-
v26.04.13.3 Stable
released this
2026-04-13 23:36:22 -04:00 | 413 commits to main since this releaseHighlights
Rich single-article Discuss
The news/RSS Discuss button now produces real conversations about articles instead of one-shot summaries. Architecture is a three-layer cache on
rss_items(context_prepared → content_full → fresh fetch) so expensive prep work happens once, upfront, at click-time — the chat itself stays as fast as any normal chat turn regardless of article size.- Small articles (≤48k chars) pass through unchanged and cache on first click
- Large articles run a parallel map-reduce step on the background model (chunked by paragraph, ~300-word factual summaries per chunk concatenated under section headers) — one-time cost, cached for repeat clicks
- Follow-up turns in both cases are zero extra cost — the article context is already in history as a synthetic
read_articletool result
The canned
"Please summarize and discuss this article."prompt is replaced with a conversational seed that invites a real discussion ("I want to talk about this article. Start with a substantive summary... I'll ask follow-ups from there.") — the old checklist-shaped prompt was making the model respond like it was completing an assignment instead of starting a conversation.The map step pins
num_ctx=16384explicitly to avoid the same silent-truncation footgun that wrecked the research pipeline in v26.04.13.2.discuss_topic(multi-article aggregation) is intentionally untouched — follow-up work will decide whether to retire it or rework it on the cached-context approach.Schema
Migration
0038adds three nullable columns torss_items:content_full,context_prepared,content_fetched_at. Additive — no backfill, cache fills in lazily.Downloads