v3.19.3

Fixed

• Attaching an image to a text-only model no longer breaks the session. DeepSeek (and other text-only models) reject image content with a 400, and the bad message used to stay in history so every later turn re-sent it — you couldn't chat until restarting. Kodiqa now detects models that can't see images, sends your message as text + a note instead of crashing, and tells you to switch to a vision model (claude, gpt-4o, or a qwen-vl/omni model). A defensive pass also strips images already in history so they can't poison later turns.
• Drag-and-drop robustness. A whole-string last-resort attach catches dropped files even when the terminal sends them as keystrokes (waking the completer) instead of a clean bracketed paste; unreadable paths now show exactly what was received instead of a bare "unknown command".