dsh-dual-model-eval

Evaluating multiple coding models inside DeepSeek Harness used to mean running them one at a time and comparing notes in a spreadsheet. dsh-dual-model-eval turns that into a side-by-side experiment right inside the Chat tab: one prompt is dispatched simultaneously to two to four configured model routes, each working in its own isolated Git worktree cut from the same base commit so nothing bleeds into another run. As the agents work, every result card streams live tool-call trajectories, elapsed time, token counts, TTFT, and decoding details directly in the interface—no log-digging required.

The real payoff is seeing the differences at a glance. Final replies render side by side with precise line-change statistics (+added, −deleted, percentages, changed-file count), and you can preview or download a single edited file or the entire candidate worktree. When a result convinces you, clicking Adopt this result commits the chosen patch locally, advances the shared baseline, and feeds the accepted conversation context into the next round; unselected candidates stay out of scope. A built-in composer gate also blocks the next round until at least one completed candidate has been adopted, so every comparison starts from a confirmed code state. All operations stay local—no remote pushes, no global Git identity changes.

This plugin is ideal for engineers and architects who code daily in DeepSeek Harness and want to benchmark several LLMs on real tasks, as well as tech leads who need reproducible, evidence-backed comparisons to guide team model selection.