dsh-strength-meter

A local token-usage statistics section for the DeepSeek Harness Settings panel: metric cards, per-model donut, a 15-week activity heatmap, and a daily token trend. Everything is folded from your local session logs through the current sessionQuery service; nothing is uploaded.

中文文档见 README.zh.md。

What changed from 1.x

DSH 0.2 replaced the old host session seams. Version 2 keeps the same UI and route path but reads the corpus through today's API:

The fold itself (lib/fold.js) is unchanged and still mirrors the durable log: request/header names the model, assistant/message.usage / assistant/chunk carry tokens, and user/assistant messages drive the activity calendar.

Install

Plugin Manager (Desktop)

Install this folder as a bundle and enable dsh-strength-meter.

CLI

dsh plugin --profile <profile> add /absolute/path/to/dsh-strength-meter

Manual

Link the package into the profile's node_modules, then add to cordis.patch.yml:

- insert:
    - id: dsh-strength-meter
      name: 'dsh-strength-meter'

Reload the page. No Harness file is patched.

Use

Open Settings → Usage stats (the section sits after the built-in sections). Pick Last 7 days / Last 30 days, or press the refresh button to bypass the 15 s response cache. The first read folds the whole corpus with a concurrency of 4; subsequent reads reuse folds whose persistence revision is unchanged.

Route

GET /plugins/dsh-strength-meter/usage.json?days=7|30&force=1

The payload contains total, models[], sessions, messages, activeDays, streak, topModel, byDay[], heatmap[][], and a debug block with corpus/folded/cached/errors.

Verify

npm test
# → fold tests, host smoke test, bundle smoke test

Live check: open the stats section; the metric cards must show non-zero totals for a profile with history, and the heatmap must render 15 columns.

Limits

• Response cache — 15 seconds unless force=1.
• Bounded corpus — at most 400 sessions per request.
• No projection dependency — the plugin deliberately folds raw logs instead of registering a sessionStats projection, so it works in any composition.

License

MIT.