dsh-token-stats

中文

Cross-session token usage statistics for DeepSeek Harness: a Settings panel with a monthly stacked cost chart broken down by provider and model (USD), month navigation, provider/model filters, and a per-model legend.

Built as a standalone bundle (the harness's installable plugin format) — no changes to the core repository needed.

Features

• Overview — usage-bearing session count, the four provider-reported token buckets (uncached input, cache read, cache write, output), the grand total, and the estimated cost split per matched price table.
• Monthly cost chart — stacked bars of each day's per-model cost for the displayed month, with </> month navigation, provider and model filters, gridlines with fixed two-decimal USD ticks, and a color-coded legend.
• Per-model pricing — usage is priced via exact provider/model table matches (seeded with a deepseek / deepseek-v4-flash row), everything else falls back to the default table. Prices are fixed constants in this version; adjust DEFAULT_PRICES / DEFAULT_PRICE_TABLES in src/client/usage-aggregate.ts and rebuild if you need different rates.
• Data honesty — sessions whose cached projections predate the per-model units (or deployments without them) are counted as "no model data" and priced at the default table, with an explicit hint.

How it works

The bundle has two halves:

• Host half (src/host/) — registers two session projection units, modelUsage (per provider/model session totals) and modelDailyUsage (per provider/model/UTC-day cells), following the @deepseek-ai/dsh-token-meter pattern. Registration is idempotent against deployments that already ship the same units (the registry shares same-stateVersion keys), so it is safe on both release builds and source checkouts that already carry them.
• Client half (src/client/) — a settings.section contribution (Token 用量) that reads the session.list projection column (zero log loads) and renders the panel. Day cells are shifted from UTC to your local days for display.

Install

Prerequisite: a DeepSeek Harness installation (the dsh CLI or a source checkout).

From npm (once published)

dsh plugin --profile <name> add dsh-token-stats
dsh --profile <name> web

From this repository

git clone https://github.com/qiushui0901/dsh-token-stats.git
cd dsh-token-stats
npm install && npm run build
dsh plugin --profile <name> add ./dsh-token-stats
dsh --profile <name> web

From a source checkout of DeepSeek Harness

pnpm dsh web --patch /path/to/dsh-token-stats/cordis.patch.yml

(with dsh-token-stats linked into the checkout's node_modules, or installed into the profile the checkout boots.)

Open http://127.0.0.1:3080 → Settings → Token 用量.

Build

npm install
npm run build

Produces lib/host.js (self-contained Node entry) and lib/client.js (browser closure-factory bundle registered via window.__ModuleLoader__.load).

Known limitations

• Model data fills forward — sessions that ran before the modelUsage/modelDailyUsage projections existed stay unattributed until their projection cache rows rewrite; their usage is priced at the default table.
• Day bucketing is UTC-shifted — the host folds cells under UTC days and the client shifts by its timezone offset rounded to whole days, so a day's cells can sit on the neighbouring local day near the boundary.
• Prices are constants — the price tables are baked into the bundle; there is no in-UI editor (the editor was removed by request).

License

MIT — see LICENSE. Built on the DeepSeek Harness plugin system (MIT), which this project gratefully builds upon.