dsh-cost-tracker
为 DeepSeek Harness 提供按模型配置价格及峰时/非峰时费率的令牌成本跟踪
插件会安装到这里；不确定时保持 web。
npx -y @deepseek-ai/dsh plugin --profile web add github:yflmq001/dsh-cost-tracker#096b1ae917ae127f08aa3f0cad4ca320e58448ef
Cost Tracker 以 dsh-cost-tracker 发布，当前版本为 1.0.0。Plugin Hub 会校验它的 manifest，并保存精确安装来源，便于复现安装结果。
🇨🇳 中文说明
Token cost tracking for DeepSeek Harness, with per-model configurable pricing and peak/off-peak rates.
dsh-cost-tracker: per-session cost bar and the /cost command
assistant/message usage) against a
per-model pricing table you configure — cache-hit / cache-miss input, output,
and optional peak-window rates.cost projection (total / peak / off-peak / per-model),
consumed by the Web UI to show a live session cost readout.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 byname. Cordis non-insert patches locate the target line byid; aname-keyed override row is silently dropped, leaving prices atmodels: {}.
The cross-session bill persists through dsh's storage domain; a profile without a storage backend keeps the bill in memory only.
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
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).
Two choices keep the numbers honest:
/cost command are available.assistant/message.usage is absent when the adapter
reports no accounting, and the session format has no compatibility promise.