8 docs tagged with "safety"

Status

Canonical rules for what the ChurchWiseAI chatbot and voice agent must always and never do — derived from battle-tested prompt fragments, the HEAR protocol, and the Safety Covenant

Multi-layer content moderation for the chatbot — blocklist patterns, LLM-based abuse detection, user restriction tracking, and crisis escalation to 988 / church staff

Multiple Claude Code windows routinely run against the same repos, the same Supabase instance, and — as of this incident — the same skill. This runbook documents the first known same-skill collision, explains what saved us, and lays down a small convention (an ACTIVE-SESSION.md sentinel) so the next collision is caught at entry instead of 45 minutes in.

How both the chatbot and voice agent detect crisis signals (suicidal ideation, domestic violence, medical emergencies), immediately route to safety resources (988), and escalate to pastoral staff

Git safety protocol for concurrent Claude agents — always create your own branch, never reuse another agent's branch, rebase before pushing, and never force-push to main/master

Complete audit of every protection mechanism for the ChurchWiseAI chatbot and voice agent, with gap analysis and the test/baseline/benchmark framework definition.

Definitive step-by-step trace of every function, LLM call, STT/TTS step, database query, and third-party service call from SIP arrival to call termination. Required reading before touching any voice code.