Dynamic model discovery via /v1/models endpoint

[Kamilbenkirane]

Context

Providers like HuggingFace serve 127+ models across 15 inference providers, and the catalog changes frequently. Currently, we use MODELS = [] (same as Ollama) and users must pass the model ID directly with explicit provider=. This works but disables parameter validation and prevents list_models(provider=Provider.HUGGINGFACE).

Several providers already define LIST_MODELS endpoints in their config files (Groq, Anthropic, Mistral, OpenAI, etc.) but none are actually called.

Proposal

Add a dynamic model discovery mechanism that can query a provider's /v1/models (or equivalent) endpoint and populate the model registry at runtime.

HuggingFace example

GET https://router.huggingface.co/v1/models returns rich metadata per model:

• id — model ID (org/name format)
• architecture.input_modalities — ["text"] or ["text", "image"] for vision
• architecture.output_modalities — ["text"]
• providers[] — inference backends with context_length, pricing, supports_tools, supports_structured_output

Considerations

• Caching strategy (avoid hitting the API on every import)
• Lazy vs eager loading
• Graceful fallback when API is unavailable
• Which providers to support initially (HuggingFace, Ollama, others?)

Related: #148, #165