[Feature] Auto-discover models for OpenAI-compatible providers

[yan-ad]

Feature Request

It would be useful if Crabcode could automatically discover available models from custom OpenAI-compatible providers instead of requiring every model to be manually defined in the configuration.

Most OpenAI-compatible APIs expose a standard models endpoint:

GET /v1/models

Crabcode could use this endpoint to automatically populate the available model list when configuring or using an OpenAI-compatible provider.

Motivation

Currently, when using a custom OpenAI-compatible endpoint, models may need to be manually configured.

This becomes inconvenient for providers that expose many models or frequently add/update their available models.

For example, if a provider exposes:

GET https://example.com/v1/models

Crabcode could fetch the models from that endpoint and make them immediately available in /models.

Suggested Behavior

For providers configured as OpenAI-compatible:

1. Use the configured baseURL and API key.
2. Request GET /v1/models.
3. Parse the returned model IDs.
4. Automatically expose them in Crabcode's model selector (/models).
5. Resolve model metadata from sources such as models.dev or AI SDK when available.
6. Allow manually configured models to override discovered models and metadata.

If the endpoint doesn't support /v1/models, Crabcode could simply fall back to the current manually configured model behavior.

Model Metadata

The /v1/models endpoint usually only provides model identifiers and does not expose enough information about the model's limits or capabilities.

After discovering a model, Crabcode could optionally look up additional metadata from models.dev or AI SDK, including:

• Input token limit
• Context window size
• Output token limit
• Supported modalities
• Reasoning capability

For example, a discovered model could internally resolve to metadata similar to:

{
  "id": "gpt-6-astra",
  "limit": {
    "input": 1050000,
    "context": 1050000,
    "output": 128000
  },
  "reasoning": true,
  "modalities": {
    "input": ["text", "image"],
    "output": ["text"]
  }
}