dsh-usage-cost

Usage and cost tracking plugin for DeepSeek Harness, inspired by Codex 0.144's "usage limit" credits reminder: watch every model call and keep a running token/cost tally, with warnings before the budget runs out.

English | 中文

What it does

• Stream-level usage accounting: hooks the llm/stream waterfall and accumulates inputTokens, outputTokens, cacheReadTokens, cacheWriteTokens, reasoningTokens from every usage chunk without buffering or altering the stream; one call counts once.
• Per-model cost estimation: prices each call in USD from a built-in table (DeepSeek: $0.28/M in, $0.42/M out, $0.028/M cache read, $0.42/M cache write; others: $2/M in, $8/M out), overridable per provider/model key in the config.
• Per-turn settlement: on agent/turn-stopping, usage accumulated since the last settlement is appended (deltas only) to ~/.dsh/dsh-usage-cost/stats.json, bucketed by date and by model.
• Budget warnings: with dailyBudgetUsd / sessionBudgetUsd set, crossing the threshold logs once per date and once per session: [dsh-usage-cost] 今日用量超预算 / 本次会话用量超预算.
• usage_report model tool: read-only report of today, the current session, the per-model breakdown and the budget state (dailyExceeded / sessionExceeded flags).
• Package-private RPC: usage-cost/status returns the same report.

Install

dsh plugin --profile web add github:DamonKoy/dsh-plugins#path:/packages/dsh-usage-cost

Restart dsh web. Active by default, records usage with no config.

Config

Optional file ~/.dsh/dsh-usage-cost.json:

{
  "prices": {
    "deepseek-official/deepseek-v4-flash": {
      "inputPerM": 0.28,
      "outputPerM": 0.42,
      "cacheReadPerM": 0.028,
      "cacheWritePerM": 0.42
    }
  },
  "dailyBudgetUsd": 2,
  "sessionBudgetUsd": 0.5
}

Config is re-read on every use, so edits apply without a restart.

Stats file

Written synchronously at every turn settlement (small file, no races); costs are estimates from the price table, not provider invoices:

{
  "byDate": {
    "2026-08-16": { "calls": 3, "inputTokens": 1200, "outputTokens": 800, "costUsd": 0.00068 }
  },
  "byModel": {
    "deepseek-official/deepseek-v4-flash": { "calls": 3, "inputTokens": 1200, "outputTokens": 800, "costUsd": 0.00068 }
  }
}

Notes

• Session tally is in-memory for the process lifetime; the daily file is the durable record.
• The llm/stream listener calls next() first and only reads chunk fields; stream, options and chunks are never modified.
• Warnings fire once per date and once per session to avoid log spam.
• Prices and budgets are user-side estimates; adjust them to your billing.

License

MIT