The dedup hot path recomputed normalize_text() 4 times per comparison
and recomputed item_text() on every inner-loop iteration. Pre-computing
n-gram sets and token sets into a _PreparedText cache cuts dedup time
by 6x (2.16s to 0.39s on 300 unique items).
Bird handle searches spawned one Node process per handle sequentially.
Now uses ThreadPoolExecutor so N handles run concurrently. Same pattern
applied to YouTube comment enrichment (was serial, Reddit was already
parallel) and the retry-thin-sources phase in the pipeline.
Clustering now pre-computes candidate text and uses prepared_similarity
for the O(n^2) grouping and MMR representative selection loops.
Minor: _is_wsl() cached with lru_cache, Bundle.add_items() uses
extend() instead of list concatenation.
End-to-end: 5.2s -> 3.7s (29% faster) on a typical 4-source query.
Add extract_transcript_highlights() that scores sentences by specificity
(numbers, proper nouns, topic relevance) and filters YouTube filler
(subscribe, welcome back, etc). Top 5 highlights shown as structured
bullets in compact output. Full transcript moved to collapsible <details>
block so the LLM reads highlights first, full text on demand.
SKILL.md updated to instruct the judge agent to quote highlights
directly in synthesis, same as Reddit top comments.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
TRANSCRIPT_MAX_WORDS raised from 500 to 5000 so the LLM gets the full
content of most videos (up to ~25 minutes). Removed the second 200-char
truncation in render.py that was reducing transcripts to a single sentence
before the judge agent ever saw them.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
- Remove duplicate detect_query_type from query.py (divergent 5-type version);
canonical 7-type version lives in query_type.py
- Fix reddit.py import to use query_type.detect_query_type
- Clean unused STOPWORDS/SYNONYMS/tokenize imports from youtube_yt, instagram,
tiktok, scrapecreators_x, bird_x after relevance consolidation
- Fix _relevance_filter default from 0.7 to 0.0 (items without relevance
should not silently pass the filter)
- Remove --dateafter from yt-dlp (returns 0 results for evergreen topics)
- Remove restrictSearchableAttributes from HN search (misses Ask/Show HN)
- Lower HN points filter from >5 to >2 (avoids filtering niche posts)
- Add error logging to select_openai_model HTTP failures
- Remove mise.toml and internal planning doc from repo
- Update module docstrings to describe current purpose, not migration history
- Update tests to import from canonical relevance module
- hackernews: use extract_core_subject instead of raw topic, add
points>5 filter and restrictSearchableAttributes=title to reduce
noise from URL-match and low-signal posts
- youtube: add --dateafter parameter to yt-dlp for server-side date
filtering (Python soft filter still handles fallback)
- reddit: skip opinion/review query variant for how_to/comparison
queries where it adds noise
- bird_x: add OR-group retry with compound terms before falling back
to word-dropping (uses X OR operator for multi-concept queries)
- query.py: add detect_query_type() and extract_compound_terms()
Replace duplicated STOPWORDS, SYNONYMS, _tokenize, and _compute_relevance
in four modules with imports from the shared relevance.py module.
Existing tests pass unchanged since modules re-export the functions
under the same names via import aliases.
Replace duplicated _extract_core_subject() in bird_x, reddit, youtube_yt,
tiktok, instagram, bluesky, and scrapecreators_x with thin wrappers that
delegate to query.extract_core_subject() with platform-specific noise sets.
Each module preserves its current behavior exactly:
- bird_x: max_words=5, strip_suffixes=True, full noise set
- youtube_yt: keeps tips/tricks/tutorial/guide/review (content types)
- reddit: preserves original smaller noise set
- tiktok/instagram: same small noise set
- bluesky/scrapecreators_x: minimal noise set
Existing tests pass without modification since _extract_core_subject()
still exists as a callable on each module.
Ran 15-way blinded comparison (5 topics x 3 versions). CROSS won all 5 topics
(4.74/5.0 avg vs HN 4.10, Base 3.73). Then improved CROSS further:
- dedupe.py: hybrid similarity (token+trigram Jaccard) at 0.40 threshold,
cross-source links went from 3 to 13 items across 5 topics
- render.py: [xref: HN5, HN4] -> [also on: HN, Reddit] for human-readable tags
- youtube_yt.py: SYNONYMS dict so "hip hop" matches "rap" (0.33 -> 0.71 score)
- SKILL.md: instruction #7 tells Claude to lead with cross-platform signals
Validation: improved CROSS scores 4.38/5.0 vs original 3.98 (+0.40), wins 4/5
topics. Biggest gains in specificity (+0.8) and format compliance (+1.0).
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
YouTube videos now get real relevance scores based on token overlap
between the search query and video title (was hardcoded at 0.7).
Uses ratio overlap with stopword removal, floored at 0.1.
Cross-source linking annotates items that discuss the same story
across different platforms (e.g., Reddit + HN + X). Items get
bidirectional cross_refs displayed as [xref: R3, HN5] in compact
output so Claude can triangulate multi-platform coverage.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
- Remove 2>&1 from SKILL.md so stderr doesn't pollute model input
- Run script in foreground (not background) with 5min timeout
- Add explicit YouTube synthesis instruction for Claude
- Remove --flat-playlist which broke date filtering (all dates were None)
- Move date filtering to Python with soft fallback for evergreen topics
- Keep 'tips', 'tutorial', 'review', 'guide' in YouTube search queries
- Increase yt-dlp timeout from 60s to 120s for full metadata fetch
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
YouTube search and transcript extraction runs automatically when yt-dlp
is installed. Searches for topic videos from the last N days, fetches
auto-generated transcripts for top results, and feeds them through the
same scoring pipeline (relevance + recency + engagement) as Reddit/X.
New files:
- youtube_yt.py: search, transcript extraction, VTT cleanup
Modified files:
- schema.py: YouTubeItem dataclass, updated Report
- normalize.py: normalize_youtube_items()
- score.py: YouTube engagement scoring (views-dominated)
- dedupe.py: YouTube deduplication
- render.py: YouTube section in compact output
- env.py: is_ytdlp_available() check
- ui.py: YouTube progress messages
- last30days.py: _search_youtube(), parallel execution with Reddit/X
- SKILL.md: YouTube in stats box, citation priority
- README.md: YouTube docs, yt-dlp requirement, Peter shoutout
Inspired by Peter Steinberger's yt-dlp + summarize toolchain approach.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Ilia Alshanetsky