dsh-catalog-refresh

The DSH model picker reads from a static snapshot baked into the pinned pi-ai package, while providers such as OpenRouter, Fireworks, and Groq ship new models on a rolling basis. DSH users are left either hand-adding custom entries or waiting for the next pi-ai dependency bump, resulting in a disconnected and lagging experience.

dsh-catalog-refresh addresses this by fetching model endpoints from each provider at boot and on a configurable cadence (default 12 hours), converting the response into pi-ai Model entries, and replacing the MODELS registry inside the same in-memory module instance that dsh-llm-pi-ai reads. It then writes a lightweight refresh stamp into the llm-pi-ai settings section, which nudges the picker to re-resolve its routes against the patched registry, so the picker updates immediately with no restart required. Rebuilt reasoning models also receive a default thinkingLevelMap so the reasoning-effort selector remains available. Providers without a configured API key are silently skipped, and the last successfully fetched catalogue is cached so an offline restart still applies the most recent refresh.

If you use DSH across multiple providers and would rather have the model picker track upstream releases in real time than maintain a hand-curated list of custom models, this plugin turns the static snapshot into a living catalogue that stays in sync with every connected provider.