dsh-usage-analytics
Personal Agent usage analytics & activity dashboard for the DeepSeek Harness (dsh) Web GUI.
中文文档 | 中文版
A usage statistics / activity dashboard for DeepSeek Harness. It adds a Usage entry at the bottom of the sidebar (next to Settings) that opens a full-screen dashboard aggregating your real Harness usage:
- Token totals (input / output / cache hits / reasoning), session activity, a GitHub-style contribution heatmap, a token trend chart, a token mix breakdown, a model share ranking, plus insights (streaks, peak day, most-used model/tool/skill, …).
- Data comes only from your local session event logs (
ctx.sessionPersistence) — no dsh core changes, no prompt content ever leaves your machine (only event metadata and numeric usage are aggregated).
Features
- Full dashboard — contribution heatmap, token trend (daily/weekly/hourly), token mix, model share, reasoning-effort / tool / skill / dynamic-plugin rankings, streaks and personal insights.
- Period filtering — every chart follows the selected range (
Today / 24h / 7d / 30d / 90d / All time), so the headline and the charts always agree.
- Model grouping by real name — models served through several providers (e.g.
aaa/…, acme-gateway/…) collapse into one row by their real model name; the provider list is kept as a hover tooltip, never shown inline.
- Ghost-session detection — forked sessions that were copied but never ran are excluded from every aggregate, so duplicated logs can never inflate your numbers.
- Provider-console accounting — the headline and heatmap use the raw convention (input incl. cache hits, matching what a billing console shows); the new-token figure is always available separately, so you can tell "new tokens" from "context re-reads".
- Local & private — loopback-only API, no telemetry, no uploads, no prompt content collected or persisted.
- Incremental & instant — per-session revision diffing + resumable folds; the dashboard serves the cached snapshot immediately and finishes scans in the background.
Understanding the numbers (important)
DeepSeek-style APIs count cache-hit prompt tokens as input in their consoles. In long-running sessions with large contexts, every tool call re-sends the whole conversation, so 99%+ of "input" can be cache reads — that is why a busy day can show billions of raw tokens while your actual new tokens are only tens of millions.
This plugin follows that convention deliberately (so the dashboard matches your provider console), but always separates the three numbers:
| Term | Meaning |
|---|
| Input | uncached (new) input tokens only |
| Cache hits | prompt tokens served from the provider cache (re-reads) |
| Output | generated tokens |
If the raw totals feel too large, look at the new-token figures (day cells, hour buckets, and the "Input" line) — that is the usage you intuitively "produced".
Install
From npm (recommended)
dsh plugin --profile <name> add dsh-usage-analytics
Then restart the web service (profile bundles load at startup).
Manual / local development
- Copy the package to
data/profiles/web/plugins/dsh-usage-analytics/ (pure JS, no build step).
- Make it resolvable from the web profile's
node_modules/@local/dsh-usage-analytics (a directory junction/symlink to step 1, or an actual copy — both work; note the two locations are not auto-synced, keep them in step).
- Add
dsh-usage-analytics to dsh.profile.bundles in data/profiles/web/package.json (plus a file: dependency if you use pnpm install).
- Restart the web service.
After the restart:
- The sidebar footer shows a Usage entry → opens the full-screen dashboard;
GET /api/dsh-usage-analytics/stats returns the aggregate JSON (?force=1 triggers a full rescan);- The browser bundle is served at
/plugins/@local/dsh-usage-analytics/client.js (per the profile bundle roster).
Usage
Click Usage in the sidebar footer. Use the period pills in the header to filter every chart (Today, 24h, 7 days, 30 days, 90 days, All time). Hover the heatmap cells for per-day details. The 刷新 button re-syncs from the session logs (first open after a cache-version bump rebuilds the aggregate in a few seconds).
Data & privacy
| Guarantee | How |
|---|
| Local only | Reads session logs through ctx.sessionPersistence; never writes into sessions; loopback-only HTTP routes with a same-origin fence |
| Metadata only | Consumes event types and numeric usage fields only — no user-authored prompt content is collected, persisted, or served |
| Failure-isolated | Every fold/listener is try/catch-contained; a failing analytics never affects the agent loop or the GUI (worst case: a stale cache served with stale: true) |
Architecture
Session events / sessions
└─> lib/aggregate.js pure-function aggregation core (foldEvent / mergeInto / computeInsights / streaks)
└─> lib/store.js incremental cache: revision diffing + readFrom(fromSeq) single-source fold + JSON persistence
└─> lib/index.js host plugin: /api/dsh-usage-analytics/stats route + background catch-up folding
└─> lib/client.js browser bundle: sidebar.footer.action + shell.overlay official slots
- Single-source fold (v3) — tokens are folded only from the persisted log via
readFrom(fromSeq); the watermark advances only from persisted reads, so double counting is impossible.
- Incremental —
sessionPersistence.listSnapshots() exposes per-session stat revisions (header-only reads); unchanged sessions are skipped entirely, changed ones re-fold only their tail.
- Cache — per-session fold results + revision watermarks are persisted to
<DSH_HOME>/usage-analytics/agg.json (atomic write); CACHE_VERSION bumps rebuild the cache automatically.
- Per-model per-day buckets (v9) — each model records daily new/raw/output/call totals so the model-share chart can follow the period selector.
- Ghost-fork detection (v6) — a forked session whose whole usage predates its own creation is treated as a copied seed and excluded.
Development
npm test # node --test: aggregation core + client bundle smoke (zero dependencies)
node scripts/verify-data.mjs # print what the dashboard would show from real session data
node scripts/smoke-host.mjs # end-to-end smoke: real persistence + plugin apply + route handler
Layout: lib/ (host + client), test/ (unit tests), scripts/ (dev verification tools, not published).
Known limitations
- No cost estimates — no local pricing table; a provider→price mapping extension point could be added later.
- Skill-level token attribution — only skill call counts are tracked; tokens cannot be attributed to a single skill.
- Session duration — approximated by the wall-clock span between the first and last event.
- Deleting a session does not immediately shrink the stats — the aggregate cache keeps folded results until a full rebuild (
CACHE_VERSION bump or clearing agg.json + restart). The dashboard itself already excludes ghost forks at snapshot time.
- Multi-window freshness — host aggregation is in-process; concurrent writers to the cache file are safe (atomic rename) but last write wins.
FAQ
Why does a single day show billions of tokens?
That is the provider-console raw convention: it includes cache-hit prompt tokens. In long sessions with ~100k–800k token contexts, every call re-reads most of the context. Your actual new tokens that day are usually two orders of magnitude smaller — see Understanding the numbers.
Are the numbers fabricated?
No — every figure is summed from assistant/message.usage events recorded in your own session logs, per API call. Nothing is estimated, extrapolated, or injected.
Why don't deleted conversations reduce the totals?
Deletion removes the log, but the analytics cache keeps the already-folded statistics until a full rebuild. This is a documented limitation (see above).
What are "ghost sessions"?
Session forks that were created but never ran — their entire log is a copied seed of a parent conversation. Counting them would double-count the parent's tokens, so they are excluded automatically.
License
Apache-2.0 — see LICENSE.
