[BUG] preview_screenshot crashes the turn with "Invalid image_url" on vision-capable models

[youlianvr]

On a vision-capable model (ChatGPT-5.6 Luna/Mimo), calling preview_screenshot crashes the turn
with an invalid image payload. The screenshot tool itself returns "ok", but the image attached
to the model request is not a valid base64 data URL:

⚠️ Provider returned error [OpenAI: { "error": { "message":
"Invalid 'input[126].content[0].image_url'. Expected a base64-encoded data URL with an image
MIME type (e.g. 'data:image/png;base64,...'), but got an invalid base64-encoded value.",
"type": "invalid_request_error", "param": "input[126].content[0].image_url",
"code": "invalid_value" } }]

Key observation

• The crash happens ONLY on models that DO support vision
• Text-only models are NOT affected: the orchestrator does not attach the screenshot to them,
  so there is nothing to reject.
• Conclusion: the bug is in the screenshot pipeline (binary → base64 encoding / transport).

Steps to reproduce

1. Start a conversation on a vision-capable model (e.g. ChatGPT-5.6 Luna/Mimo 2.5).
   Then the agent should:
2. Register a preview (htmlPath or url+pid).
3. Call preview_screenshot.

---

4. The tool returns "ok", but the next model turn fails with the invalid image_url error above.
5. Repeats on every retry in this session.

Environment

• OS: Windows 11 LTSC IoT Enterprise
• Freebuff Desktop 0.55
• Preview feature: register_preview + preview_screenshot
• Affected model: ChatGPT-5.6 Luna/Mimo-2.5. Text-only models unaffected.

---

I used Freebuff to help me write this issue.