The Dove run hit 51% manifest coverage with 17 transcripts + 9 bundles
missing. Three real bugs uncovered in sequence by the manifest jq
introspection:
1. Transcripts gated as required. The Apify TIKTOK_TRANSCRIPTS actor
only returns text for videos with captions/subtitles; haircare,
beauty, dance, ASMR, and lifestyle TikTok content (which IS the
target audience for this kind of brief) often has none. Stage 6's
prompt already handles bundle.transcript === null gracefully — it
logs "(no transcript)" and Claude analyses on caption + 30 comments
+ frame note. Made transcripts advisory by default; opt back into
strict via MANIFEST_TRANSCRIPT_REQUIRED=true.
2. Backfill-exhaustion path didn't rebuild the manifest. When Stage 5
ran out of pass1 candidates to backfill with, it trimmed
selected_video_ids to survivors and broke the loop — but never
re-evaluated the manifest against the trimmed list. coverage_pct
stayed at the pre-trim value (e.g. 51%) and the gate failed even
though every remaining id was 100% bundled. Now rebuilds.
3. Added a fast-fail when fewer than 5 videos pass the gate. Stage 8b
trends require ≥5 supporting videos per trend; running Stage 6 + 7
on 3 passing videos burns Claude budget for an unsatisfiable
downstream schema. The cli error now distinguishes "coverage too
low" from "passing-set too small" and points at the right remedy
(lower min_likes/min_plays vs. raise min_plays to filter junk).
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
508804b84f