Images are silently dropped from chat completions.

[0xThiebaut]

On the latest version (v1.7.15) for iOS, images are not submitted when attached from the Photo Library or from the Camera itself. While the behavior seems similar to #74, neither Photo Library nor Camera images seemed to work.

Files (e.g., PDF) attached through the Documents option seem to work. I was unable to test image files as these are grayed out from the Documents selection.

Request trace:

POST /chat/completions HTTP/1.1
Connection: close
Host: REDACTED
X-Real-IP: REDACTED
X-Forwarded-For: REDACTED
X-Forwarded-Proto: https
Content-Length: 590
Content-Type: application/json
User-Agent: openclient-llm/117 CFNetwork/3896.100.1.2.1 Darwin/27.0.0
Accept: */*
Accept-Language: en-GB,en;q=0.9
Accept-Encoding: gzip, deflate, br
Authorization: Bearer REDACTED

{"messages":[{"content":"Describe this image.\n\nThe following image attachments are not directly available to this model and have not been analyzed yet in this message. If analyze_images is available in the current tool definitions, it can analyze these image IDs. Otherwise, explain that you cannot inspect the images directly. Do not infer visual contents from these references; use actual analysis results when present.\n{\"image_attachment_ids\":[\"8307FA05-138D-4BDE-AE7F-9C676133F302\"]}","role":"user"}],"model":"OpenAI Family","stream_options":{"include_usage":true},"stream":true}

[artcc]

Hi @0xThiebaut, thank you for reporting this and for including the request trace — that's very helpful!

The image flows tested so far are working as expected, including sending an image to a chat model without vision: the main model calls analyze_images, the selected vision model analyzes the image, and the main model returns the result. We'd like to reproduce the behavior with your specific setup.

Could you please share a few more details?

• Server/backend: Are you using LiteLLM, another OpenAI-compatible server or proxy, or a provider directly? Which one and, if applicable, which version?
• Model configuration: What actual model/provider is behind OpenAI Family? Is this an alias or a group that routes to different models, and how is it configured?
• Detected capabilities: Does OpenClient show vision and tool/function-calling capabilities for this model? A screenshot of its details in the Models screen would help.
• Vision delegation settings: Which model, if any, is selected under Image and Vision → Vision, and is analyze_images enabled in the tool settings?

If you're using LiteLLM, the relevant entries from /models and /model/info for this model would also be useful. Please redact API keys, credentials, and any private details.

The trace contains image references but no tool definitions; on its own, it doesn't tell us why this request path was selected in your setup. These details will help us identify what's happening without guessing at the cause.

Thanks again for your help!

[0xThiebaut]

Hi, good catch! The OpenAI Family is an auto-router on LiteLLM. While it routes to models capable of both vision and tools, the auto-router itself is not marked as vision/tool capable. By using a model directly, the calls succeed.