dsh-opencode-models

The built-in model fetch button in DSH settings reads a static snapshot shipped inside the pi-ai npm package. OpenCode Go's live roster shifts constantly—new models appear, old ones retire, others go temporarily dark—so that snapshot is always a step behind reality. dsh-opencode-models takes the honest route: it calls the server's /models endpoint for the current full list, then fires a text ping and a 1-pixel image ping at every single model before deciding what actually works, and only then writes the usable set back into your configuration.

The write-back logic is deliberately conservative. If nothing changed, the config file is never touched. If more than half the models are undecidable due to network-level errors, the entire sync aborts rather than risk wiping your list. Manual metadata you have already set—contextWindow, maxTokens, custom names—survives every sync cycle. New entries record only the model id and measured vision or reasoning support; concrete token limits are resolved later by DSH's built-in catalog, so a rough guess is never frozen forever. Probe outcomes also surface reasoningEfforts declarations, making a thinking-intensity slider appear directly in the Web UI.

This plugin is built for anyone running DSH alongside OpenCode Go or any OpenAI-compatible model aggregator. It targets the opencode-go route by default, and a single config swap (baseURL plus API key) repurposes it for other gateways. Zero build step and zero runtime dependencies on both the host and browser side—install, restart DSH Web, refresh the page, and the new model-sync panel is ready in your settings.