dsh-token-viewer

CC Switch-style token consumption statistics for the DeepSeek Harness Web GUI. Read-only surfaces over the harness's host-computed projections plus one balance read; the plugin adds no prompt content, tools, or provider requests.

> Install (one command): > `` > dsh plugin add qwert702/dsh-token-viewer > `` > Restart the harness, refresh the web page, then open the Token 消耗 card in the sidebar → 用量详情.

Features

• TokenDock — a slim live strip above the composer showing the current session's billed input (uncached + cache read + cache write), output, cache hit rate, and approximate context occupancy.
• Sidebar card — DeepSeek account balance (with refresh; error-retry when the host proxy fails) and aggregate consumption across all sessions, expandable to a per-conversation list.
• Usage statistics panel (right-side drawer, a faithful port of CC Switch's usage-dashboard method):

- Per-request statistics — the host usageLog projection records one timestamped entry per reported assistant step (commit time, model, four token buckets); every figure folds these records, never cumulative session totals. - Hero — real consumption (fresh input + output + cache write + cache read), request count, total cost, over a five-card breakdown row with a cache-hit-rate progress bar. - Trend chart — requests bucketed by their own commit time (hourly for the day, daily otherwise, empty buckets zero-filled), four token series plus a dashed cost line. - Three tabs — request log (newest first; clicking a row opens that session), per-project statistics, and per-model statistics with average cost. - Range presets — today / 7d / 14d / 30d / all, resolved exactly like CC Switch (local midnight of N−1 days back).

• Per-model peak/off-peak list pricing — every request bills under its own model's provider list price (V4-Flash / V4-Pro, CNY per 1M tokens, cache writes at the cache-miss rate), split by the provider's Beijing peak windows (09:00–12:00 and 14:00–18:00, double the off-peak rate); versioned model ids match by prefix, unknown models fall back to the V4-Flash off-peak table. Prices live in MODEL_PRICING (see below).
• Balance route — GET /api/billing/balance proxies DeepSeek's /user/balance through the harness credentials service; the API key never leaves the server.

Screenshots




Repo layout

• lib/index.js — plugin host half (balance route + modelUsage / usageLog session projections), ready to load.
• lib/client.js — browser half bundle (built), discovered via package.json dsh.client.
• scripts/build-client.mjs — regenerates lib/client.js by vendoring the installed @deepseek-ai/dsh-client-ui-sidebar bundle (set DSH_SIDEBAR_BUNDLE or it probes ~/.dsh/profiles).
• test/smoke.cjs — node test/smoke.cjs: host route + projections + SSR render checks.

The TypeScript monorepo source (extracted from deepseek-ai/deepseek-harness) lives on the archive/monorepo-src branch.

Model pricing

MODEL_PRICING in lib/client.js (and lib/index.js's projection fallback) holds the current DeepSeek list prices; the panel also fetches GET /api/billing/pricing and prefers the provider's official page when reachable, falling back to the built-in table otherwise.

License

MIT