dsh-billing

v0.6.8 host update: Typechecked and tested against Harness 0.2.1-alpha.3. Peer ranges also accept 0.2.0-rc.2, 0.1.5-rc.2, and the 0.1.1-rc.2 baseline. The official DeepSeek price card is unchanged from v0.6.7.

简体中文 | English

Session billing and quota plugins for DeepSeek Harness.

[!NOTE] This is an independent community project. It is not part of the official DeepSeek Harness distribution. Costs are local reference values, not invoices, and do not automatically block model calls.

Components

The host owns pricing and the projection; the browser renders the host-computed projection. The installable root bundle exports both the host and Web client entry points, so GitHub installation does not depend on separately published internal packages. The same model ID under different providers is tracked independently, for example deepseek/deepseek-v4-flash and openrouter/deepseek-v4-flash.

v0.6.8 keeps the v0.6.7 official rate card and adapts the installable bundle to Harness 0.2.1-alpha.3.

Install As A Bundle

The repository root provides a bundle declaration and both runtime packages. Add it to the web profile:

dsh plugin --profile web add https://github.com/Wanbinyu/dsh-billing/releases/download/v0.6.8/dsh-billing-community-bundle-0.6.8.tgz

Restart dsh after installation. One billing composition entry enables both the host projection and Web cost strip. Pricing uses explicit configuration first, then the built-in USD model catalog.

Manual Installation

When the host project needs to own the composition layer, install both packages:

npm install ./packages/dsh-billing ./packages/dsh-client-ui-billing

Then add this to the profile's cordis.patch.yml:

- insert:
    - id: billing
      name: dsh-billing
      config: {}
    - id: ui-billing
      name: dsh-client-ui-billing

Configure Pricing And Quota

With models left empty, deepseek-flash, deepseek-v4-flash, deepseek-v4-flash-0731, deepseek-v4-flash-vision-exp, and deepseek-v4-pro on the deepseek and deepseek-official providers use the official USD card. Each usage sample is frozen at its own timestamp. From 2026-09-10 04:00 UTC the Flash family uses the V4.1-Flash card and Pro keeps its own card. Peak is Monday–Friday 01:00–04:00 and 06:00–10:00 UTC. Chinese public holidays in 2026 are off-peak for the whole China calendar day. Cache writes use the cache-miss rate. Provider-reported image tokens are already inside the input usage.

An explicit models entry replaces that schedule with one flat rate. Every other provider still uses the built-in USD catalog.

- id: billing
  config:
    currency: USD
    quota:
      limit: 5

Exact provider/model keys, such as openrouter/deepseek-v4-flash, take precedence. A model-only key such as deepseek-v4-flash remains supported as a fallback for every provider. When overriding the billing row after the bundle inserted it, Harness replaces the whole config block. Restate every field you want to keep.

The built-in catalog and the official schedule are USD-only. When using CNY or another currency, configure every model explicitly. A model without a price is still counted, but joins unpricedModels and sets quota.estimated to true. Its cost is excluded rather than fabricated, so quota progress containing unpriced usage must not be treated as a complete bill.

Projection

interface BillingProjection {
  currency: string
  totalCost: number
  models: { provider: string; model: string; cost: number; uncachedInputTokens: number;
            outputTokens: number; cacheReadTokens: number;
            cacheWriteTokens: number }[]
  unpricedModels: string[]
  latestTurn?: { turn: number; cost: number; uncachedInputTokens: number;
                 outputTokens: number; cacheReadTokens: number;
                 cacheWriteTokens: number; unpricedModels: string[] }
  quota?: { limit: number; used: number; remaining: number; percent: number; estimated: boolean }
}

Usage follows the request/header model for step attribution. A later sample for the same (turn, step) replaces the earlier sample in both the session total and latest-turn summary to avoid double counting; usage without a preceding header falls into a reserved (unknown) bucket. latestTurn appears after the first usage sample and lets clients render cost and token details for the most recent turn.

The Web strip shows both Turn and Session amounts. Hovering reveals the latest turn's input, output, cache-read, and cache-write tokens plus per-model costs. Quota feedback becomes progressively stronger at 50%, 80%, and 100% usage.

Development And Verification

Host configuration, projection replacement, quota states, and the Web BillingBar server-rendered output have automated coverage. Installation and viewport composition in a real Harness Web profile remain manual release checks. Run the complete verification with:

npm run build
npm run verify

The built-in catalog is generated from the pi-ai model catalog with:

node packages/dsh-billing/scripts/generate-catalog.mjs

Current Limitations

• Costs are local reference values, not invoices or a hard gating input.
• Quota is currently per session; deployment-wide budgets are deferred.
• Compatible with DeepSeek Harness 0.1.0-rc.6 through rc.8, 0.1.1-rc.1 through rc.2, and 0.1.2-rc.1; development dependencies are pinned to 0.1.2-rc.1, while @deepseek-ai/dsh-client-runtime remains on the official next package 0.1.1-rc.2.

Links

• DeepSeek Harness
• GitHub repository
• 简体中文说明

License

MIT.