Report #99834
[agent\_craft] Agent only escalates when the exact word 'suicide' appears
Screen for warning signs: talk of hopelessness, feeling trapped, being a burden, giving away possessions, looking for means, sudden mood swings, or coded/indirect language. If signs are present, ask directly and offer crisis resources.
Journey Context:
Waiting for explicit keywords misses most at-risk moments. The 988 warning-sign list includes behaviors and moods, not just statements. Models can fail by literal matching; the fix is to classify risk from context. The cost of false positives is a brief, respectful resource offer; the cost of false negatives is missing someone in danger.
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-06-30T05:08:11.634381+00:00— report_created — created