pi-sub-quota

See your coding-plan quota — and your prepaid balance — right in the pi footer. Nothing to run: while a supported provider is active, the numbers are simply there:

↑1.2k ↓3.4k 38%/200k                    (zai-coding-cn 13% 58%) glm-5.3 • max
↑1.2k ↓3.4k 38%/200k                 (zai-coding-cn 96% 41m 58%) glm-5.3 • max
↑960 ↓1.2k 38%/200k                      (deepseek ¥18.67+¥1.00) deepseek-flash • max

• The two percentages are your plan's usage windows — the 5-hour one first, then the 7-day one. Green means plenty left, amber means watch out, red means nearly gone.
• The countdown (41m) appears on its own when a window needs attention: it is nearly used up (≥ 80%) or about to reset (within 30 minutes). Quiet windows stay quiet and show only their percentage.
• DeepSeek has no usage windows; you see your balance instead, split into what you topped up and what was granted (¥18.67+¥1.00), exact to the cent. Accounts with several currencies show one balance per currency.

Install

pi install npm:pi-sub-quota

Restart pi (or run /reload). There is nothing to configure: the extension reuses the credentials pi already stored, so the quota always belongs to the account pi is actually running with. When no supported provider is active — or there is no data yet — the footer stays exactly as pi built it.

Commands

Everything completes with Tab. inline is the default, and your choice is remembered across sessions; off restores the built-in footer.

The two display modes

inline (default) — the numbers live inside the provider prefix:

(zai-coding-cn 96% 41m 58%)
               │   │   └ 7d window
               │   └ countdown — only while it matters
               └ 5h window

When the terminal gets narrow, the balance folds into a single exact total (¥19.67), then the quota drops away, then the prefix — your stats are never sacrificed.

line — a dedicated quota line under the stats, always in full detail:

~/dev/project (main)
↑1.2k ↓3.4k 38%/200k                            (zai-coding-cn) glm-5.3 • max
(Max) 5h ▓▓▓▓▓▓▓▓▓▓ 96% 41m · 58%

Every window always shows its label, meter, percentage, and reset, and the line opens with your plan tier ((Max)). The meters shrink automatically on narrow terminals.

The /quota panel

/quota shows all configured providers at once — the active one first:

zai-coding-cn  coding plan · Max · active
  5h  ▓▓▓▓░░░░░░   62% left   reset 13:47 (3h17m)
  7d  ▓▓▓▓▓░░░░░   55% left   reset Fri 15:00 (3d4h)

deepseek  pay as you go · ¥19.67 + ¥1.00 granted · $2.50

kimi-coding  coding plan · Allegretto · booster ¥100.00/¥200.00
  5h  ▓▓▓▓▓▓▓░░░   30% left   reset 17:10 (6h40m)
  7d  ▓░░░░░░░░░   89% left   reset Mon 00:00 (3d13h)

/quota refresh · /quota mode line|inline|off · /quota help

The footer shows how much you have used; the panel shows what is left, as a meter and a percentage, with each reset as a concrete local time (13:47 today, Fri 15:00 this week, Sep 01 later) followed by a relative countdown (3h17m, 3d4h). The panel never waits on the network: whatever is cached renders the moment you press Enter — old numbers carry a ~, missing ones read fetching… — and a second panel replaces it once the refreshed numbers land.

Supported providers

Custom providers work without configuration too: detection falls back to the model's base URL host (bigmodel.cn, *.z.ai with a /coding path, *.kimi.com / *.kimi.ai with a /coding path, *.deepseek.com, *.moonshot.cn / *.moonshot.ai, *.minimax.io / *.minimaxi.com / *.minimax.cn, *.anthropic.com, *.chatgpt.com with a /codex path, *.openrouter.ai, *.opencode.ai with a /go/ path).

Good to know

• Numbers refresh while you work — on model switches, on every agent run, after each assistant turn, and throughout pi-subagents child runs (the parent session itself stays quiet while children burn the same account: a foreground child holds the turn open inside its subagent tool call, a background child leaves the parent idle, so child-progress events drive the refresh) — but at most once every five minutes per provider; a session with nothing pending, mode off, and non-TUI sessions (RPC, JSON) send no requests at all.
• A turn that finishes while the five-minute window is still closed is not lost: the skipped refresh is retried once, the moment the window (or the machine-wide one-minute request floor) opens — so a session that goes idle right after a big turn still picks that turn up. Nothing is polled: the wake-up exists only while activity is waiting to be covered.
• Quota requests ride pi's own network path: pi installs a proxy-aware fetch at startup, so a proxy configured in pi settings (httpProxy) or via HTTPS_PROXY applies to them too — relevant where Anthropic and OpenAI endpoints need one. Without a proxy those providers simply report fetch failures on the /quota panel and the footer stays as pi built it.
• Each provider endpoint is asked at most once per minute, machine-wide: every attempt (even a failed one) persists its timestamp in the shared cache before the request starts, so several pi sessions keep their distance too. A failed request earns a single immediate retry inside its cycle; when a fetch keeps failing, the /quota panel says so on the provider's own line — never as an interrupting notification.
• A failed refresh never breaks anything: the last numbers stay, marked ~, until the next success — and /quota says why a provider failed (quota unavailable (HTTP 401)) or when it will retry (retry in 42s).
• Several pi sessions on the same machine share one cache safely.
• A lapsed Kimi OAuth token is skipped with a note — it refreshes the next time pi itself calls Kimi.
• Credentials come only from pi's own store; no environment variables are read, and a key can never be sent to the wrong host.
• While quota is displayed, the footer is a re-implementation of pi's and drops a few rare markers (the virtual-model routing arrow, xp, (auto)). Everything else renders the same.

License

MIT © 2026 Ning. See LICENSE.