dsh-deepseek-peak-indicator

A DeepSeek Harness web plugin that shows a small green/red status dot in the composer bar, immediately left of the model-select menu, indicating whether the DeepSeek API is currently in official peak time (red) or off-peak (green).

Screenshots

Light	Dark

What it does

Dot	Meaning
🟢 green	DeepSeek API off-peak — standard rates, usually faster
🔴 red	DeepSeek API peak time — official peak window (rates doubled, higher load)

Hovering the dot shows the current tier, when the next transition happens (in your local time), and the official schedule. Clicking the dot opens a polished dashboard panel:

• a mode pill (● PEAK / ● OFF-PEAK) with a mode-tinted hairline accent,
• a countdown card — "Time remaining" with a live 32px tabular countdown (ticking every second; amber pulse in the final minute) and a caption for the next transition,
• a live token-cost card — V4.1 Flash rates (per 1M tokens, with a 元/$ switch) for cached input, input and output, off-peak vs peak side by side — the currently active rates are highlighted, so you see what tokens cost right now,
• a 24-hour day track — a seamless red/green gradient of peak/off-peak hours in your local time, with a "now" marker (ring + line) aligned to exact hour ticks and a summary line,
• stat tiles — "In this mode" (elapsed + since) and "Next" (relative + absolute time),
• a compact two-line schedule footnote.

The panel closes on outside click, Escape, or the close button; Tab cycles the panel's controls while open. It flips to the new mode automatically at a window boundary. Motion (panel entrance, pill pop, marker glide) respects prefers-reduced-motion, and the styles use the host theme's tokens so the panel matches light and dark themes.

The tier follows the official DeepSeek schedule — Models & Pricing — plus the peak/off-peak billing notices:

2026-08-23 — Weekends (Saturday and Sunday, Beijing Time): the peak/off-peak time divisions no longer apply — all calls are charged uniformly at the off-peak rate.

2026-09-10 — V4.1 Flash pricing takes effect (04:00 UTC). From 12:00 Beijing on 2026-09-14 (04:00 UTC), V4 Pro requests are routed to V4.1 Flash and billed at Flash rates; before that, Pro keeps its own unchanged pricing. The legacy model names deepseek-v4-flash and deepseek-v4-flash-vision-exp are retired and served as V4.1 Flash at Flash prices.

Cost card rates (per 1M tokens, peak = exactly 2× off-peak — both tables are official and never inter-converted):

V4.1 Flash	off-peak	peak
input, cache hit	元0.02 · $0.003	元0.04 · $0.006
input, cache miss	元1 · $0.15	元2 · $0.30
output	元4 · $0.60	元8 · $1.20

So in Beijing time: peak = 09:00–12:00 and 14:00–18:00, Mon–Fri (01:00–04:00 and 06:00–10:00 UTC); weekends are off-peak all day. The day-of-week check is evaluated on the Beijing wall clock, exactly as the billing rule states. The indicator is computed from the local clock, so it works offline and flips exactly at the boundary (refreshed every 30s; per-second while the panel is open).

Maintenance note — when DeepSeek changes prices, update RATES_CNY / RATES_USD in lib/client.js (they are the single source of truth). While V4 Pro is still separately priced, the card shows a short auto-expiring note; it disappears by itself once Pro is routed to Flash (PRO_ROUTED_FROM).

Install

# npm (recommended)
dsh plugin --profile web add dsh-deepseek-peak-indicator

# GitHub
dsh plugin --profile web add git+https://github.com/DDA-DIGITAL/dsh-deepseek-peak-indicator.git

# local checkout (development)
dsh plugin --profile web add link:/path/to/dsh-deepseek-peak-indicator

Then restart the web profile (dsh --profile web) so the loader picks up the new bundle. After that, the client half hot-reloads on changes (the built-in HMR poller watches plugin client bundles).

Development

No build step — the browser half (lib/client.js) is the shipped artifact.

npm test    # schedule + formatting unit tests (also runnable with TZ=UTC / TZ=Asia/Shanghai)
npm run check   # syntax-check the bundles

Edit lib/client.js → the running GUI hot-reloads the change (no restart). The panel is a React component registered into the conversation.input.right UI slot; styling lives in one scoped stylesheet (dpk-* classes) injected by the client half.

How it works

• cordis.patch.yml — the dsh.bundle.patch layer: inserts the plugin row { id: deepseek-peak-indicator, name: 'dsh-deepseek-peak-indicator' } into the profile's plugin tree.
• lib/index.js — node half: empty apply, so the row is a valid loader entry (the browser does all the work).
• lib/client.js — browser half (declared via dsh.client + exports["./client"], served at /plugins/dsh-deepseek-peak-indicator/client.js): registers a React occupant into the conversation.input.right UI slot. ConversationRoot renders that slot's entries immediately before the conversation.input.model seat, which places the dot exactly left of the model menu.

No server half, no network calls, no dependencies beyond the shell's static react module.

Compatibility

Targets the current DeepSeek Harness web client slots API (conversation.input.right). If a future DSH release reshapes the composer slots, this plugin may need a compatibility bump.

License

MIT — see LICENSE.