dsh-livebench-rankings

Opening a browser, waiting for LiveBench to load, skimming the table, then switching back to your workspace—dsh-livebench-rankings collapses that loop into a single AI Rankings button tucked into the lower-left corner of the Harness input bar. It pulls the official CSV and category JSON straight from livebench.ai, detects the latest front-end release, and silently re-checks every 60 seconds by default (manual refresh anytime). On network failure it tells you it is serving cache; it never fabricates or pads scores.

Once open, you get a full-featured leaderboard: filter by institution, open-weight status, or reasoning model; sort by individual task scores; search for a specific entry; each model family keeps its highest-scoring variant by default; compare up to four models side-by-side with task-level detail; export the current view as CSV. Everything is fetched as static resources only—no remote JavaScript execution, no global fetch overrides, no access to chat history. Cache lives purely in the plugin's own directory and vanishes on uninstall.

If you evaluate models day-to-day inside Harness, write comparison notes, or just want a quick score check without the browser tab dance, this plugin fills that micro-gap. It registers no global model tools, modifies no system prompt, and sits quietly under its own route—open when you need it, forget it exists when you do not.