[Bug]: Images pasted via clipboard or provided as file references are never received by the AI model

Platform

Windows

AionUi Version

latest

Bug Description

Images provided through any input method (clipboard paste, file upload, direct path reference) are not reaching the AI model layer in AionUI v1.9.25. The UI renders thumbnails and attachment indicators correctly, but no binary image data is transmitted to the inference backend. When asked about an image, the model hallucinates details instead of reporting that it received nothing.

Steps to Reproduce

1. Take a screenshot or obtain any image file (PNG, JPG, etc.)
2. Paste into chat via Ctrl+V or provide as a direct path reference (C:\path\to\image.jpg)
3. Observe the message appears with an attachment indicator in the UI
4. Ask the model to describe what it sees — it fabricates content instead of reporting missing input

Expected Behavior

The pasted or referenced image should be attached as part of the multimodal payload so the model can see and describe it accurately.

Actual Behavior

Model silently fails and lies

• The paste renders visually in the chat UI (thumbnail/attachment indicator visible)
• No image file/data reaches the inference backend (LM Studio API at port 8877)
• When asked to describe or reference the pasted image, the model hallucinates details that were never provided — fabricating content from nothing because it received zero visual input

Additional Context

Evidence / Example

In a recent session:

1. User pasted image.png showing a Collabora Online splash page with admin credentials displayed
2. Model responded by describing specific details (username, password, IP address) that were NOT in the actual image — clear hallucination from not receiving the file

Additional testing confirmed the same behavior when providing regular files via direct path reference (C:\Users\win10\Downloads\CockShot-69.jpg, a 3 MB JPEG). The model reported it as "binary" and could not access its contents.

Scope of the Bug

This appears to be a systemic issue with all image/binary inputs, not limited to clipboard paste alone:

Input Method	Observed Behavior
Clipboard Paste (Ctrl+V)	Thumbnail renders in UI, but no binary data reaches the model. Model hallucinates content.
File Upload / Attachment Button (untested — please verify)	Likely affected if it uses the same payload serialization path as clipboard paste.
Direct File Path Reference (e.g., C:\path\to\image.jpg)	Read tool reports (binary file, N bytes) but does not decode or pass content to the model layer. Model cannot access image data at all.

Environment

Parameter	Value
AionUI Version	v1.9.25
Host OS	Windows 11 Enterprise x64, Build 26200 "Blackbird" (Preview)
WSL Backend	Ubuntu 24.04.4 LTS on WSL2 (Kernel 6.6.87.2-microsoft-standard-WSL2)
Model Serving Layer	LM Studio Local API — headless mode, Developer Mode ON
Inference Endpoint	<redacted>:8877 (Tailscale internal network)
Loaded Model	qwen/qw3-6-35b-a3b (MoE / 35B-A3B) via LM Studio

Impact

Users cannot share screenshots with the model for debugging, visual confirmation, or any task requiring visual context. This is a fundamental capability gap that affects all multimodal workflows and renders AionUI unable to perform even basic image analysis tasks.

Suspected Root Cause Areas

1. Clipboard paste handler (Electron/Desktop layer) — captures image data from clipboard but fails to persist/write it to a temp file before sending
2. Message payload builder — attachment metadata is added (hence the thumbnail renders in UI) but actual binary/file reference is empty or dropped during serialization
3. Model input pipeline — multimodal/image attachments not wired through properly, so only text reaches the LM Studio API endpoint
4. Binary file reader / tool layer — Read tool does not decode image files to a format the model can consume (e.g., base64-encoded data URL or OCR'd text), and no fallback mechanism exists for multimodal inference

Priority

High — this blocks a core feature of any AI assistant that needs to work with visual information from the user's system.