Support declaring image input capability for custom openai-compatibility models

[a137460387]

Problem

When using openai-compatibility with a custom model (e.g. mimo-v2.5-pro from Xiaomi API), the proxy returns:

{"error":{"code":"404","message":"No endpoints found that support image input"}}

The upstream API (api.xiaomimimo.com) supports both text and image input for mimo-v2.5-pro, but the proxy has no way to know this since the model is not in the built-in models list (models.router-for.me/models.json).

Current config

openai-compatibility:
  - name: mimo-0mm4
    base-url: https://api.xiaomimimo.com/v1
    api-key-entries:
      - api-key: sk-xxx
    models:
      - name: mimo-v2.5-pro
        alias: ""

Steps to reproduce

curl -s -X POST http://127.0.0.1:8317/v1/responses \
  -H "Authorization: Bearer <api-key>" \
  -H "Content-Type: application/json" \
  -d "{\"model\":\"mimo-v2.5-pro\",\"input\":[{\"role\":\"user\",\"content\":[{\"type\":\"input_text\",\"text\":\"describe this\"},{\"type\":\"input_image\",\"image_url\":\"data:image/png;base64,...\"}]}],\"max_output_tokens\":50}"

Response:

{"error":{"code":"404","message":"No endpoints found that support image input"}}

Expected behavior

A way to declare model capabilities in the config, e.g.:

models:
  - name: mimo-v2.5-pro
    alias: ""
    supports_image: true
    input_modalities: [text, image]

Or alternatively, allow a local models.json override that supplements the remote model list.

Use case

Third-party APIs (like Xiaomi mimo) that are OpenAI-compatible but not in the standard model registry need a way to declare their capabilities (image input, vision, etc.) so that clients like Codex Desktop can use them fully.

CPA Version

v7.2.28