dsh-token-stats
English | 中文
The missing global token-usage dashboard for deepseek-harness (dsh) — one-line install, then open Settings to see per-model, per-time-bucket, and per-session consumption with anomaly tracking.
Why
dsh's built-in stats cover only the current session. Total consumption across sessions, per-model breakdowns, which day was heaviest, which sessions failed — none of that exists in the stock UI. This plugin adds all of it, and gets two easy-to-get-wrong things right:
- Correct attribution: subagent tokens and anomalies roll up along the
parentSessionchain to their root main session — counts speak in main sessions only, while every subagent token lands on its owner - Zero wait: a resident background aggregate (warm-up at start → 60s incremental sweeps → 10min full rescans) means the Settings page always reads hot data
Install
npx github:caiyfa/dsh-token-statsRestart dsh, open Settings → Token 用量. Done.
- Requires: dsh (web profile), Node ≥ 22
- Several profiles:
--profile=tui,web; every profile:--profile=all - Uninstall:
npx github:caiyfa/dsh-token-stats --uninstall - After upgrading dsh, re-run the installer once if the plugin fails to load — it repairs the links
What you get
| Section | Contents |
|---|---|
| Summary card | Uncached input / cache read / cache write / output / reasoning / total / model calls |
| Per-model table | Main-session count, calls, the four token buckets, and totals for each provider / model |
| Token trend | Granularity (minute·24h / hour·7d / day·all, auto-weekly beyond 90 days) × mode (bars / 4-color stacked composition — watch your cache-hit structure evolve / cumulative curve — the budget-burn view); dashed Y contour lines with a peak label, hover for per-bucket detail |
| Anomaly marks | A green/red dot above every time bucket (red = failed turns inside it); click a red dot to focus the main sessions involved |
| Per-session anomaly groups | Collapsible per-main-session list (time + error message), collapsed by default so volume never floods the page |
How it works (30-second version)
The host half is a resident background service: it scans every session log, folds each assistant/message's provider-reported usage (bucketed by provider/model) and each turn/end error into minute/hour/day timelines. The client half registers the Settings dashboard above and reads snapshots through dsh's /api gateway. Pure in-memory aggregation — nothing written to disk or session logs, nothing injected into model prompts.
The installer does three things: deploys the package into ~/.dsh; links (junction/symlink, never copies — Remote markers ride a module-scoped WeakMap, so the plugin must share the same physical module instance with dsh) @deepseek-ai/cordis and @deepseek-ai/dsh-typert-protocol from your live dsh install; and appends a marker-delimited idempotent block to the profile's cordis.patch.yml.
Troubleshooting
| Symptom | Fix |
|---|---|
Failed to load plugins ... dsh-token-stats |
Re-run the installer; if it still fails, DSH_MODULES_DIR=<dsh node_modules> npx github:caiyfa/dsh-token-stats |
| Section renders but stays empty | Check the dsh process log for token-stats: errors |
invocation-unavailable: tokenStats/snapshot |
A peer got copied as a real directory: delete it and re-run the installer |
License
MIT