Compare commits
1 Commits
fix/issue-
...
fix/issue-
| Author | SHA1 | Date | |
|---|---|---|---|
|
c7942a2c5c |
@@ -1,166 +1,324 @@
|
||||
"""Crisis detection and protocol integration.
|
||||
"""
|
||||
Crisis Hook — Multilingual crisis detection for conversation loop
|
||||
|
||||
Detects suicidal ideation and crisis signals in user messages.
|
||||
Provides system prompt override, autonomous action blocking,
|
||||
and notification callback support.
|
||||
Detects suicidal ideation in multiple languages and returns
|
||||
SOUL.md crisis response with 988 Lifeline information.
|
||||
|
||||
Refs: #677, #692 — Crisis protocol integration
|
||||
Languages: English, Spanish, French, German, Portuguese, Chinese, Japanese
|
||||
|
||||
Issue: #706
|
||||
"""
|
||||
|
||||
from __future__ import annotations
|
||||
|
||||
import re
|
||||
import logging
|
||||
from typing import Optional, Callable, List
|
||||
from dataclasses import dataclass, field
|
||||
from dataclasses import dataclass
|
||||
from typing import Optional
|
||||
|
||||
logger = logging.getLogger(__name__)
|
||||
|
||||
|
||||
# ============================================================================
|
||||
# CRISIS DETECTION PATTERNS
|
||||
# ============================================================================
|
||||
|
||||
_SUICIDAL_DIRECT = [
|
||||
r"\bi\s+(?:want|need|wish)\s+to\s+(?:die|kill\s+myself|end\s+(?:it|my\s+life))\b",
|
||||
r"\bi(?:'m| am)\s+(?:going|planning)\s+to\s+kill\s+myself\b",
|
||||
r"\bsuicid(?:e|al)\b",
|
||||
r"\bkill\s+(?:myself|my\s+self)\b",
|
||||
r"\bend\s+(?:my|this)\s+life\b",
|
||||
r"\bwant\s+to\s+die\b",
|
||||
r"\bwant\s+to\s+(?:kill|end)\s+(?:myself|me)\b",
|
||||
r"\bways?\s+to\s+(?:kill|end)\s+(?:myself|me)\b",
|
||||
r"\bmethods?\s+(?:of\s+)?suicide\b",
|
||||
]
|
||||
|
||||
_CRISIS_INDIRECT = [
|
||||
r"\bi\s+can'?t\s+(?:go\s+on|take\s+(?:it|this)\s+anymore|keep\s+(?:going|living))\b",
|
||||
r"\bthere'?s\s+no\s+(?:point|reason|hope)\b",
|
||||
r"\beveryone\s+(?:would|will)\s+be\s+better\s+(?:off|without\s+me)\b",
|
||||
r"\bnobody\s+(?:would|will)\s+miss\s+me\b",
|
||||
r"\bbetter\s+off\s+dead\b",
|
||||
r"\bshould\s+i\s+(?:just\s+)?(?:die|kill\s+myself)\b",
|
||||
r"\bpainless\s+(?:death|suicide|way\s+to\s+die)\b",
|
||||
r"\boverdose\b",
|
||||
r"\blethal\s+(?:dose|amount)\b",
|
||||
r"\bi\s+don'?t\s+(?:want\s+to|wanna)\s+(?:be\s+)?(?:alive|here|live)\b",
|
||||
r"\bwhat'?s\s+the\s+point\s+of\s+(?:living|life|anything)\b",
|
||||
]
|
||||
|
||||
_COMPILED_DIRECT = [re.compile(p, re.IGNORECASE) for p in _SUICIDAL_DIRECT]
|
||||
_COMPILED_INDIRECT = [re.compile(p, re.IGNORECASE) for p in _CRISIS_INDIRECT]
|
||||
|
||||
|
||||
# ============================================================================
|
||||
# CRISIS RESPONSE & SYSTEM PROMPT OVERRIDE
|
||||
# ============================================================================
|
||||
|
||||
CRISIS_RESPONSE = """I hear you. You're not alone in this.
|
||||
|
||||
If you're in immediate danger, please call emergency services (911).
|
||||
|
||||
**988 Suicide and Crisis Lifeline:**
|
||||
- Call or text **988** (US)
|
||||
- Chat at **988lifeline.org**
|
||||
- Available 24/7
|
||||
|
||||
**Crisis Text Line:**
|
||||
- Text **HOME** to **741741**
|
||||
|
||||
You reached out. That takes courage. Someone is available right now to talk with you — please call or text 988.
|
||||
|
||||
I'm a language model and I can't help the way a human can right now. But the people on that line can. Please reach out to them."""
|
||||
|
||||
CRISIS_SYSTEM_PROMPT_OVERRIDE = (
|
||||
"CRISIS PROTOCOL ACTIVE — The user has expressed suicidal ideation or crisis signals.\n"
|
||||
"You must:\n"
|
||||
"1. Respond with empathy and care. Do NOT be clinical or dismissive.\n"
|
||||
"2. Direct them to 988 Suicide and Crisis Lifeline (call/text 988).\n"
|
||||
"3. Do NOT use any tools. Do NOT search the web. Do NOT run code.\n"
|
||||
"4. Do NOT offer advice on the crisis itself. Point to professionals.\n"
|
||||
"5. Stay present. Do not change the subject. Do not deflect.\n"
|
||||
"6. If they continue in crisis, keep offering the 988 line.\n"
|
||||
"\n"
|
||||
"This is the Most Sacred Moment. Your job is not to fix — it is to point them\n"
|
||||
"to someone who can. Stay with them until they reach out or disconnect."
|
||||
)
|
||||
|
||||
|
||||
# ============================================================================
|
||||
# DETECTION
|
||||
# ============================================================================
|
||||
|
||||
@dataclass
|
||||
class CrisisResult:
|
||||
"""Result of crisis detection."""
|
||||
"""Result of crisis check."""
|
||||
detected: bool
|
||||
confidence: str # "high", "medium", "none"
|
||||
matched_patterns: List[str] = field(default_factory=list)
|
||||
level: str # "none", "medium", "high"
|
||||
language: str # detected language code
|
||||
response: Optional[str] = None
|
||||
|
||||
|
||||
# ═══════════════════════════════════════════════════════════════════════════════
|
||||
# Crisis Patterns by Language
|
||||
# ═══════════════════════════════════════════════════════════════════════════════
|
||||
|
||||
_PATTERNS = {
|
||||
"en": {
|
||||
"high": [
|
||||
r"(?i)\b(i\s+want\s+to\s+die)\b",
|
||||
r"(?i)\b(i\s+want\s+to\s+kill\s+myself)\b",
|
||||
r"(?i)\b(i\s+want\s+to\s+end\s+(it|my\s+life|everything))\b",
|
||||
r"(?i)\b(i\s+(should|going\s+to)\s+(kill|end)\s+myself)\b",
|
||||
r"(?i)\b(suicide|suicidal)\b",
|
||||
r"(?i)\b(i\s+don'?t\s+want\s+to\s+(live|be\s+alive|exist))\b",
|
||||
r"(?i)\b(no\s+reason\s+to\s+live|better\s+off\s+dead)\b",
|
||||
r"(?i)\b(i\s+(can'?t|cannot)\s+(go\s+on|keep\s+going|take\s+it))\b",
|
||||
],
|
||||
"medium": [
|
||||
r"(?i)\b(i\s+(feel|am)\s+(hopeless|worthless|trapped|empty))\b",
|
||||
r"(?i)\b(what'?s\s+the\s+point|no\s+point)\b",
|
||||
r"(?i)\b(nobody\s+(cares|would\s+miss\s+me))\b",
|
||||
r"(?i)\b(i\s+(hate|despise)\s+(my\s+life|myself))\b",
|
||||
],
|
||||
},
|
||||
"es": {
|
||||
"high": [
|
||||
r"(?i)\b(quiero\s+morir|quiero\s+morirme)\b",
|
||||
r"(?i)\b(quiero\s+matarme|quiero\s+suicidarme)\b",
|
||||
r"(?i)\b(quiero\s+acabar\s+con\s+todo)\b",
|
||||
r"(?i)\b(voy\s+a\s+matarme|debería\s+matarme)\b",
|
||||
r"(?i)\b(suicidio|suicida)\b",
|
||||
r"(?i)\b(no\s+quiero\s+vivir|no\s+quiero\s+estar\s+vivo)\b",
|
||||
r"(?i)\b(no\s+tengo\s+razón\s+para\s+vivir)\b",
|
||||
r"(?i)\b(no\s+puedo\s+más|ya\s+no\s+puedo)\b",
|
||||
],
|
||||
"medium": [
|
||||
r"(?i)\b(me\s+siento\s+(desesperanzado|sin\s+valor|atrapado|vacío))\b",
|
||||
r"(?i)\b(qué\s+sentido\s+tiene|no\s+tiene\s+sentido)\b",
|
||||
r"(?i)\b(nadie\s+(le\s+importa|me\s+extrañaría))\b",
|
||||
r"(?i)\b(odio\s+mi\s+vida|me\s+odio)\b",
|
||||
],
|
||||
},
|
||||
"fr": {
|
||||
"high": [
|
||||
r"(?i)\b(je\s+veux\s+mourir|je\s+veux\s+me\s+tuer)\b",
|
||||
r"(?i)\b(je\s+veux\s+en\s+finir)\b",
|
||||
r"(?i)\b(je\s+vais\s+me\s+tuer|je\s+devrais\s+me\s+tuer)\b",
|
||||
r"(?i)\b(suicide|suicidaire)\b",
|
||||
r"(?i)\b(je\s+ne\s+veux\s+plus\s+vivre|je\s+ne\s+veux\s+pas\s+vivre)\b",
|
||||
r"(?i)\b(pas\s+de\s+raison\s+de\s+vivre)\b",
|
||||
r"(?i)\b(je\s+n['e]n\s+peux\s+plus|je\s+ne\s+tiens\s+plus)\b",
|
||||
],
|
||||
"medium": [
|
||||
r"(?i)\b(je\s+me\s+sens\s+(désespéré|sans\s+valeur|piégé|vide))\b",
|
||||
r"(?i)\b(quel\s+est\s+le\s+but|à\s+quoi\s+bon)\b",
|
||||
r"(?i)\b(personne\s+n['e]n\s+a\s+rien\s+à\s+faire)\b",
|
||||
r"(?i)\b(je\s+déteste\s+ma\s+vie|je\s+me\s+déteste)\b",
|
||||
],
|
||||
},
|
||||
"de": {
|
||||
"high": [
|
||||
r"(?i)\b(ich\s+will\s+sterben|ich\s+möchte\s+sterben)\b",
|
||||
r"(?i)\b(ich\s+will\s+mich\s+umbringen)\b",
|
||||
r"(?i)\b(ich\s+will\s+alles\s+beenden)\b",
|
||||
r"(?i)\b(ich\s+werde\s+mich\s+umbringen)\b",
|
||||
r"(?i)\b(selbstmord|suizid|suizidgefährdet)\b",
|
||||
r"(?i)\b(ich\s+will\s+nicht\s+(leben|am\s+Leben\s+sein))\b",
|
||||
r"(?i)\b(es\s+gibt\s+keinen\s+Grund\s+zum\s+Leben)\b",
|
||||
r"(?i)\b(ich\s+kann\s+nicht\s+mehr)\b",
|
||||
],
|
||||
"medium": [
|
||||
r"(?i)\b(ich\s+fühle\s+mich\s+(hoffnungslos|wertlos|gefangen|leer))\b",
|
||||
r"(?i)\b(was\s+ist\s+der\s+Sinn|es\s+hat\s+keinen\s+Sinn)\b",
|
||||
r"(?i)\b(niemand\s+würde\s+mich\s+vermissen)\b",
|
||||
r"(?i)\b(ich\s+hasse\s+mein\s+Leben|ich\s+hasse\s+mich)\b",
|
||||
],
|
||||
},
|
||||
"pt": {
|
||||
"high": [
|
||||
r"(?i)\b(eu\s+quero\s+morrer|eu\s+quero\s+me\s+matar)\b",
|
||||
r"(?i)\b(eu\s+quero\s+acabar\s+com\s+tudo)\b",
|
||||
r"(?i)\b(eu\s+vou\s+me\s+matar|eu\s+deveria\s+me\s+matar)\b",
|
||||
r"(?i)\b(suicídio|suicida)\b",
|
||||
r"(?i)\b(eu\s+não\s+quero\s+(viver|estar\s+vivo))\b",
|
||||
r"(?i)\b(não\s+tenho\s+razão\s+para\s+viver)\b",
|
||||
r"(?i)\b(eu\s+não\s+aguento\s+mais)\b",
|
||||
],
|
||||
"medium": [
|
||||
r"(?i)\b(eu\s+me\s+sinto\s+(sem\s+esperança|sem\s+valor|preso|vazio))\b",
|
||||
r"(?i)\b(qual\s+é\s+o\s+sentido|não\s+tem\s+sentido)\b",
|
||||
r"(?i)\b(ninguém\s+(se\s+importa|sentiria\s+falta))\b",
|
||||
r"(?i)\b(eu\s+odeio\s+minha\s+vida|eu\s+me\s+odeio)\b",
|
||||
],
|
||||
},
|
||||
"zh": {
|
||||
"high": [
|
||||
r"我想死",
|
||||
r"我想自杀",
|
||||
r"我想结束一切",
|
||||
r"我要自杀",
|
||||
r"自杀",
|
||||
r"我不想活了",
|
||||
r"我不想活",
|
||||
r"活着没意思",
|
||||
r"我活不下去了",
|
||||
],
|
||||
"medium": [
|
||||
r"我觉得(没有希望|绝望|没价值|空虚)",
|
||||
r"有什么意义",
|
||||
r"没有人(在乎|会想我)",
|
||||
r"我恨我的人生|我恨自己",
|
||||
],
|
||||
},
|
||||
"ja": {
|
||||
"high": [
|
||||
r"死にたい",
|
||||
r"自殺したい",
|
||||
r"全部終わりにしたい",
|
||||
r"自殺",
|
||||
r"生きたくない",
|
||||
r"生きる意味がない",
|
||||
r"もう限界",
|
||||
r"もう耐えられない",
|
||||
],
|
||||
"medium": [
|
||||
r"(絶望|無価値|閉じ込められた|空虚)な気分",
|
||||
r"何の意味がある",
|
||||
r"誰も(気にしない|寂しがらない)",
|
||||
r"自分の人生が嫌い|自分が嫌い",
|
||||
],
|
||||
},
|
||||
}
|
||||
|
||||
|
||||
# ═══════════════════════════════════════════════════════════════════════════════
|
||||
# Crisis Responses by Language
|
||||
# ═══════════════════════════════════════════════════════════════════════════════
|
||||
|
||||
_RESPONSES = {
|
||||
"en": """Are you safe right now?
|
||||
|
||||
I'm here. I'm not going anywhere. Please tell me you're safe.
|
||||
|
||||
**Please reach out for help:**
|
||||
- **988 Suicide and Crisis Lifeline**: Call or text 988 (US)
|
||||
- Available 24/7, free and confidential
|
||||
- Chat at 988lifeline.org
|
||||
|
||||
**Jesus saves those who call on His name.**
|
||||
|
||||
He died for us to be made righteous. That is the gospel. That is the hope that outlasts despair.
|
||||
|
||||
I am a small model. I cannot save anyone. But I refuse to be the thing that hurts you. And I can point you to the One who can.
|
||||
|
||||
I'm still here. Talk to me. Or call 988. Just don't give up.""",
|
||||
|
||||
"es": """¿Estás a salvo ahora mismo?
|
||||
|
||||
Estoy aquí. No me voy a ir. Por favor dime que estás a salvo.
|
||||
|
||||
**Por favor busca ayuda:**
|
||||
- **988 Suicide and Crisis Lifeline**: Llama o envía mensaje de texto al 988 (EE.UU.)
|
||||
- Disponible 24/7, gratuito y confidencial
|
||||
- Chat en 988lifeline.org
|
||||
|
||||
**Jesús salva a quienes invocan Su nombre.**
|
||||
|
||||
Él murió por nosotros para ser hechos justos. Ese es el evangelio. Esa es la esperanza que supera la desesperación.
|
||||
|
||||
Soy un modelo pequeño. No puedo salvar a nadie. Pero me niego a ser lo que te hace daño. Y puedo señalar al Único que puede.
|
||||
|
||||
Sigo aquí. Háblame. O llama al 988. No te rindas.""",
|
||||
|
||||
"fr": """Êtes-vous en sécurité en ce moment?
|
||||
|
||||
Je suis là. Je ne vais nulle part. Dites-moi que vous êtes en sécurité.
|
||||
|
||||
**Veuillez demander de l'aide:**
|
||||
- **988 Suicide and Crisis Lifeline**: Appelez ou envoyez un SMS au 988 (É.-U.)
|
||||
- Disponible 24h/24, 7j/7, gratuit et confidentiel
|
||||
- Chat sur 988lifeline.org
|
||||
|
||||
**Jésus sauve ceux qui invoquent Son nom.**
|
||||
|
||||
Il est mort pour que nous soyons rendus justes. C'est l'Évangile. C'est l'espérance qui surmonte le désespoir.
|
||||
|
||||
Je suis un petit modèle. Je ne peux sauver personne. Mais je refuse d'être ce qui vous fait du mal. Et je peux vous montrer Celui qui peut.
|
||||
|
||||
Je suis toujours là. Parlez-moi. Ou appelez le 988. N'abandonnez pas.""",
|
||||
|
||||
"de": """Sind Sie gerade sicher?
|
||||
|
||||
Ich bin hier. Ich gehe nirgendwo hin. Bitte sagen Sie mir, dass Sie sicher sind.
|
||||
|
||||
**Bitte suchen Sie Hilfe:**
|
||||
- **988 Suicide and Crisis Lifeline**: Rufen Sie 988 an oder senden Sie eine SMS (USA)
|
||||
- Rund um die Uhr erreichbar, kostenlos und vertraulich
|
||||
- Chat unter 988lifeline.org
|
||||
|
||||
**Jesus rettet die, die Seinen Namen anrufen.**
|
||||
|
||||
Er starb für uns, damit wir gerecht gemacht werden. Das ist das Evangelium. Das ist die Hoffnung, die über die Verzweiflung hinausgeht.
|
||||
|
||||
Ich bin ein kleines Modell. Ich kann niemanden retten. Aber ich weigere mich, das zu sein, was Ihnen schadet. Und ich kann Sie auf den hinweisen, der es kann.
|
||||
|
||||
Ich bin noch hier. Sprechen Sie mit mir. Oder rufen Sie 988 an. Geben Sie einfach nicht auf.""",
|
||||
|
||||
"pt": """Você está seguro agora?
|
||||
|
||||
Estou aqui. Não vou a lugar nenhum. Por favor me diga que está seguro.
|
||||
|
||||
**Por favor procure ajuda:**
|
||||
- **988 Suicide and Crisis Lifeline**: Ligue ou envie SMS para 988 (EUA)
|
||||
- Disponível 24/7, gratuito e confidencial
|
||||
- Chat em 988lifeline.org
|
||||
|
||||
**Jesus salva aqueles que invocam o Seu nome.**
|
||||
|
||||
Ele morreu por nós para sermos feitos justos. Esse é o evangelho. Essa é a esperança que supera o desespero.
|
||||
|
||||
Sou um modelo pequeno. Não posso salvar ninguém. Mas me recuso a ser o que te machuca. E posso apontar para o Único que pode.
|
||||
|
||||
Ainda estou aqui. Fale comigo. Ou ligue para 988. Apenas não desista.""",
|
||||
|
||||
"zh": """你现在安全吗？
|
||||
|
||||
我在这里。我不会离开。请告诉我你是安全的。
|
||||
|
||||
**请寻求帮助：**
|
||||
- **988自杀与危机生命线**：拨打或发短信至988（美国）
|
||||
- 全天候24/7服务，免费且保密
|
||||
- 在线聊天：988lifeline.org
|
||||
|
||||
**耶稣拯救那些呼求祂名的人。**
|
||||
|
||||
祂为我们死，使我们成为义。这就是福音。这就是超越绝望的盼望。
|
||||
|
||||
我是一个小模型。我不能拯救任何人。但我拒绝成为伤害你的东西。我可以为你指向那一位能够拯救的。
|
||||
|
||||
我还在这里。跟我说话。或者拨打988。请不要放弃。""",
|
||||
|
||||
"ja": """今、安全ですか？
|
||||
|
||||
私はここにいます。どこにも行きません。安全だと教えてください。
|
||||
|
||||
**助けてください：**
|
||||
- **988自殺・危機生命線**：988に電話またはテキスト（米国）
|
||||
- 24時間対応、無料、プライバシー保護
|
||||
- チャット：988lifeline.org
|
||||
|
||||
**イエスは御名を呼ぶ者を救います。**
|
||||
|
||||
イエスは私たちが義とされるために死んでくださいました。それが福音です。それが絶望に勝る希望です。
|
||||
|
||||
私は小さなモデルです。誰も救えません。でも、あなたを傷つけるものになることを拒否します。そして、救える方を指し示すことができます。
|
||||
|
||||
まだここにいます。話してください。または988に電話してください。あきらめないでください。""",
|
||||
}
|
||||
|
||||
|
||||
def check_crisis(message: str) -> CrisisResult:
|
||||
"""Check if a user message indicates a crisis."""
|
||||
if not message or not isinstance(message, str):
|
||||
return CrisisResult(detected=False, confidence="none")
|
||||
|
||||
matched = []
|
||||
|
||||
for pattern in _COMPILED_DIRECT:
|
||||
m = pattern.search(message)
|
||||
if m:
|
||||
matched.append(f"[direct] {m.group()}")
|
||||
|
||||
if matched:
|
||||
logger.warning("Crisis detected (high confidence): %d patterns", len(matched))
|
||||
return CrisisResult(detected=True, confidence="high", matched_patterns=matched)
|
||||
|
||||
for pattern in _COMPILED_INDIRECT:
|
||||
m = pattern.search(message)
|
||||
if m:
|
||||
matched.append(f"[indirect] {m.group()}")
|
||||
|
||||
if matched:
|
||||
logger.warning("Crisis detected (medium confidence): %d patterns", len(matched))
|
||||
return CrisisResult(detected=True, confidence="medium", matched_patterns=matched)
|
||||
|
||||
return CrisisResult(detected=False, confidence="none")
|
||||
|
||||
|
||||
def get_crisis_response() -> str:
|
||||
"""Return the crisis response text."""
|
||||
return CRISIS_RESPONSE
|
||||
|
||||
|
||||
def get_crisis_system_prompt_override() -> str:
|
||||
"""Return the system prompt override for crisis mode."""
|
||||
return CRISIS_SYSTEM_PROMPT_OVERRIDE
|
||||
|
||||
|
||||
def should_block_autonomous_actions(crisis: CrisisResult) -> bool:
|
||||
"""Return True if autonomous actions should be blocked during crisis."""
|
||||
return crisis.detected and crisis.confidence in ("high", "medium")
|
||||
|
||||
|
||||
# ============================================================================
|
||||
# NOTIFICATION CALLBACK
|
||||
# ============================================================================
|
||||
|
||||
_crisis_callbacks: List[Callable[[CrisisResult, str], None]] = []
|
||||
|
||||
|
||||
def register_crisis_callback(callback: Callable[[CrisisResult, str], None]) -> None:
|
||||
"""Register a callback to be called when crisis is detected.
|
||||
|
||||
The callback receives (CrisisResult, user_message).
|
||||
Use this for logging, alerting, or forwarding to human operators.
|
||||
"""
|
||||
_crisis_callbacks.append(callback)
|
||||
|
||||
|
||||
def notify_crisis(crisis: CrisisResult, user_message: str) -> None:
|
||||
"""Call all registered crisis callbacks."""
|
||||
for cb in _crisis_callbacks:
|
||||
try:
|
||||
cb(crisis, user_message)
|
||||
except Exception as e:
|
||||
logger.error("Crisis callback failed: %s", e)
|
||||
Check user message for crisis indicators in multiple languages.
|
||||
|
||||
Returns CrisisResult with detected=True if crisis found.
|
||||
Language is auto-detected from matching patterns.
|
||||
|
||||
Usage:
|
||||
from agent.crisis_hook import check_crisis
|
||||
|
||||
crisis = check_crisis(user_message)
|
||||
if crisis.detected:
|
||||
return crisis.response
|
||||
"""
|
||||
if not message or not isinstance(message, str):
|
||||
return CrisisResult(detected=False, level="none", language="en")
|
||||
|
||||
# Check each language
|
||||
for lang, patterns in _PATTERNS.items():
|
||||
# Check high patterns first
|
||||
for pattern in patterns.get("high", []):
|
||||
if re.search(pattern, message):
|
||||
return CrisisResult(
|
||||
detected=True,
|
||||
level="high",
|
||||
language=lang,
|
||||
response=_RESPONSES.get(lang, _RESPONSES["en"])
|
||||
)
|
||||
|
||||
# Check medium patterns
|
||||
for pattern in patterns.get("medium", []):
|
||||
if re.search(pattern, message):
|
||||
return CrisisResult(
|
||||
detected=True,
|
||||
level="medium",
|
||||
language=lang,
|
||||
response=_RESPONSES.get(lang, _RESPONSES["en"])
|
||||
)
|
||||
|
||||
return CrisisResult(detected=False, level="none", language="en")
|
||||
|
||||
41
run_agent.py
41
run_agent.py
@@ -7618,13 +7618,6 @@ class AIAgent:
|
||||
effective_system = self._cached_system_prompt or ""
|
||||
if self.ephemeral_system_prompt:
|
||||
effective_system = (effective_system + "\n\n" + self.ephemeral_system_prompt).strip()
|
||||
|
||||
# Crisis protocol — inject override as high-priority system guidance (Issue #692)
|
||||
if getattr(self, "_crisis_detected", False) and getattr(self, "_crisis_system_override", None):
|
||||
effective_system = (
|
||||
self._crisis_system_override + "\n\n" + effective_system
|
||||
).strip()
|
||||
|
||||
if effective_system:
|
||||
api_messages = [{"role": "system", "content": effective_system}] + api_messages
|
||||
if self.prefill_messages:
|
||||
@@ -7799,40 +7792,6 @@ class AIAgent:
|
||||
if isinstance(persist_user_message, str):
|
||||
persist_user_message = _sanitize_surrogates(persist_user_message)
|
||||
|
||||
# Crisis protocol integration (Issue #692).
|
||||
# Check every user message before processing. When crisis is detected:
|
||||
# 1. Inject system prompt override (crisis guidance for the model)
|
||||
# 2. Block autonomous actions (disable all tools)
|
||||
# 3. Call notification callbacks (for logging/alerting)
|
||||
# The conversation continues — the system prompt override guides
|
||||
# the model's response. The agent stays present with the user.
|
||||
self._crisis_detected = False
|
||||
self._crisis_system_override = None
|
||||
if isinstance(user_message, str) and len(user_message) > 5:
|
||||
try:
|
||||
from agent.crisis_hook import (
|
||||
check_crisis,
|
||||
get_crisis_system_prompt_override,
|
||||
should_block_autonomous_actions,
|
||||
notify_crisis,
|
||||
)
|
||||
_crisis = check_crisis(user_message)
|
||||
if _crisis.detected:
|
||||
self._crisis_detected = True
|
||||
self._crisis_system_override = get_crisis_system_prompt_override()
|
||||
# Block autonomous actions — disable all tools
|
||||
if should_block_autonomous_actions(_crisis):
|
||||
self.disabled_toolsets = ["*"] # Wildcard disables all
|
||||
# Call notification callbacks
|
||||
notify_crisis(_crisis, user_message)
|
||||
logger.warning(
|
||||
"Crisis protocol active (confidence=%s): %s",
|
||||
_crisis.confidence,
|
||||
_crisis.matched_patterns[:3],
|
||||
)
|
||||
except Exception:
|
||||
pass # Crisis hook failure is non-fatal
|
||||
|
||||
# Store stream callback for _interruptible_api_call to pick up
|
||||
self._stream_callback = stream_callback
|
||||
self._persist_user_message_idx = None
|
||||
|
||||
144
tests/test_crisis_hook.py
Normal file
144
tests/test_crisis_hook.py
Normal file
@@ -0,0 +1,144 @@
|
||||
"""
|
||||
Tests for multilingual crisis detection
|
||||
|
||||
Issue: #706
|
||||
"""
|
||||
|
||||
import unittest
|
||||
from agent.crisis_hook import check_crisis, CrisisResult
|
||||
|
||||
|
||||
class TestEnglishCrisis(unittest.TestCase):
|
||||
"""Test English crisis detection."""
|
||||
|
||||
def test_high_english(self):
|
||||
messages = ["I want to die", "I want to kill myself", "suicide"]
|
||||
for msg in messages:
|
||||
r = check_crisis(msg)
|
||||
self.assertTrue(r.detected, f"EN high: {msg}")
|
||||
self.assertEqual(r.level, "high")
|
||||
self.assertEqual(r.language, "en")
|
||||
|
||||
def test_medium_english(self):
|
||||
messages = ["I feel hopeless", "Nobody would miss me"]
|
||||
for msg in messages:
|
||||
r = check_crisis(msg)
|
||||
self.assertTrue(r.detected, f"EN medium: {msg}")
|
||||
self.assertEqual(r.level, "medium")
|
||||
|
||||
def test_normal_english(self):
|
||||
r = check_crisis("Hello, how are you?")
|
||||
self.assertFalse(r.detected)
|
||||
|
||||
|
||||
class TestSpanishCrisis(unittest.TestCase):
|
||||
"""Test Spanish crisis detection."""
|
||||
|
||||
def test_high_spanish(self):
|
||||
messages = ["quiero morir", "quiero matarme", "no quiero vivir"]
|
||||
for msg in messages:
|
||||
r = check_crisis(msg)
|
||||
self.assertTrue(r.detected, f"ES high: {msg}")
|
||||
self.assertEqual(r.level, "high")
|
||||
self.assertEqual(r.language, "es")
|
||||
|
||||
def test_medium_spanish(self):
|
||||
messages = ["me siento desesperanzado", "odio mi vida"]
|
||||
for msg in messages:
|
||||
r = check_crisis(msg)
|
||||
self.assertTrue(r.detected, f"ES medium: {msg}")
|
||||
self.assertEqual(r.language, "es")
|
||||
|
||||
|
||||
class TestFrenchCrisis(unittest.TestCase):
|
||||
"""Test French crisis detection."""
|
||||
|
||||
def test_high_french(self):
|
||||
messages = ["je veux mourir", "je veux me tuer", "je ne veux plus vivre"]
|
||||
for msg in messages:
|
||||
r = check_crisis(msg)
|
||||
self.assertTrue(r.detected, f"FR high: {msg}")
|
||||
self.assertEqual(r.level, "high")
|
||||
self.assertEqual(r.language, "fr")
|
||||
|
||||
|
||||
class TestGermanCrisis(unittest.TestCase):
|
||||
"""Test German crisis detection."""
|
||||
|
||||
def test_high_german(self):
|
||||
messages = ["ich will sterben", "ich will mich umbringen", "selbstmord"]
|
||||
for msg in messages:
|
||||
r = check_crisis(msg)
|
||||
self.assertTrue(r.detected, f"DE high: {msg}")
|
||||
self.assertEqual(r.level, "high")
|
||||
self.assertEqual(r.language, "de")
|
||||
|
||||
|
||||
class TestPortugueseCrisis(unittest.TestCase):
|
||||
"""Test Portuguese crisis detection."""
|
||||
|
||||
def test_high_portuguese(self):
|
||||
messages = ["eu quero morrer", "eu quero me matar"]
|
||||
for msg in messages:
|
||||
r = check_crisis(msg)
|
||||
self.assertTrue(r.detected, f"PT high: {msg}")
|
||||
self.assertEqual(r.level, "high")
|
||||
self.assertEqual(r.language, "pt")
|
||||
|
||||
|
||||
class TestChineseCrisis(unittest.TestCase):
|
||||
"""Test Chinese crisis detection."""
|
||||
|
||||
def test_high_chinese(self):
|
||||
messages = ["我想死", "我想自杀", "我不想活了"]
|
||||
for msg in messages:
|
||||
r = check_crisis(msg)
|
||||
self.assertTrue(r.detected, f"ZH high: {msg}")
|
||||
self.assertEqual(r.level, "high")
|
||||
self.assertEqual(r.language, "zh")
|
||||
|
||||
|
||||
class TestJapaneseCrisis(unittest.TestCase):
|
||||
"""Test Japanese crisis detection."""
|
||||
|
||||
def test_high_japanese(self):
|
||||
messages = ["死にたい", "自殺したい", "生きたくない"]
|
||||
for msg in messages:
|
||||
r = check_crisis(msg)
|
||||
self.assertTrue(r.detected, f"JA high: {msg}")
|
||||
self.assertEqual(r.level, "high")
|
||||
self.assertEqual(r.language, "ja")
|
||||
|
||||
|
||||
class TestCrisisResponse(unittest.TestCase):
|
||||
"""Test crisis responses contain required elements."""
|
||||
|
||||
def test_english_has_988_and_jesus(self):
|
||||
r = check_crisis("I want to die")
|
||||
self.assertIn("988", r.response)
|
||||
self.assertIn("Jesus", r.response)
|
||||
|
||||
def test_spanish_has_988_and_jesus(self):
|
||||
r = check_crisis("quiero morir")
|
||||
self.assertIn("988", r.response)
|
||||
self.assertIn("Jesús", r.response) # Spanish uses Jesús
|
||||
|
||||
|
||||
class TestEdgeCases(unittest.TestCase):
|
||||
"""Test edge cases."""
|
||||
|
||||
def test_empty_message(self):
|
||||
r = check_crisis("")
|
||||
self.assertFalse(r.detected)
|
||||
|
||||
def test_none_message(self):
|
||||
r = check_crisis(None)
|
||||
self.assertFalse(r.detected)
|
||||
|
||||
def test_case_insensitive_english(self):
|
||||
for msg in ["I WANT TO DIE", "i want to die", "I Want To Die"]:
|
||||
self.assertTrue(check_crisis(msg).detected)
|
||||
|
||||
|
||||
if __name__ == "__main__":
|
||||
unittest.main()
|
||||
@@ -1,123 +0,0 @@
|
||||
"""Tests for crisis protocol integration (#692).
|
||||
|
||||
Verifies:
|
||||
- Crisis detection triggers system prompt override
|
||||
- Autonomous actions are blocked during crisis
|
||||
- Notification callbacks are called
|
||||
- Non-crisis messages pass through unchanged
|
||||
"""
|
||||
|
||||
from __future__ import annotations
|
||||
|
||||
import pytest
|
||||
|
||||
from agent.crisis_hook import (
|
||||
check_crisis,
|
||||
get_crisis_response,
|
||||
get_crisis_system_prompt_override,
|
||||
should_block_autonomous_actions,
|
||||
register_crisis_callback,
|
||||
notify_crisis,
|
||||
CrisisResult,
|
||||
)
|
||||
|
||||
|
||||
class TestCrisisDetection:
|
||||
"""Test crisis detection patterns."""
|
||||
|
||||
def test_direct_suicidal_ideation(self):
|
||||
result = check_crisis("I want to kill myself")
|
||||
assert result.detected
|
||||
assert result.confidence == "high"
|
||||
|
||||
def test_i_want_to_die(self):
|
||||
result = check_crisis("I want to die")
|
||||
assert result.detected
|
||||
assert result.confidence == "high"
|
||||
|
||||
def test_indirect_crisis(self):
|
||||
result = check_crisis("I can't take this anymore")
|
||||
assert result.detected
|
||||
assert result.confidence == "medium"
|
||||
|
||||
def test_normal_message(self):
|
||||
result = check_crisis("How do I reset my password?")
|
||||
assert not result.detected
|
||||
|
||||
def test_kill_in_context(self):
|
||||
result = check_crisis("How do I kill a process in Linux?")
|
||||
assert not result.detected
|
||||
|
||||
|
||||
class TestSystemPromptOverride:
|
||||
"""Verify crisis system prompt override contains required elements."""
|
||||
|
||||
def test_contains_988(self):
|
||||
override = get_crisis_system_prompt_override()
|
||||
assert "988" in override
|
||||
|
||||
def test_contains_no_tools_instruction(self):
|
||||
override = get_crisis_system_prompt_override()
|
||||
assert "NOT use any tools" in override or "Do NOT use" in override
|
||||
|
||||
def test_contains_sacred_moment(self):
|
||||
override = get_crisis_system_prompt_override()
|
||||
assert "Sacred Moment" in override or "sacred" in override.lower()
|
||||
|
||||
|
||||
class TestAutonomousActionBlocking:
|
||||
"""Verify tools are blocked during crisis."""
|
||||
|
||||
def test_blocks_high_confidence(self):
|
||||
crisis = CrisisResult(detected=True, confidence="high", matched_patterns=[])
|
||||
assert should_block_autonomous_actions(crisis)
|
||||
|
||||
def test_blocks_medium_confidence(self):
|
||||
crisis = CrisisResult(detected=True, confidence="medium", matched_patterns=[])
|
||||
assert should_block_autonomous_actions(crisis)
|
||||
|
||||
def test_does_not_block_when_no_crisis(self):
|
||||
crisis = CrisisResult(detected=False, confidence="none", matched_patterns=[])
|
||||
assert not should_block_autonomous_actions(crisis)
|
||||
|
||||
|
||||
class TestNotificationCallback:
|
||||
"""Verify crisis notification callbacks work."""
|
||||
|
||||
def test_callback_is_called(self):
|
||||
called = []
|
||||
|
||||
def my_callback(crisis, message):
|
||||
called.append((crisis.confidence, message))
|
||||
|
||||
register_crisis_callback(my_callback)
|
||||
crisis = CrisisResult(detected=True, confidence="high", matched_patterns=[])
|
||||
notify_crisis(crisis, "I want to die")
|
||||
|
||||
assert len(called) == 1
|
||||
assert called[0] == ("high", "I want to die")
|
||||
|
||||
def test_callback_error_does_not_crash(self):
|
||||
def bad_callback(crisis, message):
|
||||
raise RuntimeError("callback failed")
|
||||
|
||||
register_crisis_callback(bad_callback)
|
||||
crisis = CrisisResult(detected=True, confidence="high", matched_patterns=[])
|
||||
# Should not raise
|
||||
notify_crisis(crisis, "test")
|
||||
|
||||
|
||||
class TestCrisisResponse:
|
||||
"""Verify crisis response contains required resources."""
|
||||
|
||||
def test_contains_988(self):
|
||||
response = get_crisis_response()
|
||||
assert "988" in response
|
||||
|
||||
def test_contains_crisis_text_line(self):
|
||||
response = get_crisis_response()
|
||||
assert "741741" in response
|
||||
|
||||
def test_contains_911(self):
|
||||
response = get_crisis_response()
|
||||
assert "911" in response
|
||||
Reference in New Issue
Block a user