dsh-scope

English · 简体中文

Usage statistics for DeepSeek Harness (dsh) — a lifetime dashboard of your token usage across every session, plus a context lens for the session you are in.

DeepSeek Harness computes rich token accounting and then shows you none of it. dsh-scope surfaces it in two places.

One of several dsh plugins developed under dsh-plugins/ — each an independent repo with its own release cadence.

Features

Usage statistics (sidebar footer)

Open from the sidebar foot — a label when the sidebar is wide, an icon on the rail.

• Lifetime summary — cumulative and peak-day tokens, longest session, current and longest streaks.
• Token activity — a 52-week heatmap in three modes: daily (each day colored on its own usage), weekly (each column summed) and cumulative (a running total across columns). One grid reads either as a year at a glance or as growth over time.
• Time range — last 7 or 30 days, driving both charts below.
• Daily token trend — one line per model, with a wrapping legend and per-day tooltip. Only models used inside the window appear as series.
• Model usage — a donut with the window total in the center and a percentage list beside it.

The whole panel is one request to the plugin's own loopback-only endpoint.

Context Lens (session header)

Window occupancy for the current session: a live occupancy ring in the header that opens into a segmented system / tools / messages composition bar, per-bucket session totals, and KV-cache hit rate. It reads the official token-meter session projections through useProjection — no RPC, no custom wire protocol.

Requirements

• dsh ≥ 0.1.0-rc.6 on your PATH (verified against 0.1.0-rc.6 and 0.2.0-rc.2)
• Node.js 22.18+ or 24.11+ (for building from source only)

Install

dsh plugin --profile <profile> add dsh-scope
dsh web                             # restart dsh to load the plugin

<profile> is the directory under ~/.dsh/profiles/ that your dsh runs from — the Desktop app uses desktop.

Install from source

git clone https://github.com/helloxkk/dsh-scope.git
cd dsh-scope
npm install && npm run build
node scripts/install.mjs <profile>   # e.g. desktop, headless, rescue
dsh web                              # restart dsh to load the plugin

The script refuses an unknown profile rather than guessing.

The installer copies lib/, cordis.patch.yml, and package.json into the profile's node_modules and appends the bundle insert to the profile's cordis.patch.yml. It is idempotent — safe to re-run after every rebuild.

Uninstall

rm -rf ~/.dsh/profiles/<profile>/node_modules/dsh-scope
# then remove the "# dsh-scope" block from ~/.dsh/profiles/<profile>/cordis.patch.yml
rm -f ~/.dsh/storages/dsh-scope-cache.json   # optional: drop the fold cache

How it works

Host half (lib/index.js) — one read-only, loopback-fenced endpoint, GET /api/dsh-scope/days.

Aggregation is incremental. Per-session fold state is cached in memory and persisted to ~/.dsh/storages/dsh-scope-cache.json, and each request folds only the events added since the last fold. Live sessions fold their in-memory tail; persisted sessions skip work entirely while the backend's opaque revision is unchanged, and otherwise read only the new suffix, with contiguity checks and a full refold when a log was truncated or rewritten. Steady-state cost stays O(new events) no matter how large the logs grow.

Fold semantics mirror dsh-token-meter's tokenUsage projection. A usage sample rides an assistant/chunk (data.chunk.type === "usage") or an assistant/message (data.usage). A repeated sample for the same (turn, step) replaces the earlier value rather than double counting, re-attributed to the later event's day and model. Model attribution follows assistant/message's data.message.source, falling back to the last request/header config.

Beyond token buckets, the fold counts turn/start and tool/call events per day (feeding the weekly and cumulative heatmap tooltips) and tracks each session's first and last event timestamps so the summary can report the longest session. The payload carries every model per day rather than a truncated top-N, so the trend chart and donut always cover all of them.

Client half (lib/client.js, loaded via the package's dsh.client declaration):

• sidebar.footer.action renders the usage trigger. The panel fetches /api/dsh-scope/days same-origin and derives every view from that single payload.
• conversation.session.header.actions (order 30, after the job list) renders the lens trigger. Its popover reads three official projections — tokenUsage (four buckets accumulated over the whole durable log), contextPressure (input-side window pressure plus the route's context window), and contextBreakdown (heuristic system/tools/message composition of the next request).

Nothing is sent anywhere. The endpoint refuses non-loopback callers and non-GET methods before doing any work, and no provider credentials are read.

Development

npm install
npm run build                          # tsdown: host ESM + browser ModuleLoader bundle
node scripts/fixture-session.mjs       # optional: write a demo session log with usage events
node scripts/install.mjs <profile>     # then restart dsh

npm run build emits two artifacts:

• lib/index.js — the Node host half. @deepseek-ai/* stays external and resolves against the profile's own installation at runtime.
• lib/client.js — the browser half, wrapped in the harness ModuleLoader factory format. Platform modules (react, @deepseek-ai/dsh-client-*) stay external and resolve through the loader's injected require; everything else, including Recharts, is inlined.

Because the harness caches plugin modules at startup, only the client half hot-reloads. After changing anything under src/index.js or src/usage.js, restart dsh.

Compatibility

Built and verified against dsh 0.1.0-rc.6 and 0.2.0-rc.2 (@deepseek-ai/dsh-base / dsh-web-app bundles). The harness is in developer preview and iterates quickly — expect compatibility-breaking changes.

The host half reads persisted logs through whichever API the installed sessionPersistence exposes: open(id, 'read') plus handle.read(offset) on 0.2.x, and the older readFrom(id, fromSeq) elsewhere. list() is normalized to { header, revision } snapshots, so both shapes fold identically.

dsh imports each plugin module once at startup, so a profile config reload cannot replace host-half code: restart dsh after rebuilding.

Credits

The usage panel's layout, its three heatmap modes, and the month-ruler alignment are derived from ZCode's 「应用用量」 surface (zai-org/ZCode, Apache-2.0). The rendering, data source and theming are this plugin's own. See NOTICE.md for the required attribution and the list of derived files.

Charts render with Recharts (MIT).

License

MIT