[Snowyyyyyy1]

The Usage page shows each allowance's readings and what is left, but users still have to estimate whether their current use will last until the next reset. This adds a forecast to each observed allowance window and shows its verdict on the Usage page and in the tray.

Rebased onto current main as one commit (44e1789a, on main cb05dc2d). Main has rewritten the allowance cards since this PR's base (packUsage, quotaFit, the past-cycle picker, colorblind dashes, Pace, quotaEta), so the verdict goes onto main's own code. The PR's earlier per-window plots and its ahead/behind chip are gone; main's Pace already says ahead or behind.

Behavior

• Forecast. QuotaHistories gives each line a forecast derived from the window's whole stored series (internal/provider/quota_forecast.go). It compares the current cycle's rate with an even burn. For windows of two days or more it also uses the well-covered past cycles, and the more conservative result wins. The state is ok, spent or none. A clear early burn warns before 8% of the cycle has passed. Cycle breaks use main's newCycle. The forecast is derived on read and is never persisted or synced.
• Usage card. quotaVerdict adds "Lasts to reset · 1.3×", "Runs out in …" or "Used up" on a line of its own after each window's reset line and Pace. A "history" marker shows when past cycles decided the answer. No verdict is shown for a missing forecast, state none, no readings, a brief account or a past cycle. Remote gateway-key cards and Sub2API key windows (gateway, provider, gui: a Sub2API key given 5h/1d/7d limits is weighed and rested by its own windows, like an account #1004) get the verdict too.
• Tray. One compact line per window under the rings. It wraps rather than being cut off, so the German line with both the multiplier and the history marker stays whole at 440px.
• Strings in zh, zh-TW, ja and de.
• Size budget. core_budget_test: app.js 1,250,000 → 1,254,000 bytes (1,252,431 with the verdict). The verdict is shared by the Usage cards and the tray, which stay in the core. The core total stays within main's 1,447,000 (test: TestCoreScriptBudget red on main since 1978eabe5 (core scripts 1,426,102 > 1,426,000) #2612).

The subsystem reference is updated in docs/subsystems/providers-accounts.md.

Validation

macOS, Go 1.27.2, Node with Playwright 1.62.1, at 44e1789a on origin/main cb05dc2d:

• go vet ./internal/provider ./internal/gui, GOOS=linux -tags nogui and GOOS=windows builds, gofmt -l internal/, node --check on app.js, git diff --check: clean.
• go test -race -count=1 -run 'Forecast|QuotaHist' ./internal/provider: ok.
• go test ./internal/gui (temporary HOME): all ok except TestProxyCheckSaysAProxyPassedBy (HTTP 503 from the sandboxed network) and TestMainWindowCorners (macOS title bar). Both fail the same way on origin/main on this machine.
• Every quota-*, usage-* and panel-* file plus i18n-keys, gui-zh-tw, gui-de, gui-ja and balance-curve (concurrency 2): 969 of 1000 pass. quota-forecast passes in full. Each failing file was then run alone here and on origin/main cb05dc2d. quota-count-wrap (1), quota-name-translated (1), usage-agents-hide (22), usage-checkin (2) and usage-fold (2) fail the same count on both. usage-curve-zoom-back's Chromium 1100px case is flaky on both: four runs each failed 5/3/3/4 on main and 1/4/5/5 here. None of these fixtures carries a forecast.

Earlier, at c51affd1 (on main 1eba266a):

• go test -timeout 45m ./internal/provider (temporary HOME): ok.
• quota-forecast.test.cjs covers Chromium and WebKit, en/zh/ja/de, light and dark, Usage at 1180 and 560px, and the tray at 440px (plus 260px for the wrap): 26/26, twice. Each check was confirmed by breaking the code on purpose and watching it fail. The breaks were: removing the past-cycle guard, showing a verdict for a missing forecast, removing the 10× cap, replacing the tray's wrap with an ellipsis (fails at 260px in German), and making quotaWindowLine ignore a pool window's own name (fails the "Every model" toggle case).

Not tried: the real app window or tray, Linux or Windows GUI, a live Sub2API key.

---

🎬 界面预览

_{CI 用这个 PR 的代码构建并真实运行 magpie（沙盒环境，配置了真实的 DeepSeek key），按改动自动操作、截图和录屏 · 44e1789 · 运行记录}

改动（按代码）：PR 在用量页的额度卡片和菜单栏托盘的额度卡片上新增预测行：撑得到重置（带余量倍数/历史标记）、预计用完或已用完，并补齐 zh、zh-TW、ja、de 文案与样式。

Note

沙盒里看不到：沙盒的实时额度只出现「撑得到重置」这类预测，无法展示「预计用完」「已用完」、无预测或 state 为 none 时不显示的分支，也无法展示托盘 440px 德语长句换行；这些需要构造额度历史或切换语言才能看到。

_{▶️ 点图打开录屏（可暂停、拖动）· 直接下载 mp4 · 红线是鼠标轨迹，红色圆环是点击}

用量页的额度预测

Codex 额度卡片：每个窗口的重置与配速后多出的预测行

「7 天」窗口的预测行末尾带「历史」标记

托盘额度页的预测行

托盘额度卡片圆环下方的按窗口预测行

[yetone]

Thanks for the careful write-up. I reviewed head 226a5e8 and ran it merged with current main (07326a3).

What holds up:

• go test ./internal/provider/ ./internal/gui/ passes.
• The forecast tests pass with -race -count=5.
• go vet passes, and so do the windows and linux nogui builds.
• On the GUI side, 443 of 448 cases pass across 60 files (quota-, usage-, panel-, balance-, codex-, upstream-, gui-ja, gui-de, i18n-keys, click-scroll), in Chromium and WebKit.

I also ran QuotaHistories over a copy of a real quota-history.json (Claude, Codex, Copilot, Devin, Grok, Kiro, ZCode, Cursor, Qoder, WorkBuddy, Command Code). Claude's 7-day window at 7% left with half the week to go comes out as "runs out in 5.3h", which is right.

What needs fixing before it can merge:

1. quota-read-at.test.cjs fails 4/4 (Chromium and WebKit, en and zh). It times out waiting for .quota-curve .qc-key.

   • That test guards #802: hovering a window's key lists its latest readings with when each was read.
   • The PR moves "Latest readings" into the plot's title, and "{name}: {n} left, read {when}" onto the dot, but leaves the test as it was. It passes 4/4 on main.
   • Please keep #802's behaviour reachable and update the test to where it now lives.
   • This file was already on your base (6892a28), so a run of the whole quota-* set would have caught it.

2. "Too few readings" is shown where it isn't true. A window with readings but no forecast falls through forecastVerdict(null) to "Too few readings". With the real history:

   • ZCode's "5 hours" has 15 readings and no forecast (no start/reset to place the cycle).
   • Cursor, Qoder, WorkBuddy and Command Code windows have readings but no forecast.
   • Codex "5 hours" has 37 readings, all at 100%, and gets state none.

   Please tell these apart: nothing used yet, no reset time to forecast against, or genuinely too few readings. Or show no verdict where there is nothing to say.

3. An early heavy burn is hidden.

   • Kiro's monthly credits read 67% left about 6% into the cycle, with an even-burn level of 94%.
   • That is under forecastMinElapsed (8%), so the state is none.
   • This is exactly the case a forecast should warn about: a third used in the first two days.
   • Please let a clear run-out through before 8% (for example, gate on an absolute minimum elapsed time, or on the even-burn ETA being well short of the reset). Add a test with that reading.

4. The "Every model" / "By family" toggle (from reading the code).

   • familyQuota's click builds a new quotaWindows box. Its plots are drawn while detached, at the 520px fallback.
   • The per-box quotaFit ResizeObserver is gone, so the labels aren't re-fitted.
   • Only a size change of #subscriptionUsage redraws them. Please redraw and re-fit after the toggle, and cover it in quota-forecast.test.cjs.

Interaction with #974 (Codex "Use credits" row): no textual conflict. Merged together, quota-curve.test.cjs "every window draws what was left over its own cycle" fails 4/4: the section after a Codex account now ends with quota-credits on. Whichever of the two merges second has to update that expectation. No conflict with #978.

[yetone]

Heads-up: #974 just merged (b37ac53) and adds a "Use credits" row (quota-credits) after a Codex account's Auto-use row on the Usage card. quota-curve.test.cjs line ~137 here rewrites the same assertion, so this will conflict on rebase. After rebasing, the second account's preceding siblings should be ["quota-windows", "quota-autoreset", "quota-credits on"] (three levels of previousElementSibling, as main now has). Please re-run quota-curve in Chromium and WebKit once you've rebased.

[yetone]

Thanks for the careful write-up. I reviewed head 226a5e8 and ran it merged with current main (07326a3).

What holds up:

• go test ./internal/provider/ ./internal/gui/ passes.
• The forecast tests pass with -race -count=5.
• go vet passes, and so do the windows and linux nogui builds.
• On the GUI side, 443 of 448 cases pass across 60 files (quota-, usage-, panel-, balance-, codex-, upstream-, gui-ja, gui-de, i18n-keys, click-scroll), in Chromium and WebKit.

I also ran QuotaHistories over a copy of a real quota-history.json (Claude, Codex, Copilot, Devin, Grok, Kiro, ZCode, Cursor, Qoder, WorkBuddy, Command Code). Claude's 7-day window at 7% left with half the week to go comes out as "runs out in 5.3h", which is right.

What needs fixing before it can merge:

1. quota-read-at.test.cjs fails 4/4 (Chromium and WebKit, en and zh). It times out waiting for .quota-curve .qc-key.

   • That test guards #802: hovering a window's key lists its latest readings with when each was read.
   • The PR moves "Latest readings" into the plot's title, and "{name}: {n} left, read {when}" onto the dot, but leaves the test as it was. It passes 4/4 on main.
   • Please keep #802's behaviour reachable and update the test to where it now lives.
   • This file was already on your base (6892a28), so a run of the whole quota-* set would have caught it.

2. "Too few readings" is shown where it isn't true. A window with readings but no forecast falls through forecastVerdict(null) to "Too few readings". With the real history:

   • ZCode's "5 hours" has 15 readings and no forecast (no start/reset to place the cycle).
   • Cursor, Qoder, WorkBuddy and Command Code windows have readings but no forecast.
   • Codex "5 hours" has 37 readings, all at 100%, and gets state none.

   Please tell these apart: nothing used yet, no reset time to forecast against, or genuinely too few readings. Or show no verdict where there is nothing to say.

3. An early heavy burn is hidden.

   • Kiro's monthly credits read 67% left about 6% into the cycle, with an even-burn level of 94%.
   • That is under forecastMinElapsed (8%), so the state is none.
   • This is exactly the case a forecast should warn about: a third used in the first two days.
   • Please let a clear run-out through before 8% (for example, gate on an absolute minimum elapsed time, or on the even-burn ETA being well short of the reset). Add a test with that reading.

4. The "Every model" / "By family" toggle (from reading the code).

   • familyQuota's click builds a new quotaWindows box. Its plots are drawn while detached, at the 520px fallback.
   • The per-box quotaFit ResizeObserver is gone, so the labels aren't re-fitted.
   • Only a size change of #subscriptionUsage redraws them. Please redraw and re-fit after the toggle, and cover it in quota-forecast.test.cjs.

Interaction with #974 (Codex "Use credits" row): no textual conflict. Merged together, quota-curve.test.cjs "every window draws what was left over its own cycle" fails 4/4: the section after a Codex account now ends with quota-credits on. Whichever of the two merges second has to update that expectation. No conflict with #978.

[yetone]

Thanks for the rework. All four points from the last review are addressed. I checked 4d92b3a:

• go test ./internal/provider ./internal/gui passes under a temp HOME.
• 56 GUI files pass in Chromium and WebKit (quota-, usage-, panel-, balance-, click-scroll, gui-ja, gui-de, i18n-keys).
• The Usage page at 1100 and 560px and the tray at 440px have no sideways scroll and no page errors in en, zh, ja or de.

Two things still block the merge:

1. It no longer merges with main. docs/subsystems/providers-accounts.md conflicts with b0a96cb (#1004, Sub2API key 5h/1d/7d windows), which landed today. Please rebase. #1004 also adds windows to Sub2API keys, so check that the forecast and the usage card behave on a key that has them.

2. German tray at 440px: the verdict is cut off. .pq-verdict > span ellipsizes (internal/gui/assets/app.css:1362), and the German margin label is long. The Claude Code 5-hour line reads "Reicht bis zum Reset · 1.3× Res…", and Gemini's reads "Reicht bis zum Reset · 1.4× Res…". title holds the full text, but the tray has room for a second line. When a line carries both the multiplier and the Verlauf marker (app.js:13749), the marker goes first.
   Repro:

   • Open ?mode=panel, lang de, viewport 440×800, Allowances tab, with the fixtures from quota-forecast.test.cjs.
   • Compare span.scrollWidth with span.clientWidth on each .pq-verdict.

   en, zh and ja fit. Fix: let the text wrap (for example white-space: normal with the b kept flex: none), or shorten the German string. Please add a check to the quota-forecast tray test that no .pq-verdict text overflows in every language.

[yetone]

Sorry for the slow review on this one.

At 5fe3470 the two blockers from the last review are addressed in the code: the German tray line wrap, and the Sub2API key windows from #1004. I read those changes but didn't run them at this head, because the branch no longer merges onto main (origin/main 5b14ebf). These four files conflict:

• docs/subsystems/providers-accounts.md
• internal/gui/tests/quota-curve.test.cjs
• internal/gui/tests/usage-accounts-fold.test.cjs
• internal/gui/assets/app.js, 4 hunks. Main has reworked the same usage-card code:
  • the card's layout pass: main runs fitUsageWall(); drawQuotaPlots; fitQuotaLabels, where this PR has packUsageNow(subscriptions);
  • the "now" line: main leaves out qc-now on thin sparks ([Bug] Linux GTK/WebKit 下托盘快捷面板（Tray Panel）的若干 UI 渲染与布局问题 #1601);
  • label fitting: main uses a per-target quotaFit ResizeObserver, where this PR has fitQuotaLabels(root);
  • quotaWindows(sub): main rewrote it, with new holdsLine, isKeyCard, quotaName and keyLimitLine. This PR changes the same function to quotaWindows(sub, brief).

Main has no forecast of its own, so the feature is still wanted. Could you rebase onto main? Please put the forecast verdict onto main's quotaWindows and key-card lines rather than bring back the older function, and keep #1601's thin-spark rule. Then re-run quota-forecast, quota-read-at, quota-curve, usage-accounts-fold and the panel tests in Chromium and WebKit, at 560px and the 440px tray, in every language. I'll review it again at the new head.