Enable screenshot/image input for custom Vision models in ZCode

ZCode may classify an auto-discovered custom OpenAI-compatible model as text-only even when the underlying model and gateway support Vision.

After attaching a screenshot, ZCode responds with a message similar to:

I'm unable to view images. The selected model does not support image input.

ZCode stores custom-provider metadata in:

~/.zcode/v2/config.json

An auto-discovered model may be recorded with an explicit text-only input modality:

{
  "modalities": {
    "input": ["text"],
    "output": ["text"]
  }
}

Setting only "supportsImages": true

is not enough. ZCode uses modalities.input

as the effective capability gate, so the working model entry needs both declarations:

{
  "modalities": {
    "input": ["text", "image"],
    "output": ["text"]
  },
  "supportsImages": true
}

First verify that the model accepts an OpenAI-compatible multimodal request through the same gateway used by ZCode. A valid message uses a content array containing text

and image_url

parts:

{
  "model": "glm-5.2",
  "messages": [
    {
      "role": "user",
      "content": [
        {"type": "text", "text": "Describe this image."},
        {
          "type": "image_url",
          "image_url": {"url": "data:image/png;base64,..."}
        }
      ]
    }
  ]
}

If this request fails, fix the model/gateway path first. This ZCode patch only corrects client-side capability metadata.

• Quit ZCode completely ( Cmd-Q

on macOS;Quit/Exit on Windows or Linux). Closing only the window is insufficient because ZCode caches model capabilities. - Download fix-zcode-custom-model-vision.py

from this gist. - Run it with the model ID used by the custom provider:

python3 fix-zcode-custom-model-vision.py glm-5.2

If multiple custom providers contain the same model ID, list the candidates:

python3 fix-zcode-custom-model-vision.py glm-5.2

Then rerun with the intended provider UUID:

python3 fix-zcode-custom-model-vision.py glm-5.2 --provider-id PROVIDER_UUID

The script:

• ignores ZCode's built-in providers;

• creates a timestamped backup;

• preserves credentials and unrelated settings;

• atomically updates the configuration;

• verifies the saved result;

• never prints API keys or provider options.

• Reopen ZCode.

• Select the patched custom model.

• Start a new task/chat. - Attach a screenshot.

• Ask for objectively verifiable OCR, such as:

Transcribe the largest visible heading exactly.

A successful HTTP request alone is not enough—the answer must demonstrate that the model saw the image.

The script prints the backup path, for example:

~/.zcode/v2/config.json.before-image-modalities-20260728-170000

Quit ZCode and restore that file over ~/.zcode/v2/config.json

.

Editing or refreshing the custom provider—or upgrading ZCode—may regenerate the model metadata and restore "input": ["text"]

. If screenshot support disappears, inspect the model entry and rerun the patch.

The durable upstream fix is for ZCode to preserve custom input modalities, expose them in the custom-model UI, or import reliable capability metadata during model discovery.