v1.3.1

VibeGauge 1.3.1

统计页：每日强度改版

• 按月日历：一行 7 天、周一开头，和日历对得上；格子里标日期，今天加描边，‹ › 翻看前几个月。颜色按全部活跃日统一分档，数据少的月份也不失真。
• 点一下看时段：切换成「近 7 天 × 24 小时」，看出一天里哪几个小时用得最多；再点切回月历，下次打开保持上次的选择。更早的记录已按天汇总，没有时段。
• 两种视图同一高度，切换时面板不跳动。

其他

• 修复记账代理自测偶发失败（并发写入顺序不固定）。

---

Stats: daily intensity, redesigned

• Month calendar: 7 days per row, Monday first, with day numbers, today outlined, and ‹ › to page back through months. Colors are graded against all active days, so sparse months don't distort.
• Tap for hours: switch to the last 7 days × 24 hours to see which hours you use most; tap again for the calendar. The choice is remembered. Older records are daily totals only.
• Both views share one height, so the panel doesn't jump when switching.

Other

• Fixed an intermittent accounting-proxy self-test failure (concurrent writes land in any order).

v1.3.0

VibeGauge 1.3.0

新功能

• 不走代理也能看官方额度
  • 火山方舟、阿里云百炼 Coding Plan：在「订阅」页点「安装并登录」，会打开终端，装好厂商官方命令行工具并登录（只需一次）。之后 5 小时 / 周 / 月额度都用官方数，替换本机估算。
  • 智谱 GLM、Z.ai、MiniMax、DeepSeek、OpenRouter、Moonshot：在「API」页登记 API key 就能查余额和额度。key 只存本机钥匙串，只发给该厂商自己的用量接口。
• Kimi Code 官方额度：开着 kimi web 时，读取它本机服务给的 5 小时 / 周 / 月额度。
• 会话上下文水位：每个活跃的 Claude Code / Codex 会话离自动压缩还有多远，今天压缩过几次。
• 待处理会话（可选，默认关闭）：打开后会往 Claude Code 写入只观察的 Hook，面板上能看到哪个会话请求了批准或在等你输入。只记录事件类型和时间，不会替你批准；关掉开关即删除。Codex 会话暂不支持。
• 订阅详情页显示月额度，以及按模型分开的额度池。

修复与改进

• 额度满了的提醒不再每隔几分钟重复弹出。
• 额度或时间数值异常时，App 不会再崩溃。
• 读取 Kimi 访问凭据前，会先确认端口由你自己账号的进程监听。

说明

• 仓库已改名为 MaxHaiCom/VibeGauge，旧地址会自动跳转。
• App 更新后第一次读取钥匙串里的 key 时，macOS 会询问一次，点「始终允许」即可。
• 一键安装官方命令行工具时，不会往你的其他 AI 工具里添加 skill。
• 腾讯云个人套餐没有官方查询接口，暂不支持。

---

New

• Official quota without the proxy
  • Volcengine Ark / Alibaba Model Studio Coding Plans: click "Install & sign in" on the Plans tab. Terminal opens, installs the provider's official CLI, and signs you in (once). The 5-hour, weekly, and monthly figures then come from the provider and replace the local estimate.
  • GLM, Z.ai, MiniMax, DeepSeek, OpenRouter, Moonshot: add an API key on the API tab to see its balance or quota. The key stays in your Keychain and goes only to that provider's own usage endpoint.
• Kimi Code quota (5-hour, weekly, monthly) from the local kimi web service while it runs.
• Session context levels: how close each active Claude Code / Codex session is to auto-compaction, and how many times it compacted today.
• Pending sessions (optional, off by default): when switched on, VibeGauge adds observe-only hooks to Claude Code and the panel shows which sessions asked for approval or are waiting for input. Only event types and times are recorded; it never answers for you. Switching it off removes the hooks. Codex sessions are not covered yet.
• Plan detail pages show the monthly window and per-model pools.

Fixes

• Full-quota alerts no longer repeat every few minutes.
• Unusual quota or time values no longer crash the app.
• Before reading Kimi's access token, VibeGauge checks that the port is held by a process of your own user account.

Notes

• The repository is now MaxHaiCom/VibeGauge; old links redirect.
• After an update, macOS asks once before VibeGauge reads a Keychain key; choose "Always Allow".
• The one-click CLI installers do not add skills to your other AI tools.
• Tencent Cloud personal plans have no official usage interface and are not supported.

v1.2.0

VibeGauge 1.2.0

新功能

• 订阅套餐按各家规则估算：plans.json 支持各窗口的重置方式（滚动 / 首次请求起算 / 每周一 / 订阅日）、模型抵扣系数、按模型分池。Coding Plan 这类一个订阅含多家模型的套餐，现在显示在「订阅」页并独占一整行（5H / 周 / 月），注明估算口径，真值以厂商控制台为准。API 页只放按量计费的 key。
• 本地模型监测：Ollama、LM Studio（含 llmster）、llama.cpp、MLX。能区分「在线 · 已加载哪些模型」「在线但空载」「进程在跑但接口没应答」。Ollama 原生流式接口的用量也能正确记账。
• 数据可信状态：每个额度窗口标注「官方回报 / 本机估算 / 可能过期 / 等待新回报」。过了重置点显示「—」，不再显示 0%。
• 预计打满提醒：官方回报的额度按当前节奏预计会在重置前用完、且持续 15 分钟时，每个周期提醒一次。
• 今日模型构成（统计页）和 Claude 请求异常（详情页，最终失败与自动重试分开计数）。
• 记账代理：自动走 macOS 系统代理或 HTTPS_PROXY，也可在 ~/.config/vibegauge/proxy.json 固定线路；代理端口可改（proxyPort）。

修复与改进

• 同一上游的多个 API key 按账户分卡，余额和套餐不再串户。
• 历史缓存升级时不会再丢失已删除日志的统计；缓存体积约缩小为三分之一。
• 冷启动扫描大幅加快，GB 级日志的内存占用有上限。
• 面板打开时数据跟随后台刷新；清理动作针对的就是面板上显示的那些进程。
• 同一行两张卡片等高，窄卡片不再溢出。

给开发者

• docs/PROTOCOL.md：所有数据文件、字段、设置和命令行参数的说明。
• --selftest 只跑离线测试；新增 --diagnose 输出本机诊断快照（已脱敏）。

---

New

• Plan estimates follow each provider's rules: plans.json now takes per-window reset kinds (rolling / from first request / Monday / subscription day), model weights, and per-model pools. Multi-model coding plans appear on the Plans tab as a full-width card (5H / W / M) that states how the estimate is made; the provider console is authoritative. The API tab keeps pay-as-you-go keys.
• Local models: Ollama, LM Studio (incl. llmster), llama.cpp, MLX — online with loaded models, online but idle, or not answering. Ollama's native streaming is now accounted correctly.
• Trust state on every quota window (reported / estimated / may be stale / awaiting update); "—" instead of 0% after a reset.
• Projected-to-run-out notification, once per cycle, for reported quotas.
• Today's model mix (Stats) and Claude request errors (detail page, failures and retries counted apart).
• The accounting proxy follows the macOS system proxy or HTTPS_PROXY, or a route pinned in proxy.json; its port is configurable (proxyPort).

Fixes

• Several keys on one upstream get separate cards; balances and plans no longer mix accounts.
• History of deleted logs survives cache upgrades; the cache is about a third of its old size.
• Much faster first scan; bounded memory on GB-scale logs.
• The open panel refreshes with background scans, and cleanup acts on what it shows.
• Cards in a row share a height; narrow cards no longer overflow.

For developers

• docs/PROTOCOL.md specifies every file, field, setting, and flag.
• --selftest is offline-only; --diagnose prints a masked local snapshot.

v1.1.2

An open-source review round focused on how VibeGauge behaves on other people's machines.

Safer process cleanup

• Background tools you started with npx and left running after closing the terminal are no longer treated as orphaned MCP servers. Official MCP servers are recognised by @modelcontextprotocol/server-*.
• Before sending any signal, VibeGauge re-checks listening ports and the auto-clean switch, and reports how many processes actually exited.

More accurate numbers

• API proxy: OpenAI Responses API usage is recorded; Gemini thinking tokens count as output; long streams (>8 MB) are no longer under-counted; interrupted requests are still logged.
• History cost no longer double-counts calls that went through the proxy and also appear in CLI logs.
• Codex "today": resumed sessions that span midnight now only count today's usage; sessions are found by modification time (resumed old sessions included), and per-project attribution works without Claude Code.
• Coding-plan estimates count only requests actually sent to the provider, and the card stays after midnight.

Forecasts & alerts

• Forecast samples use the provider's own capture time; last-cycle priors must be observed near the reset.
• Critical alerts are never held back by the cooldown; the pressure list ranks critical above warning.

Privacy & robustness

• --selftest output (which bug reports ask for) redacts secrets in command lines, including Authorization: Bearer … headers.
• IPv6 / DNS leak checks say "unknown" instead of a false ✓ when the probe itself failed.
• README now states exactly which auth-file fields are read, and that the optional proxy uses your API keys to query balances.
• Installing the proxy / connecting quota checks for python3 first and verifies the proxy is really up.
• Intel Macs: compressed-memory size is correct.

---

本版要点（面向「别人机器上」的真实情况做的一轮开源评审）

• 清理更安全：你用 npx 起、关掉终端后仍在跑的后台工具，不再被当成孤儿 MCP 清掉；发信号前复核端口与开关，按实际退出数上报。
• 数字更准：记账代理支持 Responses API、计入 Gemini 思考 token、长流式不再少记；历史成本不再重复计入代理调用；Codex 跨零点的会话只算今天；套餐估算只计真正发给厂商的请求，过零点卡片不消失。
• 提醒更可靠：危急级提醒不再被冷却挡住。
• 隐私与健壮性：自测输出对命令行里的密钥打码；泄漏体检在探测失败时显示「未知」而不是 ✓；README 如实说明读取了认证文件哪些字段；安装代理 / 连接额度前先检查 python3；Intel 压缩内存显示正确。

Full Changelog: v1.1.1...v1.1.2

v1.1.1

Fixes

• English UI: heatmap weekday labels no longer cut off; long forecast lines wrap instead of being truncated; platform names are never squeezed; the Gemini third-party pool is labelled 3P so the card fits.
• Token counts read cleaner (142k instead of 142.0k, 1830 万 instead of 1830.0 万) and no longer wrap in narrow columns.
• Network tab: wider name column, "Gateway" label, and the Gemini egress line is no longer shown in warning orange.
• New README screenshots (made-up demo data), regenerate with tools/screenshots.sh.

修复

• 英文界面：热力图星期标签不再被截断；预测行放不下时换行而不是截断；平台名不会被挤掉；Gemini 三方池英文标为 3P。
• Token 数字更清爽（上千万不带「.0」），窄列里不再折行。
• 网络页：名称列加宽，Gemini 出站信息不再显示成警告橙色。
• README 换上新截图（虚构演示数据，tools/screenshots.sh 可重新生成）。

Requires: macOS 14+ (Apple Silicon or Intel). First launch: System Settings → Privacy & Security → Open Anyway if macOS blocks it.

Full Changelog: v1.1.0...v1.1.1

v1.1.0

What's new

• Works out of the box for Claude Code & agy users — click Connect quota on the Claude / Gemini card once. VibeGauge reads the official 5h / weekly quota your CLI already hands to its status line; your status line looks the same, settings.json is backed up, and you can disconnect anytime under Mac → Settings.
• Sleep-aware weekly forecast — "% at reset" now follows your last 7 days: which hours you actually code and how much you used last cycle. An evening burst is no longer extrapolated through the night (e.g. 261% → 129% on the same data).
• Intel Macs supported — Universal binary (Apple Silicon + Intel).
• English / 简体中文 — follows your system; switch anytime at the panel's bottom-right.
• Update notice — once a day VibeGauge checks GitHub for a newer release and shows it in the panel. It never downloads or installs anything; turn it off in Settings.
• Gemini 3rd-party pool keeps showing its 5h window after it resets (0%) instead of disappearing.

Install / update

1. Download VibeGauge.zip below, unzip, drag VibeGauge.app into /Applications (replace the old one).
2. First launch: the app is ad-hoc signed (not notarized). If macOS blocks it, go to System Settings → Privacy & Security → Open Anyway.

Requires: macOS 14+ (Apple Silicon or Intel).

shasum -a 256 -c VibeGauge.zip.sha256   # optional: verify the download

---

更新内容

• Claude Code / agy 用户开箱即用：在 Claude / Gemini 卡片上点一次 「一键连接额度」，VibeGauge 就能读到 CLI 交给状态栏的官方 5h / 周额度。原状态栏显示不变，settings.json 自动备份，随时可在 系统 → 设置 断开还原。
• 懂作息的周额度预估：「到重置会用到多少」改按你自己近 7 天的作息（几点在用）和上个周期的用量推算，晚上猛用一阵不会再被外推成通宵都在烧（同一份数据 261% → 129%）。
• 支持 Intel Mac：Universal 通用包。
• 中文 / English：默认跟随系统，面板右下角随时切换。
• 新版本提示：每天查一次 GitHub 有没有新版本，在面板底部提示；只提示、从不自动下载安装，可在设置里关闭。
• Gemini 三方池的 5H 窗口重置后显示 0%，不再整个消失。

Full Changelog: v1.0.0...v1.1.0

v1.0.0

First public release of VibeGauge — a native macOS menu-bar monitor for AI coding (MCP orphan reaper, 5h/weekly quota tracking, token & prompt-cache stats).

Install

1. Download VibeGauge.zip below, unzip, drag VibeGauge.app into /Applications.
2. First launch: the app is ad-hoc signed (not notarized). If macOS blocks it, go to System Settings → Privacy & Security → Open Anyway.

Requires: Apple Silicon Mac (M1+), macOS 14+.

Verify the download (optional):

shasum -a 256 -c VibeGauge.zip.sha256

---

首个公开版本。下载 VibeGauge.zip 解压拖入「应用程序」；首次打开被拦截时到「系统设置 → 隐私与安全性 → 仍要打开」。需要 Apple Silicon（M1+）+ macOS 14+。

Full Changelog: https://github.com/MaxHaiCom/vibe-gauge/commits/v1.0.0