dsh-model-arena

dsh-model-arena

Run the same prompt through every model you have, and see the difference without squinting.

Install

dsh plugin --profile web add dsh-model-arena

Or from source:

dsh plugin --profile web add github:hj01857655/dsh-model-arena

What it does

• Run a prompt through N models. Same prompt, identical parameters, outputs collected side by side.
• Compare. Line-level diff between any two model outputs — additions, deletions, and unchanged lines, with a summary of how many lines changed.
• Score. Auto-scorer (length, latency, reference match) + manual user ratings (1–5). Scores are persisted per run, so you can rate a model later and the leaderboard updates.
• Leaderboard. Aggregate scores across runs into a ranked table — win rate, average score, and total runs per model.
• Benchmark suites. Save a named set of prompts + models, re-run after changes, get a regression report that tells you which models improved and which got worse.
• Panel. Settings page listing recent runs with drill-down to comparison view, leaderboard, and regression reports.

CLI

dsh-model-arena list                              # list recent runs
dsh-model-arena compare      # diff two models in a run
dsh-model-arena rate  --model M --rating 4   # rate a model output

Panel

A Model arena page inside Settings. The host registers GET /api/model-arena.panel and rebuilds the payload from the store on every request. A headless host simply skips the route.

The panel shows:

• Recent runs with model outputs side by side
• Line-level diff viewer for any pair of models
• Auto-scores and user ratings per model output
• Leaderboard across all runs
• Regression reports for benchmark suites

Layout

Path	Role
src/index.ts	Host half — apply(ctx), exposes the modelArena service
src/arena.ts	Core arena logic — run, compare, rate, leaderboard, regression
src/scoring.ts	Auto-scorer and leaderboard builder
src/diff.ts	Line-level diff algorithm
src/store.ts	Durable .arena/ store
src/arena-view.ts	Payload → view model, shared by host route and browser half
src/cli.ts / src/bin.ts	Command-line entry, usable without dsh
src/routes.ts	Panel route on the host's web connection
src/client/view.tsx	Browser half — pure rendering
src/client/index.tsx	Browser half — fetch and settings registration
src/client/locales.ts	zh / en dictionaries

Develop

npm ci
npm run typecheck
npm run build
npm test

License

MIT