dsh-model-tester

After wiring up several providers and models in DeepSeek Harness, the most nagging question is usually the same: which ones actually work, how fast are they, and is that intermittent error a real problem or just a misconfiguration? dsh-model-tester turns that into a single click. A panel appears at the bottom of the Settings → Models page; pressing it fires one genuine llm.stream request per model (3-way concurrency, 30-second per-model timeout) and lays out availability, decode TPS, first-token latency, and total elapsed time on a single row. Failed rows expand to show the reason, error code, and HTTP status, so you are not left guessing.

Beyond the one-shot full sweep, the panel offers finer-grained controls for everyday troubleshooting. Each provider header carries a dedicated Test This Channel button that runs only that provider's models, and every model row has its own prominent Test / Retest button for repeated spot-checks. Inactive or unconfigured providers, along with auto-registered entries whose model id or name contains vision, are automatically hidden to keep the list focused. Model inventory prioritizes what you explicitly configured in settings and falls back to the adapter catalog when nothing is set.

It is aimed at Harness users who connect to multiple upstreams and need a fast, reproducible way to gauge availability and throughput, or who want to batch-verify everything after switching providers. No build step is required — two plain JS files load directly, and you are ready to go after restarting the web process.