dsh-metrics-panel

When running LLM workloads in DeepSeek Harness, token spend, cache hit rates, and real costs tend to scatter across event streams and console logs, leaving no single place to see the full picture at a glance. dsh-metrics-panel solves exactly that: it injects a floating panel into the DSH web interface, captures every API call s usage, parameters, and results from the session event stream in real time, and turns the numbers buried in logs into interactive overviews, paginated request details, and aggregate charts.

The panel offers ten sections covering cost and token distribution by hour, a paginated request list (with service endpoint, sampling parameters, tool calls, and a redacted HTTP example), an error log with error rates, and usage/cost aggregates broken down by model, provider, and tool. Pricing supports three tiers -- cache hit, non-hit input, and output -- with per-provider peak/off-peak windows for automatic tier switching, and ships with DeepSeek s official list prices as defaults. Theme can be set to light, dark, or system-following, and five chart palettes persist across sessions.

If you are driving multi-model, multi-provider Agent workloads in DeepSeek Harness and want to know how many tokens each turn consumed, how much the cache saved, which model is the most expensive, and which requests failed, this panel is the usage dashboard you have been looking for.