dsh-ui-cost-meter

dsh-ui-cost-meter

English | 日本語

Real-time spend for the DeepSeek Harness Web GUI: one pill under the composer, beside the shipped turn/token pills, that says how much the session has cost so far.

⏱ 1 turns 71 steps · 285 tok/s    🗄 7.3M tok · Cache hit 99%    $ 0.42

The amount rides the session projection seam, so it is whole-log (paging and compaction cannot change it), it survives reloads, and it is priced from the durable provider usage the harness already logs. The pill is shaped like the shipped stats pills: clicking it opens a breakdown panel that repeats their dialog's title, rule, and label/value rows, naming each cache bucket, the unpriced tokens, and the route that priced them.

Install

dsh plugin --profile web add dsh-ui-cost-meter

Under a source checkout, add link:<path> works the same way. The plugin declares its own bundle patch, so the row appears in the profile without editing cordis.patch.yml.

Configure

Every field lives on the plugin row, and every price is per one million tokens in whatever currency symbol names.

No configuration is needed. The plugin ships a rate table for every model OpenCode Go, OpenCode Zen, and Command Code served when it was generated:

Write rates only for the routes where the shipped number is wrong for you — a peak-hour price, a different plan, a provider the table has never heard of. Your key replaces the shipped one for that route:

- id: cost-meter
  name: dsh-ui-cost-meter
  config:
    symbol: '$'
    rates:
      deepseek-official/deepseek-flash:
        input: 0.15
        output: 0.6
        cacheRead: 0.003
        cacheWrite: 0

The shipped keys carry the provider names opencode-go, opencode, and command-code. A profile that names the same provider something else gets no default for it and needs its own rates entries.

What the shipped table does not promise

One rate per route can only carry one number, so the table holds the off-peak price each provider publishes:

• DeepSeek's list price and Command Code's V4.1 Flash and V4 Flash are exactly double at peak, which runs 01:00-04:00 and 06:00-10:00 UTC, Monday through Friday. Command Code bills V4 Flash Fast flat, and doubles everything on V4.1 Flash Fast except the cache read.
• No provider here bills a cache write, so cacheWrite is 0, except on OpenCode Zen where the published price has one.
• Command Code's GOAT plan is not a per-token invoice: the $10/month plan grants $60 of DeepSeek V4.1 Flash credits a month, and V4.1 Flash Fast draws on the same allowance, so the pill reads as credits burned rather than an amount owed. OpenCode Go's $10/month plan likewise meters its rates against the plan's limits.
• Ten OpenCode Go ids and two OpenCode Zen ids that the endpoints serve have no published price anywhere upstream. The table leaves them out on purpose rather than guess.

A route with no matching rate is not guessed at: its tokens are counted as unpriced, contribute nothing to the total, and make the pill mark its amount with a trailing +. The panel then names how many tokens were left unpriced, so a missing rate is visible instead of silently wrong.

Where the numbers come from

The host half registers the costMeter session projection. It folds the durable log: a request/header sets the route, and each settled assistant/message (or assistant/attempt, which preserves a billed attempt that left no surface message) contributes the provider usage embedded in its stream. A re-reported sample for the same turn and step replaces the earlier one; llm/retry-started closes that slot so a retried attempt adds instead. Input, output, cache-read, and cache-write tokens are then priced by the route's rates, with the shipped table folded under the row's own rates.

The client half registers one entry on the conversation.composer.dock slot and renders it into the shipped stats row ([data-composer-stats]) with a React portal, so the cost pill shares that row's font, spacing, and vertical rhythm instead of starting a row of its own. Shells without that marker (DSH 0.2.1 and later) get the pill placed straight into the dock's own row, unwrapped.

The pill and its panel carry the shell's own declarations — the same --dsw-* tokens, sizes, radius, and two-column row grid as the shipped pills and their dialog — because a third-party bundle can only reach the platform modules the shell seeds, not its CSS modules.

Caveats

• Not a billing record. The figures are the harness's own token accounting times the rates in play. The provider's invoice is the authority; a wrong rate is a wrong pill.
• One rate per route. A provider that changes its price mid-session, or a route whose price depends on the request size, needs a rate that reflects what you actually pay.
• Currency is yours to name. Nothing here converts between currencies.

Development

bun install
bun run build     # lib/index.js (host half) and lib/client.js (browser half)
bun test          # pricing fold, formatting, and the built bundle
bun run check     # type check
bun run lint      # biome

src/rates.generated.ts is the price list the plugin ships, and bun run rates:update rebuilds it. The generator reads models.dev plus each provider's /models endpoint for OpenCode, and each Command Code model page for Command Code. Ids upstream does not price are skipped and named on stderr, so refreshing the table is one command and a commit.

lib/ is generated and ignored by git. build.ts bundles the host half as ESM (with zod and @deepseek-ai/schemastery left external) and the browser half as CJS wrapped in the window.__ModuleLoader__.load({ id, factory }) loader format the client module system serves.

License

MIT