Summary

The garbage output persists after #247. Need more diagnostic data to isolate the root cause. This adds comprehensive one-time logging:

1. Raw messages — role, content types, text preview (before chat template)
2. Full rendered prompt — 2000 chars (was 300)
3. Image metadata — mode, size, format
4. Generation config — max_new_tokens, temperature, constrained, model type, device
5. First generation output — 500 chars + token count
6. Input tensor shapes — input_ids, attention_mask, pixel_values, image_grid_thw

Key hypothesis: If pixel_values is MISSING from the inputs, the model isn't seeing the screenshot. This would explain degenerate output regardless of prompt correctness — the model is effectively blind.

What to look for in the logs

TRL prompt msg[0] role=system content=You are a GUI automation agent...
TRL prompt msg[1] role=user content_types=['image', 'text'] text=Goal: create-desktop-folder...
TRL prompt text_input (N chars): <|im_start|>system\nYou are a GUI...
TRL prompt image: mode=RGB size=(1920, 1080) format=PNG
TRL generation config: max_new_tokens=512 temperature=0.7...
TRL first generation output (N tokens): Thought: # # # # #...
TRL input shapes: input_ids=torch.Size([1, N]) pixel_values=torch.Size([1, ...]) ...

If pixel_values shows MISSING, that's the bug.

🤖 Generated with Claude Code