dsh-model-metadata

When you wire a custom gateway into DSH, the settings file usually carries nothing but a bare model id. DSH looks up metadata by pairing the gateway name with the id verbatim, so a self-hosted gateway never matches the built-in catalog. The consequence: context window defaults to a blunt 256 K, reasoning levels come up empty, and image input is rejected outright for models like gpt-4o or deepseek-v41-flash whose names are perfectly catalogued.

dsh-model-metadata does exactly one thing: it strips the gateway prefix, takes the name after the last slash, and walks a lookup chain (official vendor route, built-in pi-ai catalog, opencode aggregator, models.dev snapshot) to fill in context length, output cap, reasoning levels, and vision support. It only writes fields you left blank; anything you declared by hand stays untouched. If the settings document is missing or unparseable, it refuses to enrich rather than risk overwriting your values. The editor gains two controls DSH does not ship natively—reasoning level and vision—with an optional per-row Write button that commits only those two fields. The catalog snapshot auto-refreshes at startup, on panel open, and on a 24-hour cadence, so new model releases are picked up without manual intervention.

Ideal for users who point DSH at a self-hosted or third-party gateway yet use industry-standard model names. If you are on an official first-party route the plugin has no visible effect. Fully proprietary names (your-internal-model-v3) will show a clear no-match notice with nearby catalog siblings for reference—no values are borrowed on a hunch.