Model catalogs are static — OpenRouter shows 333 of 431 models, DeepSeek 3

[kaiomp]

What happens. @deepseek-ai/dsh-llm-pi-ai 0.1.2-rc.1 serves each provider's model list from the JSON files bundled in @earendil-works/pi-ai (dist/providers/data/*.json). For OpenRouter that is 333 models; GET https://openrouter.ai/api/v1/models returns 431 today. DeepSeek is a fixed list of 3. New models only appear after a pi-ai release plus a harness release.

Why it matters. OpenRouter adds models weekly. Users with an OpenRouter key expect the picker to match the site.

Proposal. Give catalog routes an optional live source:

providers:
  openrouter:
    apiKeyEnv: OPENROUTER_API_KEY
    catalog: live        # fetch /models on boot and every 24h; static list is the fallback

Mapping is small: id, name, context_length → contextWindow, top_provider.max_completion_tokens → maxTokens, architecture.input_modalities → input, supported_parameters containing reasoning → reasoningEfforts: {off:, low: low, medium: medium, high: high}. For DeepSeek, GET /models with the key gives the ids; names and limits can stay in the static file.

Workaround today. A plugin can write the live list into llm-pi-ai.providers.openrouter.models through ctx.settings.mutate; resolveRouteModels merges the static defaults under known ids, so metadata survives. flykit (dsh-flykit) does this. It would be better in the adapter.

[kaiomp]

Follow-up: dsh-flykit now does the live catalog as a plugin. It fetches /models from OpenRouter on boot, so the picker shows 430 models against the 333 bundled. Nine-second clip below.

Related: #5656 is the other side of the same wall. The bundled catalog decides both which models exist and what reasoning options they have.