The eight reports that list people get the AI summary too, sent as
anonymised totals built on the screen (aiReportData.ts): no names,
contact details, notes or comments; consent and medical only as counts.
The server refuses personal columns on those reports and logs each
summary in the access log (ai_report; sensitive for consent and medical).
Roadmap §18: 2.0.1, then crosscode.co.za, then 2.1.0 (AI helpers on the
Consent log and Access log).
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Found by the first real eval run (Sonnet 5.5 through the local Hub,
19/19 checks): without an event the drafts still used {{event.title}}
and added an event-details block. The instructions now offer those
placeholders only when an event is given; the eval checks for it, and
adds up split themes when counting feedback. Route tests for the five
AI helpers in v2.test.ts; the eval script loads backend/.env.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Ls1vMeckMWrKPfc1MMyMJF
src/ai/helpers.ts: instructions, schemas and inputs for draft event,
summarise report (people-free reports only), draft message, summarise
feedback (no names, access-logged) and improve text. Contact details
are swapped for placeholders before sending and restored after; every
helper gets today's date; weekday clashes and break-even over capacity
are checked in code. v2 routes /ai/draft-event and friends.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>