feat: add LLM API (llmapi.ai) as a built-in provider with dynamic model fetching

[MarkosBK]

What is LLM API?

LLM API is an OpenAI-compatible API gateway that routes
requests to 160+ models across multiple providers (OpenAI, Anthropic, Google,
and others) through a single endpoint:

• Base URL: https://api.llmapi.ai/v1
• Auth: Bearer token (LLMAPI_API_KEY)
• Docs: https://docs.llmapi.ai

Why a built-in provider instead of a custom config entry?

LLM API exposes GET /v1/models (no auth required) with rich metadata per
model: name, context_length, input_modalities (vision), reasoning support.
The list changes frequently as new models are added.

A static config entry in crush.json would always be outdated. The right
approach is to fetch the model list dynamically at startup — the same pattern
already used by the Hyper provider.

Proposed implementation

• internal/agent/llmapi/ — package with constants and BaseURL()
• internal/config/llmapi.go — sync mechanism that fetches /v1/models,
  maps the response to catwalk.Provider (name, context_length,
  SupportsImages, CanReason), caches the result
• Uses the existing openai-compat fantasy provider — no new SDK needed
• Gracefully degrades: falls back to cache, then simply doesn't appear if
  offline

Implementation

Working proof-of-concept in my fork. Happy to open a PR once there's
approval here.

[MarkosBK]

@andreynering
Hey! Just wanted to follow up - I already have a working implementation ready as a PR. The proof-of-concept is solid: dynamic model fetching from /v1/models, proper metadata mapping (vision, reasoning, context length), and graceful fallback on network errors.

Would love to get a quick review and merge this in — it's a clean addition that keeps model list always up-to-date without any maintenance overhead.

Happy to address any feedback!

[andreynering]

Hey @MarkosBK,

New providers should be added to Catwalk.

A PR is welcome.