Why

Each DSH session tells you what that one session used. After a while the questions you actually have are different: how much did I use this month? Which model do I lean on? How much went through each relay versus the official API? Did the tokens go to system prompts, tool calls or reasoning? And what did all of that cost?

dsh-profile-stats adds a profile page to the sidebar that answers those questions from every session log on this machine.

• Works out of the box: no API key, no setup. The first open scans this machine's session logs; after that the host tops up silently in the background every five minutes.
• Numbers you can trace: token counts are the providers' own figures as logged, and every section is summed from the same daily entries, so they agree with each other.
• Stays on your machine: logs are read only through DSH's ctx.sessionQuery; the one outbound request is a check of this package's latest version on npm.
• Zero dependencies: no runtime dependencies and no other plugins required.

Install

In DSH open Plugins → Add plugin and paste the package name:

dsh-profile-stats

A person icon appears in the sidebar. The first open shows scan progress and can take a while on a machine with many sessions; the result is cached, so later opens only read sessions that are still running or new.

What's on the page

Header: range, export, refresh

• Range: today, this week (Monday to today), this month (the 1st to today), all time (the default) or custom dates. It sets the period for the token and spend numbers, the token stats, the projects and the provider split; the activity grid keeps showing the whole year but outlines the days in range. Token by model has its own switch.
• Export: the selected range as CSV (one row per day: input, output, cache read, cache write, then a column per provider) or JSON (the full snapshot).
• Refresh: rescan now. When a newer release exists, an Update to x.y.z button shows up here (registry installs only).

Five headline numbers

Card	Meaning
Tokens	Tokens in the selected range. The small trend on the right is the last 7 days as bars for today, the last 4 weeks as bars for this week, a daily line for this month and the 7-day average over the last 90 days for all time; above it is the change against the same days of the previous period (red up, green down)
Spend	Spend in the selected range — see How spend is worked out
Longest task	The longest span from a turn's start to its end, over all history
Active streak	Consecutive days with usage up to today, with your best beside it
Cache hit today	Today's hit rate across all providers, to two decimals: cache read ÷ (uncached input + cache read + cache write)

Token activity

A day grid of the last 12 months (53 weeks, starting on Monday) with three views:

• Daily: each day coloured by the quartile of non-zero days;
• Weekly: one bar per week;
• Cumulative: the running total, whose last point is your total usage.

Hover or focus any cell with the keyboard to see its date and usage.

Token by model

• Its own granularity switch: daily (last 7 days), weekly (last 4 weeks, the default) or monthly (last 6 months).
• Tokens draws stacked columns; Share draws a pie.
• Different spellings of one model collapse into a family: deepseek-v4-1-flash-261017 and deepseek-v4.1-flash are both DeepSeek V4.1 Flash, while version and tier stay apart, so v4-flash and v4.1-flash never mix. Other vendors keep their own family names, e.g. gemini-2-5-pro → Gemini 2.5 Pro.
• The top six families get their own colour; the rest are merged into "Other models (N)".

Token stats

Where the tokens went: a ring chart with the total in the centre and eight content categories beside it — system prompt, tool definitions, user messages, injected context, skill injections, tool calls, reasoning and answers — each with its share and an estimated amount.

Providers report four usage numbers per step (uncached input, cache read, cache write, output), not what the content was. The plugin replays each session log, estimates from content sizes how the step's context was made up, and splits the step's real usage in that proportion. The totals are exact; the split is an estimate, so amounts carry a ≈. The method follows the dsh-context plugin.

Projects

Tokens, share and spend per working directory. The subagent column is the part used by subagent sessions — the lighter segment of each bar. The top eight projects get a row and the rest are merged into "Other projects (N)"; hover a name for the full path.

Provider split

Tokens, share, that provider's own cache hit rate and spend per billing route (the provider id in the logs), for the header range. Every provider gets its own row.

The card itself is read-only. The button at its top right opens Provider settings, a page that lists every provider you have used (by all-time usage) with:

• Display name: a name that stays on your machine (hovering a renamed row on the card shows the original id);
• Billing method: the multiplier, plan and what this plan month has used so far are shown here, not on the card — click it to change it (see the next section);
• Remove: a route you no longer use is left out of the whole page, as if it had never been used (totals, spend, activity, models, token stats, projects). The page rescans once, in a few seconds. Removed routes are listed at the bottom of the settings page and can be brought back.

How spend is worked out

Spend takes two steps: every call is first valued at the list price, then each provider's billing method turns that value into what you actually paid.

Step 1: value at list price. Each call is priced at the list price in force when it was made.

• DeepSeek: built-in CNY list prices, including the 2026-08-17 prices and the 2026-09-10 Flash cut. Only providers billed at List price (DeepSeek direct) pay peak rates — Beijing weekdays 9–12 and 14–18, double the off-peak price; weekends and the 2026 statutory holidays are off-peak.
• Every other billing method is valued at the flat peak price at any hour: relays and coding plans charge one price all day, and relays quote their discounts against the peak price.
• Other vendors' models: a bundled LiteLLM price table (models sold directly by their vendors only), converted from USD to CNY.
• Unpriced: usage before 2026-08-17 and models without a known list price. The spend card's hover says when there is any.

Step 2: apply the billing method. Settings are stored only on your machine.

Billing method	Spend	Typical use
List price	The list-price value	Direct official API; deepseek-official and deepseek-account default to this
Discount	List-price value × your multiplier	Relays that bill a fraction of the list price
Monthly	Usage converted at list price; the settings page shows usage since the renewal day against the fee	Coding plans and other subscriptions
Not mine	Left out of spend	Paid by your company, free quotas
Not set	Left out of spend; the cell reads "—"	Every other provider until you set one

The USD → CNY rate defaults to 6.7168; write usdCny at the top level of billing.json to change it.

FAQ

Why don't the numbers match my provider's console?

Three usual reasons:

1. Scope: only this machine's DSH sessions are counted; other devices or tools using the same key are not.
2. Double-counted cache: many relay consoles show input + cache hit + output as the total, but their input already includes the cache hits, so the cache is counted twice.
3. Day boundaries: days here are cut in this machine's time zone; a console may use another.

Compare one billing route at a time using its row in the provider split, and compare token counts before percentages. Measured numbers and the full checklist are in docs/INTERNALS.md.

How is the cache hit rate calculated?

Cache read ÷ (uncached input + cache read + cache write). A cache write is billed and is not a hit, so it is in the denominator only; without writes this equals DeepSeek's hit ÷ (hit + miss). None of DeepSeek, OpenAI, Anthropic or Volcengine Ark defines a hit-rate formula in its API documentation; this one follows Langfuse.

Can I change the name and avatar on the page?

By default the page shows your operating-system account name. To use something else, create $DSH_HOME/dsh-profile-stats/profile.json:

{ "name": "Your Name", "handle": "@your-handle", "avatar": "Y" }

handle gets a leading @ if you leave it out, and avatar defaults to the first letter of name. Refresh the page to see it; no restart needed.

Why do the numbers keep moving a little?

Sessions that are still running keep adding usage; two scans a few minutes apart usually differ by less than 1%. A session log that cannot be read is skipped and counted, never filled in by a guess.

The panel disappeared after a failed update. What now?

Reopen Plugins and install dsh-profile-stats again. The cache and your settings live in $DSH_HOME, outside the package, so nothing is lost.

Privacy and local files

[!NOTE] Everything is read from your own session logs through ctx.sessionQuery. Token counts are the providers' own figures as logged; only the token-stats categories (split by content size) and the spend (priced at list prices and your billing settings) are estimates. The only outbound request is a six-hourly read of this package's own version from the npm registry, used by the in-panel update button. No telemetry, no credentials, nothing uploaded.

The plugin writes only these files, all outside the package:

File (default folder ~/.dsh/dsh-profile-stats/)	Contents
cache.json	Scan cache, so later opens don't rescan every log
billing.json	The billing method you set for each provider
aliases.json	The names you gave providers, and the ones you removed
profile.json	Optional: your own display name and avatar
update.json	The last in-panel update result
timing.log	Scan timings; rotated to timing.log.1 past 256 KB

Requirements

Requirement	Value
dsh	>= 0.2.1-rc.5 < 0.3.0 (verified on 0.2.1-rc.5)
Other plugins	none
Runtime dependencies	none

Uninstall

Disable or remove the bundle in Plugins. To remove its data as well, delete the $DSH_HOME/dsh-profile-stats/ folder.

Development

• Local preview: node tools/preview.mjs, or double-click preview/index.html. It renders the repo's own client.js against built-in sample data, with switches for light / dark theme, window width, language and the empty, scanning and error states.
• Tests: npm test. No dependencies, about a second; CI runs it under UTC, Asia/Shanghai and America/Los_Angeles.
• Releases are batched by GitHub Actions every four hours; never bump the version by hand.

More in CONTRIBUTING.md; metric definitions, the data flow, snapshot fields and troubleshooting are in docs/INTERNALS.md.

License

MIT © 2026 Lovenndme. The bundled price table comes from LiteLLM (MIT) and the token-stats estimate follows dsh-context; see THIRD_PARTY_NOTICES.md.