Live-tested again: 800 tokens still wasn't enough for Sonnet's bullet density
(6 dense bullets, cut off mid-sentence again). Capped bullet count to 4-5,
added a hard 35-word single-sentence limit per bullet, told it not to add a
closing summary, and raised the cap to 1200 for genuine headroom.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Live-tested on the dev stack: with the more verbose Sonnet model, the old
550-token cap cut the last bullet off mid-sentence. Bumped to 800 and told
the model to keep each bullet to 1-2 sentences so it isn't relying on the
cap alone to stay in bounds.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
- Fix gather_competitor_data() tier filter: it queried tier IN ('primary','secondary'),
values that never exist (real values are 'own'/'competitor'/'market'), so the
cheapest-competitor comparison has always silently returned nothing.
- Feed the previous insight back into the prompt so the model can note what's
changed/resolved instead of repeating itself.
- Extend forecast horizon from 14 to 30 days; add a per-day revenue table
alongside the existing occupancy table.
- Annotate the occupancy table with UK (England) bank holidays.
- Add a same-channel market-movement section (B.com vs B.com, rack vs rack)
diffing rates against the last insight's snapshot, threshold £3.
- Add a parsed headline field + insight history list on the Dashboard,
collapsed to headline/age and expandable to full content.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Python FastAPI ML backend kept intact; auth replaced with central hnf_session cookie verification. Frontend rebuilt on React 18 + TS + Vite with stack design system, Plotly charts retained. Shared Postgres via DATABASE_URL; schema applied on startup.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
