

  • swift build
  • node --check cloudflare/src/worker.js
  • make package
View this release on GitHub ↗

v1.3.2

Highlights

  • Added right-click menu bar cycling for used short-interval quota windows.
  • Improved automatic menu bar fallback so expired or exhausted selections rotate to the soonest reset among active, already-used models.
  • Skipped full, unused quota windows during active-model cycling.
  • Documented multi-account ChatGPT setup and the v1.3.1 ChatGPT short-window chart fix.

Verification

  • Built and packaged dist/AIQuotaBar.dmg with app version 1.3.2 build 7.
View this release on GitHub ↗

AI Quota Bar v1.3.1

What's fixed

  • Treat ChatGPT/Codex 5h quota as a short reset window so it uses the same quota chart as other sub-day limits.
  • Keep Weekly ChatGPT/Codex quota on the standard progress bar because its reset window is longer than one day.
  • Update the dropdown screenshot for the new ChatGPT/Codex quota display.
View this release on GitHub ↗

AI Quota Bar v1.3.0

What's new

  • Add ChatGPT/Codex GPT coding quota monitoring via the ChatGPT Codex usage endpoint.
  • Display Codex primary and secondary quota windows as 5h and Weekly, with percentage remaining and reset time/date.
  • Store provider credentials in one Keychain JSON item to reduce repeated permission prompts after migration.
  • Preserve percent suffixes in the menu bar for percentage-based providers such as ChatGPT and GLM.
  • Update the dropdown screenshot and README setup guidance.
View this release on GitHub ↗

AI Quota Bar v1.2.1

Fixes

  • Expand the menu bar dropdown height to use up to 90% of the visible screen height.
  • Keep the menu layout capped at the AppKit host view level so long provider/model lists remain scrollable without being artificially short.

Build

  • Version 1.2.1, build 4.
View this release on GitHub ↗

AI Quota Bar v1.2.0

What's Changed

  • Renamed the app and repository from MiniMax Usage Monitor to AI Quota Bar.
  • Repositioned the app around monitoring AI model coding plan quota across providers.
  • Updated the Swift package, app bundle metadata, update checker, docs, and build packaging for the new name.
  • Added migration for legacy preferences and Keychain credentials from the previous bundle/service identifiers.
  • Updated README usage docs, including the GLM/Z.ai flow for signing in, opening the quota page, using browser developer tools, and copying the quota/limit request as curl.
  • Removed old checked-in debug app and DMG artifacts from the repository.

Verification

  • Built the release product with make build.
  • Packaged and ad-hoc signed AIQuotaBar.app with make package.
  • Verified the app signature and Info.plist metadata.

Download

Download the attached DMG and drag AIQuotaBar.app into Applications.

View this release on GitHub ↗

MiniMax Usage Monitor v1.1.1

Bug Fixes

  • Area Chart Fill Path Fix: Fixed incorrect area fill rendering where the left vertical edge was not properly drawn from the first data point to the x-axis (y=0). The previous implementation relied on closeSubpath() which doesn't draw a line segment, resulting in missing area fill at the left edge of the chart.

Documentation

  • Added run experience guide for debugging and testing
View this release on GitHub ↗

v1.1.0

Highlights

  • Added automatic update checks (once every 24 hours).
  • Added update notifications when a newer release is available.
  • Fixed packaging issues that could cause the app to be reported as damaged.
  • Fixed app icon display in the bundled .app.
  • Added the Applications shortcut in the .dmg installer window.

Update Behavior

  • The app performs a background update check at most once per day.
  • If a new version is found, a macOS notification is shown.
  • Clicking the notification opens the release page.
  • Manual update checks in Settings are still supported.

Install

  • Download MiniMaxUsageMonitor.dmg from this release.
  • Drag MiniMaxUsageMonitor.app into Applications.
View this release on GitHub ↗

v1.0.0

v1.0.0 - First Release

MiniMax Usage Monitor is a macOS menu bar app for monitoring MiniMax API usage and quota.

Features

  • Menu bar widget displaying remaining quota
  • Detailed usage view with per-model breakdown
  • Quota trend charts for short-interval models
  • Configurable refresh interval
  • Warning notifications when quota runs low
  • Secure API key storage via Keychain

System Requirements

  • macOS 14+
  • MiniMax API Key

Installation

  1. Download MiniMaxUsageMonitor.dmg
  2. Open the DMG file
  3. Drag the app into Applications folder
  4. Launch the app from Applications
  5. Click the menu bar icon and select Settings to configure your API Key
View this release on GitHub ↗