# Codex OAuth image-to-image smoke test

This is a live smoke test for the PR that forwards `image_url` as a Responses `input_image` part to the OpenAI Codex-authenticated image backend.

## Configuration

```yaml
image_gen:
  provider: openai-codex
  model: gpt-image-2-high
```

Auth source: `~/.codex/auth.json` ChatGPT/Codex OAuth token, read through Hermes' existing `_read_codex_access_token()` helper.

## Source image

Generated with the existing text-to-image path:

```python
provider.generate(
    "A tasteful, wholesome portrait of a beautiful young adult woman ...",
    aspect_ratio="portrait",
)
```

## Edit request

The source image path was passed back to the same provider as `image_url`:

```python
provider.generate(
    "Edit the provided portrait while preserving the exact same adult woman identity, face, hairstyle, expression, lighting, camera angle, and background. Change only the outfit to a modest Japanese school-uniform-inspired JK outfit: navy sailor-style blouse collar, simple ribbon tie, navy pleated skirt below the knee. Keep it wholesome, non-sexual, realistic photography.",
    aspect_ratio="portrait",
    image_url="/Users/baicai/.hermes/cache/images/openai_codex_gpt-image-2-high_20260701_021616_67fb90c5.png",
)
```

## Result shape

```json
{
  "success": true,
  "image": "/Users/baicai/.hermes/cache/images/openai_codex_gpt-image-2-high_20260701_023558_15fb310e.png",
  "model": "gpt-image-2-high",
  "aspect_ratio": "portrait",
  "modality": "image",
  "provider": "openai-codex",
  "size": "1024x1536",
  "quality": "high",
  "input_image_count": 1
}
```

The important behavior is that the result reports `modality: "image"` and `input_image_count: 1`, proving Hermes did not silently fall back to text-to-image.