@stmol/dsh-rub-cost

dsh plugin --profile web add @stmol/dsh-rub-cost

dsh-rub-cost

A DeepSeek Harness plugin: it shows the cost of the current session in rubles in

the conversation dock, next to the built-in token counters

(conversation.composer.dock).

The capture above still shows the former "API cost" row, which the card replaced

with the price list; swap in a fresh screenshot the next time the card is opened.

The pill is a button. Clicking it opens a card styled after the built-in "Token

usage" popover; a second click, a click outside the card, or Escape closes it.

During peak hours the ruble amount on the pill is printed in amber — the same

warning colour the tariff row uses.

Installation

Published on npm as @stmol/dsh-rub-cost.

Any DSH bundle source works: an npm package, a Git repository, a tarball, or a

local folder.

• GUI: Settings → Plugins → Add plugin, then @stmol/dsh-rub-cost, or the

repository address.

• CLI: dsh plugin add @stmol/dsh-rub-cost.

The install name is the name field in package.json; the name in

cordis.patch.yml and the factory id in client.js have to match it, or the

bundle will not build. There are no automatic updates: remove the plugin and add

it again. The plugin runs with your own privileges.

What the card shows

Model, the current tariff and the countdown to its switch, the API price list in

dollars per 1M output tokens at the tariff in effect,

the CBR rate with its publication date, the cost of the session including

subagents, the same session without a cache and the savings, and the DeepSeek

Platform wallet balances. The session cost itself is the ruble amount on the pill

and in the card header; the card does not repeat it in dollars.

Every ruble amount is shown with two decimals. The price list also keeps two and

adds a third or fourth digit when needed to preserve a finer rate. The token breakdown stays out of the

card, because the neighbouring "Token usage" cell already prints it.

Peak hours cost twice the off-peak rate. The tariff row names only the window in

effect, for example peak hours (04:00-07:00) highlighted in amber, or

off-peak. Peak hours are Beijing weekdays 09:00-12:00 and 14:00-18:00, printed

in your browser's timezone — 04:00-07:00 and 09:00-13:00 at UTC+3, and the same

windows shifted by your own offset; weekends are off-peak around the

clock. The tariff itself is never computed in your zone: the rule stays in

Beijing time, and only the printed clock moves, so the amount is the same

wherever the card is opened. The countdown follows from that rule alone, so on a

Friday it runs into the next week: 63h 00m just after 13:00 MSK, 61h 47m at

14:13 MSK. The rule assumes the following Monday is an ordinary working day; a

PRC holiday moves the real switch, as the first limitation below explains.

How it is computed

cost = (uncachedInput + cacheWrite) × cacheMiss + cacheRead × cacheHit + output × output

Tokens are priced with DeepSeek's official dollar list per 1M tokens, then

converted at the CBR USD/RUB rate. The account is billed in dollars, so the

dollar list is the basis; the yuan list on the Chinese documentation page is the

same list at the fixed 20/3 ¥/$ rate. A usage report can show amounts in CNY

while a granted yuan balance is being spent, which is that same list again.

Tokens come from the host tokenUsage projection and the model from

modelSelection.lastUsed; the plugin never folds the session journal itself. The

price rows print the same model's MODEL_PRICES entry for the tariff in effect,

so they move only when the tariff or the model does. The rate comes from a mirror

of the CBR rates and is cached in localStorage: with no network the last known

rate is used, and with no cache the cell shows a dash and the reason in the

tooltip.

The pill and the card header are a running total: the cost of the current session,

not of the account and not a per-request price. Subagents are priced from their

own route and summed into a separate "Total with subagents" row; while the card is

open their spend is polled every 15 seconds. If even one child cannot be read, that

row is hidden rather than shown incomplete.

Known limitations

• PRC statutory holidays are not modelled. The plugin keeps one year-agnostic

rule — Beijing weekday plus peak window — so it needs no annual edit, and a

holiday weekday is therefore priced as a peak day. DeepSeek bills those days at

the off-peak rate, so on such a day the amount and the countdown are both too

high until the holiday is over. The official dates live in the State Council

notice published every November, which is the only source; the client bundle

ships no year-specific table on purpose.

• The ruble total is not a sum of per-request bills: the whole session is repriced

at the tariff in effect when the card renders, so a peak/off-peak switch doubles

or halves the amount without a single new token. The tariff and the price rows

name what is in effect, which is exactly what the countdown counts down to.

• tokenUsage holds a whole-session total with no per-route split, so after a

mid-session model change everything is estimated with the last used model.

• Cache writes are billed at the cache miss rate: DeepSeek publishes no separate

price, and on DeepSeek routes that counter is zero anyway.

• Subagents use their own last used model at the current tariff, and a child that

has not chosen a route yet falls back to the parent's.

• The balance request pins x-client-version. The version is hardcoded in

client.js and has to be updated by hand after a DSH upgrade.

• The clock zone comes from the browser, not from the DeepSeek account: the

account profile (remote.account.getProfile) exposes no timezone. When t

…(truncated)