dsh-token-price — DSH 实时 Token 花费 + 账户余额显示插件

[spoon-man569]

给 DeepSeek Harness（DSH）Web 聊天界面加一个实时显示 Token 花费和账户余额的小插件

功能：

每轮花费：每条 AI 回复的统计行（时间 · 用时 · 首 token · tok/s）末尾显示本轮花费，跟随悬停出现

总花费：输入框下方统计条区域显示当前会话累计花费
账户余额：实时显示 DeepSeek 账户余额（CNY，来自官方 GET /user/balance）

计费贴近官方：按官方峰谷定价（北京时间 9:00-12:00、14:00-18:00 为高峰，其余半价）计算，并区分缓存命中
实时汇率：价格表按官方美元价维护，USD→CNY 汇率由浏览器实时抓取（免费 API 双源 + 缓存），不需要手动维护汇率

仓库

https://github.com/spoon-man569/dsh-token-price

欢迎 Star

[1320800521]

Nice plugin — putting per-turn cost next to token/latency stats is much more useful than a separate billing page.

One implementation detail I would add before relying on the number for long-running sessions: version the pricing rule by effective timestamp and model alias instead of keeping one timeless price table.

For an auditable cost record, I would persist at least:

• requested model alias, e.g. deepseek-v4-pro
• resolved/version label when it is known
• cache-hit input tokens
• cache-miss input tokens
• output tokens
• pricing-rule version / effective time
• calculated cost and currency

That keeps historical sessions explainable when DeepSeek updates pricing or moves an alias to a newer backend. I would also keep the /user/balance fetch independent from token-cost calculation, so a pricing-table update cannot affect balance correctness.

More broadly, this is one reason I think the useful unit for coding agents is becoming model × harness × workflow, not the model alone: the harness is where accounting, tool execution, permissions, retries, workspace state, and observability become operational behavior.

I put my current V4-Pro-0813 + DeepSeek Harness developer notes (official API surface, pricing/limits, compatibility, and adoption boundaries) here:

https://www.xbstack.com/en/ai/tools-lab/deepseek-v4-pro-0813-harness/?utm_source=github&utm_medium=referral&utm_campaign=deepseek_v4_pro_0813_harness&utm_content=discussion_1878

[spoon-man569]

Thanks for the detailed look! Fair hit on the pricing versioning — DeepSeek just moved to peak/off-peak and aliases can retarget to newer backends, so a long-running session would get repriced against whatever table is current if prices shift. I already track per-turn cache-hit / cache-miss / output tokens and request time in the projection, so stamping the pricing rule (version + effective time) with the computed cost is really the missing piece — it's on the list for the next pass.

Small clarification: balance is actually already kept out of the costing path — it's a host-side /user/balance call that just rides along the projection, while costs are computed in the browser against the price table, so updating prices can't touch it.

And I agree with that the useful unit is becoming model × harness × workflow, since the harness is where accounting, tooling, and observability actually become operational behavior. Will read through your V4-Pro-0813 notes.

[1320800521]

@spoon-man569 Thanks for the clarification — if /user/balance is already fully outside the costing path, I would leave that separation exactly as-is.

Since you already persist per-turn token classes and request time, the smallest durable addition seems to be a pricing_rule_id (or rule version/effective timestamp) plus the computed cost snapshot on each turn. Then historical sessions never need to be re-priced against a newer peak/off-peak table, while you can still recompute deliberately when auditing a rule change.

The edge case I would test in the next pass is one long-running session that crosses a pricing effective-time boundary or an alias retarget: turns before and after the boundary should retain the rule that was effective when each request was made. That would make the accounting trail much easier to explain later.

[lieve-c]

你的连接进不去
@spoon-man569

[spoon-man569]

不好意思，我没有设置好地址，现在已经修复

[lieve-c]

可以看看我的https://github.com/lieve-c/dsh-beautify

[spoon-man569]

感觉很强啊，star一下致敬

[lieve-c]

你的我也用上了 厉害👍🏻

[spoon-man569]

目前v0.1版本，还有部分明显的bug，比如不同对话之间价格的对齐，好用的话可以关注后续更新😀

[1320800521]

感谢提醒。我刚重新检查了之前那篇页面，现在公网访问已经正常，返回 HTTP 200：

https://www.xbstack.com/en/ai/tools-lab/deepseek-v4-pro-0813-harness/

前面那次更像是当时的部署或边缘访问异常，不是链接已经下线。如果你那边现在仍然打不开，可以把具体报错或地区告诉我，我再继续查路由/缓存问题。