[pcbeingused333]

Images in a Claude request (the image blocks Claude Code sends when a screenshot is pasted) were dropped silently by the Claude -> OpenAI converters: the chat converter only read text, thinking, tool_use and tool_result blocks, and the Responses converter did the same, so the model answered without seeing the image. The Gemini converter already forwards them.

• Chat: a user message with images now gets its content as a list of parts, in order: {"type": "text", ...} and {"type": "image_url", "image_url": {"url": ...}}. A base64 source becomes a data:<media_type>;base64,<data> URL and a url source is forwarded as is. Messages without images keep the plain string content they had.
• Responses: the same blocks become input_image parts ("image_url": <url>) next to the input_text parts.

Image blocks inside tool_result content are not converted, since OpenAI tool messages are text only.

Tests: claude_images_test.go covers base64 and url images in both converters and that image-less messages are unchanged; the image cases fail on master. go vet and go test ./internal/... pass. Not run against a live OpenAI or compatible upstream.

This change was prepared with AI assistance (Claude Code).