
hi-fangj/dsh-models-radar
dsh-models-radar
Model capability radar plugin for the DeepSeek Harness Web GUI
Install
npx @deepseek-ai/dsh plugin --profile web add github:hi-fangj/dsh-models-radarRestart `dsh web` after install. Bundle APIs can change during the developer preview.
README badge
[](https://dshhub.dev/plugins/dsh-models-radar)Paste this into your README. The star count updates with every catalog sync.
From the README
Excerpt from hi-fangj/dsh-models-radar, cleaned of badges and images.
dsh-models-radar
<h1 align="center">dsh-models-radar · Model Capability Radar</h1> <p align="center"><b>Crowd-benchmarked capability scores from deng.codexradar.com, inside the DeepSeek Harness: overview, trend, and cost on one screen.</b></p><a href="./README.zh.md">中文文档</a> · <a href="#usage">Usage</a> · <a href="#data-and-privacy">Data & Privacy</a> · <a href="#troubleshooting">Troubleshooting</a>
A model capability radar plugin for the DeepSeek Harness Web GUI. It reads public benchmark data from deng.codexradar.com, adds a Model Radar page to Settings, and shows the selected session model's live DeepSWE score below the composer.
Screenshots
Settings · capability overview — best-effort-per-base ranking with per-row Harness attribution (Codex / DSH / ZCode / Grok / Kimi Code); click any row to switch the charts below to that tier.
Capability popover — opened from the composer readout: cross-base comparison plus the current tier's details, with a live "current" mark following the session model.
Highlights
- Score attribution at a glance. Every base-model row carries a Harness badge (Codex / DSH / ZCode / Grok / Kimi Code, site palette), and the tier selector options read
model · effort · harness; unmatchable bases get no badge — never a guess. - Best-effort-per-base ranking. The capability overview groups by base model with a fixed
0–110absolute-scale magnitude bar and a 24h trend signal per row; expand a row for the base's full reasoning-effort ladder. - 24h / 7d dual-window IQ trend. Tab between two time windows, each independently scaled with its own full stats (net change, low, average, high); the curve is colored by capability band.
- Cost × IQ from three angles. Tabs for composite cost (the site's own 2.5×-price-for-1.35×-speed trade-off, normalized per chart), time cost, and price cost; color = base, shape = reasoning effort, same-base tiers joined by ladder lines. Upper-left = more efficient.
- Live readout beside the composer. Exact
model@reasoningEffortmatching through DSH's official per-session model directory, updating immediately on model switches; click it to open the capability popover for cross-base comparison. - Lightweight, credential-free, offline-tolerant. The browser never hits upstream directly (same-origin host proxy), freshness windows mean zero upstream requests inside a window, the latest local snapshot serves as fallback, and no credentials are requested or submitted.
Features
…

