dsh-token-attention — Token Check for DeepSeek Harness

English | 中文说明

Token Check is a token-attention management panel for DeepSeek Harness (DSH). It records every conversation's token usage and estimated cost by task / day / week / month, recognizes task types, and — combined with DeepSeek's peak/off-peak pricing and DSH's context mechanics — tells you when to run a task, whether to switch sessions, and whether to write a hand-off file.

It is not a money-saving tool: it never throttles or cancels requests. It presents facts and advice; you decide.

Features

• Five views in DSH Settings — Overview / Sessions / Tasks / Cost / Settings
• Per-task ledger — token consumption (input hit / miss / output / reasoning) and estimated cost, rolled up by day / week / month
• Task-type recognition & advice engine — best execution window, switch-session, hand-off, context, cost and model suggestions
• Attention KPI — attention score, hit rate, context occupancy, compaction count per session
• Peak/off-peak pricing — DeepSeek off-peak half-price windows (peak 9:00–12:00 / 14:00–18:00), estimated cost breakdown by bucket
• Multi-model breakdown — cost and hit rate per model (e.g. deepseek-v4-flash / pro / grok)
• Ledger export — Markdown / CSV download, one-click hand-off template
• Editable settings — pricing table, peak hours, advice thresholds; new records are priced immediately
• Zero runtime dependencies — local SQLite via node:sqlite; no network calls

Screenshots

Overview — today's usage, cost, hit rate, attention score, model mix, advice card, and trend

Session card — per-session model breakdown and attention metrics

Tasks — task ledger with Markdown / CSV export

Cost — peak split, cost leverage breakdown, and model comparison

Settings — pricing table, peak hours, and advice thresholds

Install

Once published to npm:

dsh plugin --profile web add dsh-token-attention

Restart DSH, then open Settings → 词元管理 (Token Check). Data is stored at DSH_HOME/token-attention/token_records.db.

Development install (link)

npm install && npm run build

Then register the plugin in the web profile (%APPDATA%\dsh-desktop\harness\profiles\web\package.json):

{
  "dependencies": { "dsh-token-attention": "link:/path/to/dsh-token-attention" },
  "dsh": { "profile": { "bundles": ["...", "dsh-token-attention"] } }
}

Run pnpm install inside profiles\web, restart DSH, and verify:

curl http://127.0.0.1:<port>/token-attention/api/health   # → {"ok":true,...}

How it works

• Collection: the node half subscribes to DSH's session/event stream (usage, goal, subagent, compaction, turn) and writes normalized records to SQLite; a startup backfill replays existing sessions with an idempotent last_seq cursor.
• Aggregation: hourly incremental rollups into daily_agg, with a daily full rebuild for correction.
• Advice: src/shared/advice.ts evaluates context occupancy, turns, hit rate, cost and model mix against your thresholds and suggests concrete actions.

Configuration

Settings are editable in the panel and saved to SQLite:

…