dsh-cost-governor

English | 中文

dsh-cost-governor is a plugin for DeepSeek Harness (dsh) that provides real-time token auditing, monetary cost tracking, dynamic model routing, and context compaction gating.

Powered by Cordis, it integrates directly with the Context lifecycle to monitor usage across agent sessions.

Overview

The plugin tracks prompt, completion, and prompt-cache hit tokens across sessions. It calculates monetary costs using configured model pricing matrices (including DeepSeek-V3, DeepSeek-R1, OpenAI, and Claude), recommends or degrades models based on budget constraints, and emits compaction signals when context token usage crosses defined thresholds.

Features

• Token and cost ledger: Tracks prompt, completion, and cache-hit tokens per session with estimated costs.
• Model pricing matrix: Pre-configured rates for DeepSeek, OpenAI, and Anthropic models, with support for custom rates.
• Dynamic model routing: Recommends appropriate models based on task complexity and automatically downgrades when budget limits are breached.
• Context compaction threshold: Emits compaction triggers to Cordis (ctx.compaction) when context token usage reaches configured percentages.
• Agent tool integration: Registers the get_cost_stats tool for real-time cost breakdown queries during agent sessions.

Installation

Add dsh-cost-governor to your DeepSeek Harness installation or workspace:

dsh plugin add github:echo-escape/dsh-cost-governor

Or install via pnpm / npm:

pnpm add @echo-escape/dsh-cost-governor

Configuration

Configure the plugin in your session's cordis.yml:

plugins:
  cost-governor:
    enabled: true
    config:
      currency: "USD"
      maxSessionBudgetUsd: 5.0
      compactionThresholdPercent: 80
      modelRates:
        deepseek-chat:
          inputPerM: 0.27
          outputPerM: 1.10
          cacheHitInputPerM: 0.07
        deepseek-reasoner:
          inputPerM: 0.55
          outputPerM: 2.19
          cacheHitInputPerM: 0.14

Usage

The plugin automatically registers the get_cost_stats tool on ctx.tools. An agent or caller can execute this tool to retrieve structured cost statistics:

{
  "status": "success",
  "currency": "USD",
  "totalTokens": 150000,
  "promptTokens": 100000,
  "completionTokens": 50000,
  "cacheHitTokens": 20000,
  "totalCostUsd": 0.0806,
  "recordCount": 5,
  "byModel": {
    "deepseek-chat": {
      "promptTokens": 100000,
      "completionTokens": 50000,
      "costUsd": 0.0806
    }
  }
}

Development

Run tests and type checking:

pnpm run typecheck
pnpm run test

Build the package:

pnpm run build

License

MIT