A rule only filtered articles that arrived after it was saved, so the
stories it was meant to remove stayed in the feed for up to two days.
- Saving a negative rule now also removes its matches from the feed
(checkbox, on by default). If the form's dry run of the same rule has
finished, its results are used; otherwise a check runs and removes the
matches as soon as it ends.
- A saved rule's dry run has "Remove these N from the feed".
- Removed articles are filtered with the rule and reason (Filtered,
Rescue), stories left without articles disappear, and "Put back all"
restores them without recording a correction example.
- Dry runs check what's in the feed: articles published within the
feed's time window (at least 24 h, up to 400), and honour enabled
"always keep" rules (keyword exceptions skip the model entirely).
API: POST /api/rules/:id/apply { jobId? }, POST /api/rules/:id/restore.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The reader sorted by when a story entered the feed, so stories that were
processed late (a backfilled source, a retry, the queue after quiet
hours) jumped to the top: two-day-old Android Central and Ars Technica
items sat above today's news.
- Reader (In feed and All stories): newest first by the story's earliest
report, always. The "Most important first" sort is gone.
- Feed file (RSS and JSON): items are picked, windowed and ordered by
publication time as well (or the latest report when updated stories
move to the top). The "most important first within each day" order
setting is gone.
- Phone toolbar: Unread and Bookmarked sit beside the importance filter.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Settings > Fetching & filtering > Quiet hours: a local window (e.g.
23:00-05:00, may run past midnight) with no scheduled fetching and no
background processing, so no extraction and no AI calls. New items wait
in the feeds; when the window ends the scheduler and worker pick
everything up on their next tick.
- Editions, digest creation and cleanup still run on time (no AI); a
digest due inside the window gets its opening when it ends. The
no-opening fallback after 6 hours now only applies while the LLM is
actually unavailable, so a long quiet window doesn't trigger it.
- Things started by hand still run: Fetch now, adding a source, the
Playground, the connection test.
- /api/status reports quiet hours; a banner shows until when, and how
many items are waiting. The settings panel shows the window length and
warns when the digest time falls inside it.
- Off by default. Tests for the window (past midnight, DST) and for
nothing being processed until it ends.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
A new source imported only the last 24 hours, so a tech site added on a
Sunday brought in one or three items next to a world-news feed's fifty.
- While a source is younger than the maximum age, its newest items from
that window are taken until it has 10. Existing young sources fill on
their next fetch.
- Deleting a source forgets its URLs in the seen ledger, so adding the
same feed again imports its current items, and drops its feed override
(source ids can be reused).
Importance is one scale for all news, so tech and science stories score
lower than world news and a single minimum can't balance them.
- Feed controls > Sources: per source, its own minimum importance
(higher or lower than the feed's) and its own daily maximum; the
per-source limit moves here as the default.
- The 7-day preview shows stories a day in and out per source.
The reasoning-effort hint says to leave it on the default when token
counts don't drop (measured on a proxy that ignores it).
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Before the AI (free):
- Seen-URL ledger: every article URL is remembered for 180+ days, so an
item still listed in a feed after its article was pruned (rejected ones
go after 14 days) is never fetched or judged again.
- Wire copy: an identical headline from another outlet within 48 h joins
the existing story (or is filtered like its twin) with no triage,
grouping or summary call, and does not trigger a rewrite.
Triage:
- Up to 8 articles from one source are judged in one call, so rules,
instructions and examples are sent once; the worker waits up to a
minute for a fetch's other articles to finish extracting. Articles the
model skips are judged singly.
- 800 characters of article text instead of 1500.
Grouping without embeddings:
- Free pre-filter on shared names (4-letter stems, so Trump/Trumpov and
Bahrain/Bahrein agree) and numbers; no call when no recent story shares
one, and at most 12 candidates instead of 60.
Title and description:
- Max article characters default 6000 (was 12000); a multi-source rewrite
shares that budget.
- New "Reasoning effort" setting, sent as reasoning_effort. Measured on
production: Haiku 5.5 spends ~590 hidden reasoning tokens per summary.
No wasted retries:
- 401/403/404 from the chat endpoint (bad key, unknown model, wrong URL)
pauses LLM work with a banner instead of failing every article; it
resumes on a settings change or after 10 minutes.
- Simpler structured-output modes are tried only on 400/422, so a
misconfiguration costs one request, not three.
Visibility:
- Activity: calls and tokens per pipeline step, calls per published
story, and work saved without AI. Playground: tokens per model.
Tests for each saving and for the pause on misconfiguration.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Chat-only OpenAI-compatible servers (API proxies) answer /embeddings with
404. Every kept article then failed at the grouping step, so nothing was
ever summarized or published.
- Grouping now has two engines: embeddings (as before) and the triage
model, which gets the new article and a numbered list of the last 48 h
of stories and picks the same event or none. It works across languages.
- The triage-model engine is used when no embedding model is set, and
automatically after /embeddings answers 400/404/405/422/501 for the
configured endpoint and model (remembered until LLM settings change),
so the article that hit the 404 is grouped in the same step.
- Test connection explains a missing embeddings endpoint instead of a bare
"HTTP 404"; an empty embedding model is no longer reported as a failure.
- Settings copy: the embedding model is optional.
- Mock LLM: option for no embeddings endpoint, and story-match answers;
new test reproducing a chat-only server with a chat model in the
embedding field.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
UI
- Fix the LLM settings Models card: the hidden model <datalist> was the
first element after the floated card title, so the field grid never
cleared the float and was squeezed into a narrow column. Every element
after a card title now clears it.
- Save bar floats over the content only while there are unsaved changes.
- Units (s, min, hours, sources) sit inside the input, so inputs in one
row keep equal widths; even spacing between stacked fields in cards.
- Phones: main nav fits on one row (scrolls on very narrow screens, with
the current section kept in view); settings sections are one swipeable
row; feed filters use an even two-column grid with the checkboxes kept
together; story-card actions stay right-aligned when they wrap.
- Editions view: stories that entered the feed in Live mode are labelled
"From Live mode, before editions" instead of "Breaking and added by
hand"; breaking and manual additions get their own groups.
- Activity: clearer stat labels.
Server
- Unwritable data directory: a plain startup message with the exact
chown command for the running user, instead of a SQLite stack trace.
- Daily digest: with fewer than 3 eligible stories it keeps checking
until the day's window closes, instead of giving up on the first tick
(a fresh install started after 06:30 never got a first-day digest).
- "Build now" with too few stories no longer replaces today's digest.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Covers queueing on publish, the breaking bypass, releasing the latest due
edition once with the most important stories, and sweeping queued
stories after switching back to Live.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Merges RSS/Atom/JSON sources into one feed. A local OpenAI-compatible
LLM filters articles against plain-language negatives, groups reports
of the same event into stories, and writes a one-sentence Croatian
title and a description of at most two sentences for each story.
Original articles are only ever linked as published.
- Pipeline: fetch, dedupe, Readability extraction, keyword + AI triage
with exceptions and learned corrections, embedding clustering with an
LLM same-event check, summary with format check and Croatian
language guard, regeneration for developing stories
- Neutral importance scoring with coverage boost and manual override
- Feed controls: Live or Editions, daily/category/source limits,
breaking bypass, 7-day preview; daily digest with offline fallback
- Circuit breaker that pauses and resumes the queue on LLM outages
- Express 5 + node:sqlite server, React 19 + Vite web UI (dark, PWA)
- Feeds under /feeds/ for a single Pangolin rule; optional feed token
- Dockerfile runs typecheck and the full test suite before building;
Gitea workflow builds, scans and pushes the image
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>