dsh-cost-chip

English | 中文

Token usage and money, where you already look: one compact line under every completed answer, and the running session total above the composer — priced the way DeepSeek actually bills, with peak/idle rates applied per turn.

What it does

DeepSeek Harness already counts tokens, but the numbers live in a popover. This plugin puts them on screen and turns them into money:

• Per-turn line under each completed answer: 本轮 12.3k tok · ¥0.0345 (Turn 12.3k tok · ¥0.0345 in English).
• Session line above the composer: 会话 2.11M tok · ¥0.2498, updating as the turn streams and settles.
• Hover for the breakdown: uncached input, cache read, output, the model, and whether that reading was priced at the peak or idle rate.
• Each turn keeps its own price. DeepSeek bills peak and idle hours differently, so a reading is priced by its own timestamp instead of the current clock; a session that crosses 12:00 or 18:00 prices every turn correctly.
• A bundled skill. Installing the plugin also registers task-token-cost, so you can ask the agent for cross-session reports, per-day totals, or CSV/JSON — the same numbers, aggregated over every session on disk.

Zero dependencies, zero network calls, no files written: everything is read from the session projections the harness already maintains.

How to use

1. Install it (below) and open any conversation. Finished turns get the line automatically; the composer line appears as soon as the session has a model call.

2. Point at either line to see the split. Nothing to configure, no panel to open.

3. To go beyond the current session, just ask in natural language — the bundled skill handles it:

   最近三天每个任务花了多少 token 和钱？
   把这个月的用量导出成 CSV
   哪个会话最贵？
   

   The skill answers from $DSH_HOME/sessions, so it also covers sessions you are not looking at, and can emit per-turn, per-day or per-model totals.

4. Price change? Edit PRICES at the top of client.js for the on-screen readings and skills/task-token-cost/prices.json for the reports. Models missing from the table fall back to the Flash rate and are marked with ≈.

Screenshots

The turn line sits at the end of every finished answer — tokens for that turn, and what it cost at that turn's own peak/idle rate:

The composer line keeps the running total for the whole session, next to the harness's own speed and cache-hit pills:

Hovering either line opens the split — uncached input, cache read, output, model, and the rate that was applied. The figures above are real readings from a deepseek-flash session, priced at the idle rate. The ≈ on the session reading means the session is longer than the loaded transcript: the token count is still the whole session, while the money covers the turns currently in the window.

Install

• Market (recommended once listed): Plugins → Discover, search dsh-cost-chip, install.
• From this repository:

  dsh plugin add github:Yearngg/dsh-cost-chip
  

• From npm, after it is published there:

  dsh plugin add dsh-cost-chip
  

Requires DeepSeek Harness 0.2.1-rc.1 or newer — it uses the conversation.chat.turnTail and conversation.composer.dock slots. Reload the page after installing; no restart needed.

Pricing

Prices are CNY per 1,000,000 tokens and mirror the public DeepSeek list (checked 2026-10):

Model	Idle: miss / hit / output	Peak: miss / hit / output
deepseek-flash	¥1 / ¥0.02 / ¥4	¥2 / ¥0.04 / ¥8
deepseek-v4-pro	¥4.5 / ¥0.15 / ¥13.5	¥9 / ¥0.3 / ¥27

Peak hours are Beijing time, Monday–Friday, 09:00–12:00 and 14:00–18:00; everything else is idle. Cache writes are not billed. Figures are estimates: they follow the list price, not your invoice, and statutory holidays are priced as peak because the plugin cannot know the holiday calendar.

The bundled skill

skills/task-token-cost/ ships the task-token-cost skill and its CLI:

node <skill-dir>/scripts/token-cost.mjs report              # per-session tokens + cost
node <skill-dir>/scripts/token-cost.mjs report --by turn    # per-turn detail
node <skill-dir>/scripts/token-cost.mjs report --since 7d --csv report.csv
node <skill-dir>/scripts/token-cost.mjs prices              # show the active price table

The plugin registers it as a runtime skill, so it outranks a same-named copy under ~/.dsh/skills. Two skills with one name would shadow each other; delete the user-level copy if you installed one earlier.

Privacy

The readings come from the running session's tokenUsage projection and each completed turn's folded usage; the skill reads $DSH_HOME/sessions/**/session.v4.jsonl.zstd. Nothing leaves the machine and nothing is written.

Development

The package is a plain Cordis bundle: cordis.patch.yml inserts the token-cost row, index.js registers the skill, client.js registers the two slots. No build step — a git install runs the source directly. Edit client.js, reload the page, done.

License

MIT

---

中文

作用：把 DeepSeek Harness 已经统计的 token 用量直接摆到界面里，并折算成金额。

• 每条回答完成后的尾部加一行 本轮 12.3k tok · ¥0.0345；
• 输入框上方显示整个会话累计 会话 2.11M tok · ¥0.2498；
• 鼠标悬停看拆分：未命中输入 / 缓存命中 / 输出、模型、当时按高峰还是闲时计价；
• 每轮按它自己的时间戳定价，跨 12:00、18:00 的会话不会算错；
• 顺带把 task-token-cost 技能注册进来，可以让 agent 跨会话出报表、导 CSV/JSON。

零依赖、不联网、不写文件。

截图：

每条回答末尾的本轮读数——本轮的 token 和按该轮时段单价算出的金额：

输入框上方的会话累计，排在宿主自带的 tok/s 和缓存命中率旁边：

指着任一行就能展开拆分（未命中输入 / 缓存命中 / 输出、模型、当时按高峰还是闲时）。上面两张是 deepseek-flash 会话里的真实读数，按闲时价计。会话行前面的 ≈ 表示这个会话比已加载的对话窗口更长：token 数仍是整个会话的，金额只覆盖窗口里已有的轮次。

怎么用：

1. 安装后打开任意会话，完成的轮次自动出现那一行，输入框上方随会话出现累计行，无需配置。
2. 鼠标指上去看明细。
3. 想看更大范围，直接用自然语言问 agent，例如「最近三天每个任务花了多少」「导出这个月用量」「哪个会话最贵」——由内置技能回答。
4. 调价改 client.js 顶部的 PRICES（界面）和 skills/task-token-cost/prices.json（报表）。

安装：

dsh plugin add github:Yearngg/dsh-cost-chip   # 本仓库
dsh plugin add dsh-cost-chip                  # 发布到 npm 之后

也可以在市场里搜索 dsh-cost-chip 一键安装（收录后）。需要 dsh 0.2.1-rc.1 及以上，装完刷新页面即可，不用重启。

口径：金额 = Σ 每次调用（未命中输入×未命中价 + 缓存命中×命中价 + 输出×输出价）÷ 1,000,000，高峰为北京时间工作日 9:00–12:00、14:00–18:00，其余闲时；缓存写入不计费；法定节假日无法识别，按高峰计。