Shift the acceptance bar from "must be uplifting" to "will a reader finish this
calm or a little better, never worse." Keep neutral-but-absorbing (discoveries,
explainers, clever builds, useful insight), and reject anxiety-inducing content —
especially the comparison traps (inferior/behind/FOMO/hustle/status). Scores still
back the verdict. Lets us pull from mainstream sources and filter, rather than
relying on niche good-news outlets.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
client._chat() JSON-parses every response (for the classifier), so the plain-text
summary was rejected ("model did not return JSON") even though the model returned
a perfect summary. Split out _raw_content() and add chat_text() for free-form
output; summaries use it. _chat keeps parsing for classification.
Three-layer organization: primary topic (one per article, for ranking and
brief balance) + grouping tags (1-4 per article from a controlled vocabulary,
the organic "wandering" axis) + tonal flavor.
- taxonomy: add technology + learning topics; 4 calm tag families
(Discovery & Wonder, People & Kindness, Solutions & Progress, Mind & Craft)
defined in code, not the DB; ALLOWED_TAGS union + coerce_tags validation.
- db: article_tags(article_id, tag) join table + tag index.
- llm: tags added to the classifier json_schema (enum-constrained, maxItems 4)
and system prompt; normalize_scores coerces tags; upsert_article_score
replaces a row's tags atomically on every (re)classification.
- queries: feed gains a tag filter and exposes tags via group_concat; tag_counts.
- api: Article.tags, feed tag param, and /api/families with per-tag counts.
- tests: coerce/normalize/upsert/tag-filter/reclassify-replace/tag_counts +
/api/families. 99 passing.
Corpus reclassify (re-tag + new primary topics) runs separately against the
local LLM. Frontend (B2) pairs with this; the live site is unchanged until then.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
- Add pytest suite (34 tests) covering scoring thresholds, dedup clustering +
representative selection + time window, brief source/category diversity,
avoid-term phrase matching, and text canonicalization/truncation.
- Rewrite _select_diverse with an explicit, tested contract (best-first, one
per source, backfill, then inject a second category by evicting the
lowest-ranked pick).
- classify_articles now returns attempted/succeeded/skipped (ClassifyReport) so
silent model failures are visible in both the cycle and classify output.
- Fix clean_text truncation to stay within max_len (ellipsis no longer
overshoots).
- New filters.py: canonical FilterPrefs shape (include/mute topics+flavors,
avoid_terms, pauses) and pure word/phrase-boundary matching engine seeding
Calm Filters. Not yet wired into the API.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
- cycle now prints per-article classify progress (flushed) so the long step is
clearly alive rather than appearing hung.
- An exclusive flock guards the cycle so a manual run and the systemd timer (or
two timer ticks) cannot overlap and contend on the database and model.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
- LocalModelClient.embed() calls the OpenAI-compatible /embeddings endpoint
(local nomic model); base_url shared with chat, model via GOODNEWS_EMBED_MODEL.
- New article_embeddings table and articles.duplicate_of column (+ migration).
- dedup module: embeds missing articles, clusters near-identical stories within
a date window by cosine similarity (pure-stdlib, vectors normalised once), and
marks all but the highest-ranked member of each cluster as a duplicate.
- 'dedup' CLI command; cycle now runs poll -> classify -> dedup -> brief.
- Feed and brief queries hide duplicates, so a story carried by multiple
outlets shows once.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
- poll_due_sources(): polls only sources whose last successful poll is older
than their poll_interval_minutes (or never polled), finally giving that
config field meaning.
- classify gains only_unclassified to spend the LLM solely on new (heuristic)
articles, so a frequent scheduled run stays cheap.
- 'cycle' command runs poll-due -> classify-new -> rebuild-today's-brief, with
each step non-fatal so a down model endpoint or empty day never aborts it.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
- New taxonomy module: single source of truth for 6 topics x 5 flavors,
shared by the LLM response schema (enum-constrained) and validation.
- Classifier now assigns one topic + one flavor per article; json_schema
enums force valid values, with coercion as a safety net.
- article_scores gains topic/flavor columns via an idempotent migration.
- New 'list-category' command to browse by topic and/or flavor, ranked by
composite score.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
- Use json_schema structured output (newer LM Studio rejects json_object),
escalating through json_schema -> json_object -> text and pinning the
first format the server accepts to avoid wasted round-trips.
- Make per-article failures non-fatal and commit incrementally so a single
timeout no longer discards the whole batch.
- Raise default timeout to 180s (configurable via GOODNEWS_LLM_TIMEOUT) for
larger local reasoning models.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
thejayman77