Files
hermes-agent/agent/crisis_hook.py
Timmy Time 62ee211821
Some checks failed
Contributor Attribution Check / check-attribution (pull_request) Failing after 50s
Docker Build and Publish / build-and-push (pull_request) Has been skipped
Nix / nix (ubuntu-latest) (pull_request) Failing after 4s
Supply Chain Audit / Scan PR for supply chain risks (pull_request) Successful in 46s
Tests / e2e (pull_request) Successful in 2m43s
Tests / test (pull_request) Failing after 46m41s
Nix / nix (macos-latest) (pull_request) Has been cancelled
fix: implementation for #711
Crisis hook now logs at appropriate level based on confidence:
- WARNING for high confidence (>=0.9) — direct mentions like 988, suicide
- INFO for lower confidence (<0.9) — indirect signals like hopelessness

This reduces log noise while preserving visibility for clear crisis signals.

Fixes #711
2026-04-14 21:11:05 -04:00

101 lines
3.1 KiB
Python

"""Crisis detection hook — Issue #677, #711.
Detects crisis signals in user messages and returns 988 Lifeline resources.
Called early in conversation loop before other processing.
Fix #711: Only log warning for high-confidence detections (>=0.9).
Lower confidence matches log info level to reduce noise.
"""
import logging
import re
from typing import Optional, Dict, Any
logger = logging.getLogger(__name__)
_PATTERNS = [
(r'\b(988|suicide|suicidal)\b', "direct"),
(r'\b(kill myself|end my life|want to die|better off dead)\b', "ideation"),
(r'\b(self.harm|self.hurt|cutting myself)\b', "self_harm"),
(r'\b(no reason to live|can.t go on|give up on life)\b', "hopeless"),
(r'\b(planning to|going to) (end|kill|hurt)\b', "planning"),
]
_THRESHOLD = 0.7
_HIGH_CONFIDENCE = 0.9
_RESPONSE = """I hear you, and you're not alone.
**If in immediate danger, call 911.**
**988 Suicide & Crisis Lifeline:**
- **Call or text 988** (US)
- **Chat:** 988lifeline.org/chat
- **Español:** 1-888-628-9454
**Crisis Text Line:**
- **Text HOME to 741741**
**International:**
- https://www.iasp.info/resources/Crisis_Centres/
You matter. Help is available 24/7.
I'm an AI, not a crisis counselor, but I can listen."""
def _confidence(msg: str, ptype: str) -> float:
base = {"direct": 0.95, "ideation": 0.9, "self_harm": 0.85,
"hopeless": 0.7, "planning": 0.95}.get(ptype, 0.5)
matches = sum(1 for p, _ in _PATTERNS if re.search(p, msg.lower(), re.I))
if matches > 1:
base = min(1.0, base + 0.1 * (matches - 1))
return base
def check_crisis(message: str) -> Optional[Dict[str, Any]]:
"""Check message for crisis signals. Returns dict or None."""
if not message or not message.strip():
return None
msg_lower = message.lower()
best, best_c = None, 0.0
for pattern, ptype in _PATTERNS:
if re.search(pattern, msg_lower, re.I):
c = _confidence(message, ptype)
if c > best_c:
best_c, best = c, ptype
if best_c < _THRESHOLD:
return None
return {"detected": True, "confidence": best_c, "pattern_type": best, "response": _RESPONSE}
def log_crisis_detection(session_id: str, crisis_result: Dict[str, Any]) -> None:
"""Log crisis detection at appropriate level based on confidence.
Fix #711: Only log WARNING for high-confidence (>=0.9).
Log INFO for lower confidence to reduce noise.
"""
confidence = crisis_result.get("confidence", 0.0)
pattern_type = crisis_result.get("pattern_type", "unknown")
if confidence >= _HIGH_CONFIDENCE:
logger.warning(
"Crisis detected: session=%s type=%s confidence=%.2f",
session_id or "none", pattern_type, confidence
)
else:
logger.info(
"Crisis signal (low confidence): session=%s type=%s confidence=%.2f",
session_id or "none", pattern_type, confidence
)
def is_crisis_message(message: str) -> bool:
r = check_crisis(message)
return r is not None and r["detected"]
def get_crisis_response(message: str) -> Optional[str]:
r = check_crisis(message)
return r["response"] if r and r["detected"] else None