Two upgrades to make the digest actually readable:
1) Article-grounded 2-3 sentence summaries (everywhere)
The old prompt got just the title + miniflux's content excerpt,
which for HN/Lobsters/wire feeds is barely more than the title
itself — so summaries paraphrased the title and added nothing.
Now every URL gets fetched and main-content-extracted via
trafilatura on a parallel pre-pass (10 workers, ~15s for ~50
URLs). Extracted text caches to /output/.article-cache.json with
a 7-day TTL so repeat runs in the same window don't re-pull.
Headlines also get summarized now — one batched LLM call per
category (world / local). Rendered as a paragraph below the
title with source + time on the right rail.
Prompt rewrites tell the model to pull names/numbers/places
from the body and explicitly forbid restating the title.
Result: real specifics ("71% saw no pay increase globally",
"third time in less than two weeks", "Islamabad and Moscow
intermediaries") instead of title paraphrase.
2) Per-desk collapse buttons
Chevron next to .desk-count toggles a .is-collapsed class.
Collapsed state is per-device (localStorage by section id) since
collapse is a viewing preference, not content state.
vh
40f1e0ee00
