dsh-live-perf-gauges

Text streaming in DeepSeek Harness Web gives you characters on screen but little feel for the inference happening underneath. dsh-live-perf-gauges drops an automotive-dashboard-style live performance strip right below the composer so you can read the reasoning rhythm and spot bottlenecks at a glance.

It tracks five dimensions in real time: decode tok/s (refreshed every 250 ms, preferring the provider cumulative usage an adapter exposes), TTFT (elapsed time from step start to the first streamed text, reasoning, or tool-argument delta), logical prompt rate (input plus cache read and write divided by end-to-end TTFT, deliberately not labelled vLLM kernel prefill throughput because it includes DSH overhead, network, and queue time), stream-event diagnostics (per-second delta events and generated tokens per event, making MTP and speculative-decoding acceptance visible), and output-token progress. Every metric is projected purely from timestamps, counts, and final token usage already present in the session event stream; the plugin makes no network calls and never stores message text, prompts, tools, or credentials.

If you run the DSH 0.1.x Web profile and care about inference latency, speculative-decoding efficiency, or want quantified feedback while tuning prompt structure, the strip is ready to use with zero build-time steps: restart the web service, hard-refresh the browser, and the dashboard appears the moment the next model step begins.