Compact snapshot format for token-efficient LLM consumption #46
Description
Problem
The ARIA snapshot format is verbose - every link, button, image, and text node gets its own indented line. For a page with 50 links in a sidebar, that's 100+ lines of tree structure (link line + url line each) that an LLM must process.
Example of current verbosity for a single article card:
- link "Article Title":
- /url: /path/to/article
- link "author profile":
- /url: /author
- img "author profile"
- button "Author Name profile details": Author Name
- link "Feb 23":
- /url: /path/to/article
- time: Feb 23
- heading "Article Title" [level=2]:
- link "Article Title":
- /url: /path/to/article
- link "# tag1":
- /url: /t/tag1
That's 15 lines for one card. With 15 cards on a page plus sidebar/footer, the snapshot explodes.
Proposal
Add a --snapshot-compact flag that produces a denser format optimized for LLM consumption:
- Collapse link + url into one line:
link "Title" -> /path - Omit redundant duplicate links (same URL appearing multiple times in a card)
- Omit decorative images (
imgwithout meaningful alt text) - Inline simple children:
heading [h2] "Article Title" -> /path
Example compact output for the same card:
link "Article Title" -> /path/to/article
by Author Name, Feb 23
tags: #tag1
Impact
Rough estimate: 50-70% reduction in snapshot token count for content-heavy pages, directly reducing agent costs per page visit.