dsh-cost-tracker

🇨🇳 中文说明

Token cost tracking for DeepSeek Harness, with per-model configurable pricing and peak/off-peak rates.

What it does

• Prices every finalized LLM call (assistant/message usage) against a per-model pricing table you configure — cache-hit / cache-miss input, output, and optional peak-window rates.
• Publishes a per-session cost projection (total / peak / off-peak / per-model), consumed by the Web UI to show a live session cost readout.
• Flags calls that land in a configured peak window.
• Models without a pricing entry are surfaced as "unconfigured" with a placeholder, so you know to add them.

Installation

Install into a dsh profile straight from GitHub — the plugin ships a dsh.bundle layer, so dsh plugin add enables it automatically:

dsh plugin --profile web add github:yflmq001/dsh-cost-tracker

Pricing starts empty (models: {}); fill in your models under the plugin's config (see below). To override defaults, add a cost-tracker row to the profile's own cordis.patch.yml — later layers win by row id.

⚠️ When overriding, address the row by its id (cost-tracker), not by name. Cordis non-insert patches locate the target line by id; a name-keyed override row is silently dropped, leaving prices at models: {}.

The cross-session bill persists through dsh's storage domain; a profile without a storage backend keeps the bill in memory only.

Configuration

Pricing lives under the plugin's config (Schemastery-validated). All prices are currency units per million tokens; you fill them in manually (no scraping).

- id: cost-tracker
  config:
    models:
      deepseek-v4-flash:
        inputMiss: 1.0        # cache-miss input
        inputHit: 0.02        # cache-hit input (omit if no cache tier)
        output: 2.0
        peak:                 # peak tier: hours + toggle + prices
          hours: ["09:00-12:00", "14:00-18:00"]   # Beijing time
          enabled: true       # false = peak off (base rates always); omit = on
          inputMiss: 3.0
          inputHit: 0.10
          output: 9.0
      gpt-4o:                 # any model the harness can reach
        inputMiss: 2.5
        output: 10.0

Session projection

The plugin registers the cost projection:

{
  totalCost, peakCost, offpeakCost,
  callCount, unconfiguredCalls, unconfiguredModels,
  byModel: { [model]: { inputHit, inputMiss, output, total } },
}

Cache writes are priced at the miss rate (providers bill the write at the miss tier; the written tokens become cache hits on a later call).

Billing correctness

Two choices keep the numbers honest:

• Cache writes are billed at the miss rate, not the hit rate. Providers charge cache writes at the cache-miss input tier (e.g. DeepSeek $0.14/M vs $0.0028/M hit) — pricing a write as a hit undercounts by ~50x whenever a session writes fresh context. Some cost plugins fold cache writes into the hit bucket; this plugin does not.
• Unknown models are never silently priced. A model with no pricing entry is surfaced as "unconfigured" with a placeholder in the projection and UI, so a guessed default can't hide in your bill. You add the price, or you see the gap.

Known Limitations and Deferred Work

• Per-session cost is a projection over the durable log (replayed); the cross-session global bill and /cost command are available.
• Non-DeepSeek models need their pricing filled in manually; there is no automatic price lookup.
• Developer preview: assistant/message.usage is absent when the adapter reports no accounting, and the session format has no compatibility promise.