dsh-consumption-badge

A persistent DSH (DeepSeek Harness) Web UI plugin that puts a live consumption badge next to every assistant message, and opens a detail panel with the session's spend curve. The UI is bilingual (follows the shell language, switches live), costs are estimated for 33 model entries across 10 providers, amounts can be shown in any of 17 major currencies, and one dropdown picks the model you want to look at. All of it lives in Settings → Usage.

⛽ 3.2k | ≈¥0.004 | Off-peak | Balance = ¥42.50

• Collapsed pill – that assistant message's own token count, its own estimated cost (always prefixed ≈), its pricing tier (red Peak, green Off-peak, attributed to the turn's request start), and the DeepSeek account balance; injected into conversation.chat.assistant-actions right after the shipped copy/branch controls. The two numbers are this turn's, not the session's: the slot renders once per turn (TurnTailNodeView) and hands the badge that turn's final messageId; the host answers with a messageId → message map (which carries the turn number) next to the per-turn fold, and the badge renders that turn's folded total. Mapping is resolved in two steps so the pill is right even before the host half is restarted: the host map first, then the data-turn-tail attribute the shipped conversation view puts on the turn wrapper. If neither resolves (or the turn has no record yet) the pill falls back to the session-wide figures instead of showing a false zero, and its tooltip names the scope being shown. The balance segment renders Balance = ¥42.50 when the account API answers, … while loading and — on failure (the reason rides the tooltip); balance uses the same 60 s cache as the modal, refreshed once a minute. The modal itself still shows the whole session, unchanged.
• Detail modal (centered, blurred backdrop, Esc or backdrop-click to close): cumulative tokens with hit/miss/output split, estimated cost, current pricing tier, DeepSeek account balance, a model dropdown in the title bar, and a per-turn dual-axis line chart (left axis: tokens, right axis: money) where each model gets its own colored line and the legend names them; the hover tooltip shows the model (with its color dot), input-hit / input-miss / output tokens, the turn's cost and its peak/off-peak flag. Picking a model filters the cards + chart to that model and turns the balance hint into "about N tokens left at these prices".
• Settings page · Usage (settings.section, id balance-dsh-consumption, order 26): a top row of four cards — Total tokens (with hit/miss/output split, all-history totals), Estimated cost, Current window (peak/off-peak) and Account balance (reusing the badge's cached balance) — then all-session history as one line chart, filterable by time range (Today / Last 7 days / Last 30 days / Last 12 months), by workspace (All workspaces + every registered workspace + Unassigned), and by an explicit calendar range that overrides the dropdown. Bucket size follows the range: hourly for today and any single-day calendar range, daily for multi-day ranges, and monthly for the last 12 months. Left axis (tokens) is hidden until you toggle it in the legend; the right axis (¥) is always on. Hovering a point shows its total, its peak/off-peak split and the models involved; the footer sums the current selection. No data in range → a plain "No usage recorded in this range" note instead of an empty chart. The cards are all-history totals (independent of the filters); the footer is the filtered total, so the two numbers differ by design.
• Personalization integration (auto-detected, no configuration): when dsh-plugin-personalization is installed alongside this plugin, a Usage popup block is appended to its settings page — same pp-block / pp-field / pp-modes layout and the same Pill primitive as the shipped personalization rows — offering four backgrounds for the consumption modal: Default / Glass / Custom image / Solid color.
  • Default keeps the themed surface; Glass is real backdrop-filter glass — both the veil behind the modal and the panel itself are theme-derived (--dsw-alias-bg-base), so light themes get a light haze and dark themes a dark one instead of the panel painting a dark slab over the page; Custom image takes an image URL or a local image plus an Opacity slider (0–100 %, default 85 %) that sets the background layer's own opacity — the same semantic as the other personalization background rows, so lowering it fades the image over the themed panel instead of dimming it; Solid color takes a hex colour and derives a light/dark ink palette so the cards, chart and tooltips stay readable (the palette also overrides the tooltip surface).
  • A Glass switch appears for Custom image and Solid color, so glass composites with them: on an image it blurs the image itself (frosted photo), on a solid it makes the colour translucent over a blurred page. A Glass strength slider (0–48 px, default 22 px) sets the blur radius of the panel. The overlay behind the modal is deliberately never blurred.
  • Clear wipes the image, the URL field and the file picker in one go. The image row shows a thumbnail plus the live save state (Saving… / Saved on this device / error: …), and failures are logged with a [consumption-badge] tag instead of failing silently. The picker is a button (Choose image… / Change image…), not a raw file input, so saving never leaves a misleading "no file chosen" label behind.
  • Local images are stored on this machine with no size limit: the picked file is written as a Blob into IndexedDB (dsh-consumption-assets → images), and the preference only keeps a small key — so a 20 MB wallpaper no longer blows the localStorage quota. On startup the blob is read back into a blob: URL for the modal; Clear deletes the stored record as well. If IndexedDB is unavailable the file is read as a data URL instead, and the row says so.
  • The choice is a local preference (dsh-consumption:modal-style) and is applied to the open modal immediately. Nothing about the conversation or sidebar backgrounds is touched.
• Update notice (same shape as dsh-plugin-personalization's): ~2.5 s after the page loads the plugin asks https://registry.npmjs.org/dsh-consumption-badge/latest for the newest version (24 h cache, 5 s timeout, fails silently). If the installed version is older — and that version was not ignored — a non-blocking card appears in the bottom-right corner with current → latest, the exact dsh plugin --profile web update dsh-consumption-badge command, and three actions: Copy update command, Ignore this version (remembered on this machine) and Remind me later (this page load only). It never blocks the conversation and never covers controls.

Everything is an estimate for reference only — this plugin is not a billing system of record.

Models and pricing

Costs are estimated from a flat price table of 31 model entries (current generations first, legacy entries last so old logs still resolve). Unknown ids fall back to a family guess (pro/opus/max → the stronger tier, everything else → DeepSeek V3) and are marked as estimates.

Provider	Current	Legacy (kept for old logs)
DeepSeek	deepseek-flash = DeepSeek-V4.1-Flash-1017, deepseek-v4-pro = DeepSeek-V4-Pro-0813 (official ids/versions)	DeepSeek V3.2
OpenAI	GPT-5.6 Sol, GPT-5.6 Terra, GPT-5.6 Luna, o3	GPT-4o, GPT-4o mini
Anthropic	Claude Opus 5.5, Sonnet 5.5, Haiku 5.5	Claude Sonnet 4.6, Opus 4.x
Google	Gemini 3.8 Flash, Gemini 3.1 Pro, Gemini 3 Flash	Gemini 2.5 Flash
Alibaba	Qwen3.8 Max, Qwen3.8 Flash	Qwen3.7 Plus
Zhipu	GLM-5.3, GLM-5.1	GLM-4.6
Moonshot	Kimi K3.1 Pro, Kimi K3	Kimi K2.5
xAI	Grok 4.7	Grok 4.3
Meta	Llama 4 Maverick	—
Mistral	Mistral Large 4	—

Rules that follow from the table:

• Every turn is priced with the model it actually used — a session that mixed models is not priced with a single rate. Sessions, buckets and charts all keep the per-model split.
• Only DeepSeek has peak/off-peak pricing (×2 in the Beijing-time peak windows). Other vendors are flat; their card hint says so.
• Prices are CNY per 1M tokens. Non-DeepSeek entries are the vendor's USD list price × 7.1 (input-cache-read / input / output), taken from OpenRouter's public model catalogue; the two harness-native DeepSeek rows keep DeepSeek's official CNY table. Edit MODELS at the top of lib/client.js — each entry carries price {hit, miss, out}, peakPricing, legacy and a chart color.
• Model ids are resolved transparently. deepseek-flash is the id DeepSeek's API currently routes to DeepSeek-V4.1-Flash-1017 and deepseek-v4-pro to DeepSeek-V4-Pro-0813 (per api-docs.deepseek.com), and deprecated ids (deepseek-v4-flash*, deepseek-chat) route to the same families. Every model label in the UI can therefore be checked against the log: the settings cost card, the modal cost card, the legend tooltip and the per-turn tooltip all print Family name (Official MODEL VERSION · log id <id>), so a mismatch is visible instead of hidden.
• The Model dropdown (panel title bar and Settings → Usage, sharing one preference stored in localStorage: dsh-consumption:model-filter) lists models grouped by provider in <optgroup>s; it filters the statistics and acts as the pricing fallback for turns whose log entry carries no model. All models turns both behaviours off.
• Charts draw one line per model in that model's color (33 distinct colors), with a legend of model chips; the token axis stays a single dashed neutral line.

Currency

The price table is always CNY per 1M tokens (that is what DeepSeek bills in); the currency setting only decides how amounts are displayed. Pick it in Settings → Usage → Currency:

Aspect	Behavior
Currencies	CNY (default) plus USD, EUR, GBP, JPY, HKD, TWD, KRW, SGD, AUD, CAD, CHF, INR, RUB, BRL, THB, MXN
What converts	every amount in the plugin: pill (`≈# dsh-consumption-badge

A persistent DSH (DeepSeek Harness) Web UI plugin that puts a live consumption badge next to every assistant message, and opens a detail panel with the session's spend curve. The UI is bilingual (follows the shell language, switches live) and amounts can be shown in any of 17 major currencies — pick one in Settings → Usage.

⛽ 3.2k | ≈¥0.004 | Off-peak | Balance = ¥42.50

• Collapsed pill – that assistant message's own token count, its own estimated cost (always prefixed ≈), its pricing tier (red Peak, green Off-peak, attributed to the turn's request start), and the DeepSeek account balance; injected into conversation.chat.assistant-actions right after the shipped copy/branch controls. The two numbers are this turn's, not the session's: the slot renders once per turn (TurnTailNodeView) and hands the badge that turn's final messageId; the host answers with a messageId → message map (which carries the turn number) next to the per-turn fold, and the badge renders that turn's folded total. Mapping is resolved in two steps so the pill is right even before the host half is restarted: the host map first, then the data-turn-tail attribute the shipped conversation view puts on the turn wrapper. If neither resolves (or the turn has no record yet) the pill falls back to the session-wide figures instead of showing a false zero, and its tooltip names the scope being shown. The balance segment renders Balance = ¥42.50 when the account API answers, … while loading and — on failure (the reason rides the tooltip); balance uses the same 60 s cache as the modal, refreshed once a minute. The modal itself still shows the whole session, unchanged.
• Detail modal (centered, blurred backdrop, Esc or backdrop-click to close): cumulative tokens with hit/miss/output split, estimated cost, current pricing tier, DeepSeek account balance, and a per-turn dual-axis line chart (left axis: tokens, right axis: ¥) with a hover tooltip showing the model, input-hit / input-miss / output tokens, and the turn's peak/off-peak flag.
• Settings page · Usage (settings.section, id balance-dsh-consumption, order 26): a top row of four cards — Total tokens (with hit/miss/output split, all-history totals), Estimated cost, Current window (peak/off-peak) and Account balance (reusing the badge's cached balance) — then all-session history as one line chart, filterable by time range (Today / Last 7 days / Last 30 days / Last 12 months), by workspace (All workspaces + every registered workspace + Unassigned), and by an explicit calendar range that overrides the dropdown. Bucket size follows the range: hourly for today and any single-day calendar range, daily for multi-day ranges, and monthly for the last 12 months. Left axis (tokens) is hidden until you toggle it in the legend; the right axis (¥) is always on. Hovering a point shows its total, its peak/off-peak split and the models involved; the footer sums the current selection. No data in range → a plain "No usage recorded in this range" note instead of an empty chart. The cards are all-history totals (independent of the filters); the footer is the filtered total, so the two numbers differ by design.
• Personalization integration (auto-detected, no configuration): when dsh-plugin-personalization is installed alongside this plugin, a Usage popup block is appended to its settings page — same pp-block / pp-field / pp-modes layout and the same Pill primitive as the shipped personalization rows — offering four backgrounds for the consumption modal: Default / Glass / Custom image / Solid color.
  • Default keeps the themed surface; Glass is real backdrop-filter glass — both the veil behind the modal and the panel itself are theme-derived (--dsw-alias-bg-base), so light themes get a light haze and dark themes a dark one instead of the panel painting a dark slab over the page; Custom image takes an image URL or a local image plus an Opacity slider (0–100 %, default 85 %) that sets the background layer's own opacity — the same semantic as the other personalization background rows, so lowering it fades the image over the themed panel instead of dimming it; Solid color takes a hex colour and derives a light/dark ink palette so the cards, chart and tooltips stay readable (the palette also overrides the tooltip surface).
  • A Glass switch appears for Custom image and Solid color, so glass composites with them: on an image it blurs the image itself (frosted photo), on a solid it makes the colour translucent over a blurred page. A Glass strength slider (0–48 px, default 22 px) sets the blur radius of the panel. The overlay behind the modal is deliberately never blurred.
  • Clear wipes the image, the URL field and the file picker in one go. The image row shows a thumbnail plus the live save state (Saving… / Saved on this device / error: …), and failures are logged with a [consumption-badge] tag instead of failing silently. The picker is a button (Choose image… / Change image…), not a raw file input, so saving never leaves a misleading "no file chosen" label behind.
  • Local images are stored on this machine with no size limit: the picked file is written as a Blob into IndexedDB (dsh-consumption-assets → images), and the preference only keeps a small key — so a 20 MB wallpaper no longer blows the localStorage quota. On startup the blob is read back into a blob: URL for the modal; Clear deletes the stored record as well. If IndexedDB is unavailable the file is read as a data URL instead, and the row says so.
  • The choice is a local preference (dsh-consumption:modal-style) and is applied to the open modal immediately. Nothing about the conversation or sidebar backgrounds is touched.
• Update notice (same shape as dsh-plugin-personalization's): ~2.5 s after the page loads the plugin asks https://registry.npmjs.org/dsh-consumption-badge/latest for the newest version (24 h cache, 5 s timeout, fails silently). If the installed version is older — and that version was not ignored — a non-blocking card appears in the bottom-right corner with current → latest, the exact dsh plugin --profile web update dsh-consumption-badge command, and three actions: Copy update command, Ignore this version (remembered on this machine) and Remind me later (this page load only). It never blocks the conversation and never covers controls.

Everything is an estimate for reference only — this plugin is not a billing system of record.

), modal cards and both charts' right axis, the settings cards and footer totals, and the account balance (when the balance API reports CNY) | | Rates | 1 CNY = n target currency, cached 6 h. Primary source open.er-api.com, fallback @fawazahmed0/currency-api via jsDelivr — fetched through the host route /api/consumption/rates, and directly from the browser when that route is unavailable (both endpoints send access-control-allow-origin: *) | | Offline | if no live rate can be fetched, a built-in fallback table is used and the settings hint says so | | Preference | localStorage key dsh-consumption:currency; the settings hint shows the active rate, its source and a timestamp | | Unchanged | the pricing table, peak/off-peak rules and every stored record stay in CNY — switching currency never rewrites data |

costCNY = (hit/1e6)*hit + (miss/1e6)*miss + (out/1e6)*out   [×2 at peak]
shown   = costCNY × rate[currency]

Localization

The plugin ships Chinese and English copy and follows the DSH shell language live — no reload, no second settings page.

Aspect	How it works
Source of truth	The client-side locale service (@deepseek-ai/dsh-client-locale, ctx.get('locale'))
Dictionary	DICT_EN in lib/client.js, keyed by the Chinese copy itself; under zh no dictionary is needed because the key is the text, and any key missing from DICT_EN falls back to that Chinese text (never a blank string)
Live switching	Slot outlets re-render on a locale switch (the shell subscribes to the locale revision), so tt() picks up the new language during the next render; the personalization page block renders in its own React root and subscribes to the locale face itself
Slot labels	The action-row label and the settings tab title are thunks, resolved by the slot system at read time, so the tab renames itself on a switch
Late service	If the locale service is provided after this plugin's apply() ran (boot order differs per composition), the first render retries the bind
Missing service	Without the locale service everything falls back to Chinese — the plugin still works standalone
Host errors	lib/index.js never returns localized prose: failures carry a language-neutral code + detail (and an English error for logs). The client maps code → copy in the active language, and prints the host's error only for an unknown code

Adding a language means adding one dictionary next to DICT_EN; the lookup order is active locale → key, so untranslated keys degrade to Chinese rather than breaking.

Requirements

Capability	Why
DSH 0.1.0-rc.7 or compatible	slot API conversation.chat.assistant-actions, shell.overlay
DEEPSEEK_API_KEY in the DSH credential store	account balance (GET /user/balance)
curl on PATH	the balance request; on Windows 10+ it ships as C:\Windows\System32\curl.exe

The API key is resolved per request through the Cordis credentials service and passed to curl on stdin (curl -K -), never through argv, and never reaches the browser.

If the key or curl is missing, only the balance card degrades (it shows the reason); tokens, cost and the chart keep working.

Install

dsh plugin --profile web add dsh-consumption-badge

Then restart the DSH process — a host-side plugin (and the browser plugin roster) is composed at process start, so a page refresh is not enough.

Verify:

cd "$DSH_HOME/profiles/web"
dsh --profile web --dump-config | grep consumption-badge     # expect exactly one row
# after restart:
curl -s http://127.0.0.1:<port>/api/consumption/account      # expect {"ok":true,...,"model":"..."}
curl -s -o /dev/null -w '%{http_code}\n' \
     http://127.0.0.1:<port>/plugins/dsh-consumption-badge/client.js   # expect 200

If the row does not mount

Some deployments compose dsh.profile.bundles into --dump-config without mounting those rows at runtime. The reliable path is a roster row in the profile's own patch layer — add it to $DSH_HOME/profiles/web/cordis.patch.yml and remove dsh-consumption-badge from dsh.profile.bundles so the id is inserted exactly once:

- insert:
    - id: consumption-badge
      name: 'dsh-consumption-badge'

The package also ships its own cordis.patch.yml (dsh.bundle.patch) with the same row for deployments where the bundle layer does mount. Keep only one of the two active.

What it reads

Data	Source
Per-turn tokens	assistant/message.usage events in the session log, folded per turn (cacheReadTokens → hit, inputTokens + cacheWriteTokens → miss, outputTokens → output; multi-step turns are summed)
Turn timing	turn/start event time (peak/off-peak is attributed to request start)
Model per turn	request/context / request/header.config; falls back to the agent default model, then deepseek-flash
Session totals	the shipped tokenUsage projection when present, otherwise the folded turns
Balance	GET https://api.deepseek.com/user/balance via curl, 60s cache

Turn records are also mirrored into browser localStorage under dsh-consumption:<sessionId>:<turn> so the chart survives reloads. Persistence writes are scheduled off the render path (requestIdleCallback), and the aggregation is O(turns).

The settings page reads those same records (no second store, no new field): workspace attribution is resolved at read time from DSH's own accounting ledger (WorkspaceView.sessionIds from useWorkspaces), so historical records are attributed without any backfill. Records whose session no longer exists land in the Unassigned bucket. Above 5000 records the aggregation runs in a Web Worker (built from the pure aggregator's own source, so both paths are the same code); if the browser refuses the Worker it silently falls back to the main thread, where 6000 records aggregate in ~6 ms.

Pricing model

Off-peak prices in CNY per 1M tokens; peak multiplies everything by 2.

Model	input cache hit	input cache miss	output
deepseek-flash	0.02	1	4
deepseek-v4-pro	0.15	4.5	13.5

cost = (hit/1e6)*hit_price + (miss/1e6)*miss_price + (out/1e6)*out_price   [×2 at peak]

Peak = Beijing time (UTC+8), Monday–Friday, 09:00–12:00 and 14:00–18:00, excluding statutory holidays. Everything else, plus weekends, is off-peak.

Unrecognised model ids are priced as deepseek-flash (or deepseek-v4-pro when the id contains pro) and the UI marks the figure as an estimate. The table lives in PRICES at the top of lib/client.js — fork and edit to follow a price change.

The fixed-date holiday list (CN_FIXED_HOLIDAYS, MM-DD) covers New Year, Labour Day and the first three days of National Day. Lunar holidays (Spring Festival, Qingming, Dragon Boat, Mid-Autumn) must be added by hand.

HTTP surface

The host half registers three exact, read-only routes on the DSH web server:

Route	Returns
GET /api/consumption/turns?sessionId=<id>	per-turn {turn, hit, miss, out, time, model} plus per-message {id, turn, step, hit, miss, out, time, model} (the id is the finalized message id the badge receives as messageId); log-length cached
GET /api/consumption/account	current provider/model + whether the key is configured
GET /api/consumption/rates[?force=1]	{ok, base:"CNY", rates:{USD:…}, source, fetchedAt, stale} — 16 curated currencies, 6 h cache; stale:true means the built-in fallback table
GET /api/consumption/balance[?force=1]	{ok, available, currency, total, granted, toppedUp}; is_available:false renders a red "insufficient balance"

Failures are {ok:false, code, detail, error}: code is language-neutral (no-credentials, no-key, key-read-failed, no-subprocess, curl-empty, spawn-failed, bad-response, balance-request-failed, missing-session-id, log-unavailable), detail is raw text, error is an English fallback for logs.

These routes are unauthenticated. If your DSH web server binds 0.0.0.0, anyone on the network can read the balance and token figures; keep it on 127.0.0.1 unless you need otherwise.

Prompt for your AI agent (copy-paste)

Paste this into Claude Code / Codex / DSH and let it do the install and verification:

Install the DSH plugin "dsh-consumption-badge" into my web profile and prove it works.

1. In $DSH_HOME/profiles/web run: `dsh plugin --profile web add dsh-consumption-badge`.
   (For a local checkout use `dsh plugin --profile web add "file:<abs path to the package>"`.)
2. Check `dsh --profile web --dump-config` contains exactly ONE `id: consumption-badge` row.
   `dsh plugin add` also appends the package to `dsh.profile.bundles`; if that row does not
   actually mount after the restart, move it: remove it from `dsh.profile.bundles` and add
   `- insert: [{id: consumption-badge, name: 'dsh-consumption-badge'}]` to the profile's
   `cordis.patch.yml` instead. Never leave both.
3. Restart the DSH process (a browser refresh is NOT enough — host plugins are composed at boot).
4. Verify the host half: `curl -s http://127.0.0.1:<port>/api/consumption/account` must return
   `{"ok":true,...}`. Verify the client half:
   `curl -s -o /dev/null -w '%{http_code}' http://127.0.0.1:<port>/plugins/dsh-consumption-badge/client.js`
   must return `200`.
5. Open any session and confirm the `⛽ <tokens> | ≈¥<cost> | Peak|Off-peak | Balance = ¥<total>` pill in the assistant
   action row; click it and confirm the modal shows the four cards, the dual-axis chart, the
   balance, and that Esc closes it.
6. Keep Settings → Usage open, then switch the shell language between Chinese and English
   (Settings → General → Language) and confirm every visible string in the badge, the modal, the
   settings page and the personalization block switches without a reload.
7. On the Usage page switch the time range (Today / Last 7 days / Last 30 days / Last 12 months),
   switch the workspace (All workspaces → a specific workspace → Unassigned), set the calendar range
   and confirm both the chart and the footer totals update, and that the Token axis only appears
   after clicking its legend chip.
8. If `dsh-plugin-personalization` is installed, open Settings → Personalization and confirm a
   Usage popup block appears at the bottom; switch it to Glass / Solid color / Custom image and
   confirm the ⛽ modal (click a badge) re-renders with that background.
9. Report which step failed and paste the boot log lines that mention `consumption-badge`.

Rules while doing this: never write to the DSH credentials file yourself (ask me for
DEEPSEEK_API_KEY), never bind a second server to the same port, and do not try to persist the
plugin by "running" it — a session-scoped dynamic plugin always dies with the process.

Notes for AI agents

• This is a profile-level plugin: package.json carries dsh.bundle.patch and dsh.client (platform: web). It cannot be shipped as a dynamic/session plugin — harness.handle and host.call are dynamic-runner-only and do not exist for a bundled plugin; the client half talks to the host half over the same-origin routes listed above.
• The client bundle is hand-written in window.__ModuleLoader__.load({id, factory}) form and only require("react") — React must stay external (a second React instance breaks slot rendering).
• lib/index.js imports nothing at all (no @deepseek-ai/cordis), so it does not depend on the profile being able to resolve Cordis; it exports plain name / inject / apply.
• Do not hard-code user-visible English or Chinese in new JSX: call tt('…') with the Chinese copy as the key and add the English entry to DICT_EN. check.mjs fails if any tt() key has no English translation or if a translation still contains Chinese.
• Self-check without booting DSH: node check.mjs (320 assertions over the route contract, the client bundle contract, the pricing/aggregation maths and the bilingual copy). It lives outside the published files list.

Known limitations

• All amounts are estimates — price table, peak/off-peak rules and model detection can drift from your actual invoice.
• One chart point per turn; multiple steps in a turn are summed.
• Balance parsing depends on DeepSeek's /user/balance response shape; failures only affect the balance card.
• The chart covers turns present in the session log fold (whole log), while the pill's total uses the shipped tokenUsage projection — both are whole-session, but a heavily-compacted session may differ by the tokens lost to compaction.
• Only zh and en are shipped; any other shell locale falls back to zh (the dictionary keys).

License

MIT