Summary

End-to-end image/vision support across the stack, plus a stable browser element ref system for browser automation:

• User image upload: attach images to chat messages from the UI; forwarded to the LLM as image content blocks.
• GCU screenshot passthrough: browser tool screenshots are forwarded to the LLM as image content alongside the text result.
• Vision capability detection: new capabilities.py module checks whether a model supports images (allow/deny lists per provider and model name).
• Automatic image stripping: image blocks are stripped before non-vision model calls to prevent API errors.
• Vision fallback: when the primary model lacks vision, a fallback vision model describes images as text and injects the description.
• Aria ref system: refs.py annotates aria_snapshot() output with [ref=eN] markers so the LLM can target browser elements by stable ID instead of fragile CSS selectors.

Closes

• Closes [Feature]: An option to upload an Image #6579 — user image upload in chat UI
• Closes [Feature]: GCU screenshot passthrough to vision LLM #6678 — GCU screenshot passthrough to vision LLM
• Closes [Feature]: LLM vision capability detection and automatic image stripping #6679 — LLM vision capability detection and automatic image stripping
• Closes [Feature]: Vision model fallback — describe images as text when primary model lacks vision #6680 — Vision model fallback (describe images as text for text-only models)
• Closes [Feature]: Ref system for aria snapshots — stable element IDs for browser automation #6681 — Ref system for aria snapshots

Test plan

• make test passes (unit tests for capabilities, refs, screenshot normalization)
• Upload an image in the chat UI — verify it reaches the LLM
• Run a browser automation task with a vision-capable model — verify screenshots are forwarded
• Configure a text-only model (e.g. DeepSeek-R1) — verify images are stripped/described without API errors
• Use selector="eN" ref in a browser tool call — verify it resolves correctly