メインコンテンツへスキップ
G

dsh-money-board

github798383392/dsh-money-board

Session cost board for DeepSeek Harness: reads save-token's token counters and prices them with the official DeepSeek table, showing session spend, compression savings and the peak/off-peak delta above the composer.

インストール

dsh plugin --profile web add github:github798383392/dsh-money-board

README

dsh-money-board

Session cost board for DeepSeek Harness — turns token counters into money.

save-token tells you how many tokens you saved. dsh-cost-meter tells you what you spent. Neither tells you what the savings were worth. This plugin closes that gap: it multiplies the real token counters by DeepSeek's official price table and puts the answer above the composer.

● 省钱看板 谷 ¥3.0554 · 省 ¥0.0276~1.3778 · 错峰 ¥3.0554

Hover for the breakdown:

deepseek-flash · 谷时 · 09:00 转峰时 ×2
本会话花费                          ¥3.0554
未命中输入                          ¥0.4949   16.2%
缓存读                              ¥1.5742   51.5%
输出                                ¥0.9863   32.3%
压缩省下                     ¥0.0276~1.3778   0.9%
错峰可省                            ¥3.0554
150K 输入 · 17.1M 缓存 · 89.9K 输出 · 命中 98.6% · 260 请求 · 压缩 11 次

Requirements

  • DeepSeek Harness (dsh) with its Web GUI
  • dsh-plugin-save-token — provides the token counters (/save-token/api/dashboard)
  • dsh-cost-meter — provides the official price table and peak/off-peak windows

The badge degrades gracefully and says which one is missing.

Install

dsh plugin --profile web add dsh-money-board

Or install it from the DSH plugin market / Settings → Plugins → Add plugin, then refresh the page.

How it works

  1. The client reads GET /save-token/api/dashboard (same origin, every 30 s).
  2. It posts that payload to the host half at POST /money-board/api/compute.
  3. The host imports dsh-cost-meter's own pricing.js and uses DEFAULT_PRICE_TABLE(_CNY), DEFAULT_PEAK_WINDOWS, peakPhaseAt() and costOf() to price every bucket.

No prices are hardcoded. When dsh-cost-meter syncs official pricing, this plugin follows automatically. Peak/off-peak uses the same constants: UTC 01–04 and 06–10 (09:00–12:00 and 14:00–18:00 Beijing), with weekends and Chinese public holidays fully off-peak.

Configuration

cordis.patch.yml → config:

keydefaultmeaning
currencyCNYCNY uses the CNY table, USD the USD table
modelautoauto reads the model from the profile's agent-default-model; or pin a model id

Why the savings are a range

save-token exposes only an aggregate avoidedTokens, so the plugin bounds it:

  • low — every avoided token priced as a cache read (¥0.02/1M). History replays bill at exactly that rate, so the truth sits near the low end.
  • high — every avoided token priced as a cache miss (¥1/1M). Only holds if the removed content was brand new and sent once.

On DeepSeek that gap is ~50×, which is the point: compression barely moves the bill; peak/off-peak pricing moves it by 2×.

Limitations

  • Counters live in the save-token process and reset when DSH restarts; the badge then shows an idle state.
  • Depends on save-token's private route and on dsh-cost-meter's lib/pricing.js. If either changes shape, the badge reports pricing/save-token missing instead of guessing.
  • Refresh rate is 30 s; this is not per-request metering.
  • Client half is unverified only in the sense that visual regressions are the author's, not the API's: the compute endpoint is covered by the CLI in the same repository.

License

MIT © YG

関連プラグイン