dsh-dashboard

When running multi-turn inference in DeepSeek Harness, token burn, latency swings, and tool-call patterns are buried in logs and hard to glance at from the conversation UI. DSH Dashboard adds a Statistics tab right beside Chat and Trajectory, surfacing all of that as a live panel without switching terminals or grepping output.

The panel covers four layers: a token layer with input including cache reads, output tokens, and cache-hit rate; a timing layer with LLM time, tool time, first-token latency, and decode rate; a tool-activity layer that breaks down per-tool calls, failures, and average duration rendered as proportional four-color segments; and an insights layer summarizing tool success rate, think-versus-tool time split, tokens per turn, and the input-to-output ratio. Every number updates live as the session progresses, and the locale follows the host automatically in both English and Chinese. All figures are derived client-side from the conversation snapshot tool-result nodes and the built-in sessionStats and tokenUsage projections, so there is no extra remote call.

If you are developing or running inference on the DeepSeek Harness v0.1.2 line and want an at-a-glance view of per-session resource cost and performance without leaving the chat window, this plugin fills the dashboard gap.