dsh-billing

Per-model cost accounting and session quota progress plugins for DeepSeek Harness (dsh).

• dsh-billing (host plugin) — prices provider-reported token usage per model from your Config, folds it into the billing session projection, and reports optional per-session quota progress. Replay-aware, config-independent fold state, deterministic micro-unit rounding.
• dsh-client-ui-billing (web GUI plugin) — renders a read-only cost strip in the composer dock (order 15, after GoalBar): session cost in your currency, a quota progress bar with used/limit text, an unpriced-model warning, and the per-model breakdown in a tooltip. Renders nothing until usage exists.

They follow the standard DeepSeek Harness host-projection / client-surface split (the session-stats / ui-goal precedent): the host owns pricing and the fold, the browser renders the host-computed projection.

Install

Both packages are plain npm packages with @deepseek-ai/dsh-* peer dependencies (installed with dsh).

npm install dsh-billing dsh-client-ui-billing

In a dsh profile, add the rows to cordis.patch.yml (see the cordis primer):

- insert:
    - id: billing
      name: 'dsh-billing'
      config:
        models:
          deepseek-v4-flash:
            input: 0.2        # currency per 1M uncached input tokens
            output: 0.8       # currency per 1M output tokens
            cacheRead: 0.05   # currency per 1M cache-read tokens (absent = 0)
            cacheWrite: 0.1   # currency per 1M cache-write tokens (absent = 0)
        currency: USD
        quota:
          limit: 5            # optional per-session cost cap (currency)
    - id: ui-billing
      name: 'dsh-client-ui-billing'

Prices are examples — set them to your contract rates. A model without an entry still has its tokens counted but prices at zero and joins unpricedModels, so you learn about missing prices instead of silently under-billing. Restart dsh after editing the profile.

Projection

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