dsh-token-stats

If you have been running DeepSeek Harness for a while, you probably noticed that per-call token costs, model-level burn rates, and cache hit rates are scattered across logs with no single place to look. dsh-token-stats fills that gap: it intercepts every llm/stream call, records the real usage figures (input, output, cache read/write, and reasoning tokens) rather than estimates, and appends them to a durable local JSONL ledger that survives restarts and hot reloads.

On top of that ledger, the plugin aggregates across four dimensions—provider, model, session, and date—and pairs the numbers with cost estimates driven by OpenRouter live pricing. It also tracks cache hit rate, success rate, and time-to-first-token (TTFT). A draggable floating button in the bottom-right corner lets you toggle between global and current-session views at a glance; clicking it open reveals a full-screen dashboard with KPI cards, a request-share donut chart, a 31-day trend line, and per-model breakdowns. Models can also query the stats directly through the token_stats tool, and a backfill script can import calls made before the plugin was installed from historical session files.

If you are juggling several providers, trying to figure out which pipeline is eating your budget, or simply want a clear receipt for every conversation, this plugin fits the bill nicely. It is pure JavaScript with no build step—drop it into the packages directory, restart, and it is ready to go.