DeepSeek Harness LiveBench Panel Plugin
The `dsh-livebench-panel` developed by Vithrive is a plugin for DeepSeek Harness (DSH), designed to simplify local LiveBench evaluation workflows through a web interface. The plugin adds a LiveBench tab to the Trajectory view in DSH Web, supporting model parameter configuration, Baseline probe comparison, automatic retries, and real-time score retrieval. Installation requires LiveBench to be installed locally (Python 3.11 is supported), and the plugin retrieves the API key by parsing Harness environment variables. Functionally, it integrates reasoning intensity settings, protocol routing (e.g., Anthropic/Claude-specific protocols), and automatic
Read Moredsh-blind-arena: An Agent Arena for Local Blind Evaluation on DSH Web
`dsh-blind-arena` is an MIT-licensed local DSH plugin released by `changer-changer`, designed to eliminate identity bias in model comparison through a blind evaluation mechanism. This plugin transforms DSH Web into a local Agent arena, supporting 2 to 4-way contests under the same task, same commit, and isolated worktrees. Its core features include real-time progress monitoring, strict anonymity contracts (public records exclude identity fields), shared verification commands, blind evaluation post-reveal mechanism, offline export of self-contained HTML reports and patches, winning branch promotion, and a zero-cost demo mode without API keys. Installation requires Node.js ^22.19.0 or higher and specific DSH components; run `npm
Read More