dsh-performance-guard

Multiple rounds of a long DSH session can degrade silently: scrolling stutters, tool-card expansion lags a beat, background watchers and timers quietly eat CPU. The root cause might be browser rendering, the Host event loop, the tool-call chain, or a particular group of third-party plugins. DSH Performance Guard aligns a 60-second browser-side capture (FPS, Long Task, Layout Shift, Resource Timing, DOM size) with Host process CPU, RSS, heap, and event-loop latency onto a single per-second timeline, so you can see browser, Host, tool, model, and session-event activity side by side.

Beyond observation, the plugin offers reversible third-party plugin isolation. It generates a safe profile that removes a suspect half of the plugin pool, re-runs the same workload, and applies paired-difference statistics: a Student 95% confidence interval on the mean difference and a bias-corrected Hedges g effect size. Steps that fail to reach significance stop as inconclusive rather than forcing a verdict, and only repeated significant improvement narrows the suspect list to a single plugin. Nothing is monkey-patched, no third-party source is rewritten, and the DSH conversation renderer is left untouched.

This is for DSH users who run long multi-turn sessions, load several third-party plugins, and occasionally hit latency they cannot pin down. There is no APM platform to deploy: open the settings panel, start a baseline, reproduce the sluggish interaction, follow the persisted next-step guidance to isolate and compare, and export a local JSON or HTML report when done.