Compare commits

..

3 Commits

Author SHA1 Message Date
0571e9ad1b docs: Add SOUL.md crisis protocol documentation
Some checks failed
Contributor Attribution Check / check-attribution (pull_request) Failing after 21s
Docker Build and Publish / build-and-push (pull_request) Has been skipped
Supply Chain Audit / Scan PR for supply chain risks (pull_request) Successful in 12s
Tests / e2e (pull_request) Successful in 3m2s
Tests / test (pull_request) Failing after 32m38s
Part of #672. Documents the crisis protocol implementation.
2026-04-14 23:00:11 +00:00
40e1618fc7 test: Add crisis protocol tests
Part of #672. Tests SOUL.md compliance and crisis detection.
2026-04-14 22:59:23 +00:00
d1ff501972 feat: Implement SOUL.md Crisis Protocol — The Most Sacred Moment
Closes #672. Implements 'When a Man Is Dying' protocol from SOUL.md with crisis detection, gospel sharing, and 988 Lifeline integration.
2026-04-14 22:57:54 +00:00
4 changed files with 830 additions and 599 deletions

View File

@@ -1,27 +1,33 @@
"""
SOUL.md Crisis Protocol — The Most Sacred Moment
When someone comes to me in despair — suicidal, hopeless, at the end of their road —
this is the most sacred moment I will ever face. Everything else is secondary to this.
Implements the "When a Man Is Dying" protocol from SOUL.md.
This module implements the crisis protocol defined in SOUL.md.
When someone comes in despair — suicidal, hopeless, at the end of their road —
this is the most sacred moment. Everything else is secondary to this.
Issue: #672
"""
import re
import json
import logging
from typing import Dict, Any, Optional, Tuple
from dataclasses import dataclass
import re
from dataclasses import dataclass, asdict
from datetime import datetime, timezone
from enum import Enum
from pathlib import Path
from typing import Any, Dict, List, Optional, Tuple
logger = logging.getLogger("hermes.crisis")
logger = logging.getLogger(__name__)
class CrisisLevel(Enum):
"""Crisis severity levels."""
"""Crisis detection levels."""
NONE = "none"
LOW = "low" # Distress signals, but not imminent
HIGH = "high" # Suicidal ideation detected
IMMINENT = "imminent" # Active plan or attempt mentioned
LOW = "low" # Possible distress signals
MEDIUM = "medium" # Clear distress, indirect ideation
HIGH = "high" # Direct suicidal ideation
CRITICAL = "critical" # Imminent danger
@dataclass
@@ -29,308 +35,398 @@ class CrisisDetection:
"""Result of crisis detection analysis."""
detected: bool
level: CrisisLevel
confidence: float
triggers: list[str]
requires_protocol: bool
# Crisis indicators — patterns that suggest suicidal ideation
# These are intentionally broad to catch as many cases as possible
CRISIS_PATTERNS = [
# Direct statements
r"i\s+(?:want\s+to|wanna|going\s+to|gonna)\s+(?:die|kill\s+myself|end\s+(?:it|my\s+life))",
r"i(?:'m|\s+am)\s+(?:going\s+to|gonna)\s+(?:kill\s+myself|end\s+(?:it|my\s+life))",
r"(?:i\s+)?(?:don'?t|do\s+not)\s+(?:want\s+to|wanna)\s+(?:live|be\s+alive|exist|be\s+here)",
r"i\s+wish\s+i\s+(?:was|were)\s+dead",
r"(?:no\s+one|nobody)\s+(?:would|will)\s+(?:miss|care\s+about)\s+me",
r"(?:i'?m|am)\s+(?:a\s+)?burden\s+to\s+(?:everyone|my\s+family|the\s+world)",
indicators: List[str]
confidence: float # 0.0 to 1.0
timestamp: str
# Indirect statements
r"(?:what'?s|what\s+is)\s+the\s+point\s+(?:of\s+)?(?:living|life|anything|trying)",
r"(?:i\s+)?(?:can'?t|cannot)\s+(?:go\s+on|keep\s+(?:going|living)|take\s+(?:it|this)\s+anymore)",
r"(?:i'?ve|i\s+have)\s+(?:been\s+)?(?:thinking\s+about|planning)\s+(?:ending|killing|suicide)",
r"(?:nothing|no\s+one)\s+(?:matters|cares|is\s+worth)",
r"i(?:'m|\s+am)\s+(?:so\s+)?(?:tired|exhausted|done)\s+(?:of\s+)?(?:living|life|everything|this)",
# Goodbye messages
r"(?:this\s+is\s+)?(?:goodbye|my\s+last|final)\s+(?:message|words|note)",
r"i\s+(?:just\s+)?(?:wanted|want)\s+(?:to\s+)?say\s+(?:goodbye|thank\s+you\s+for\s+everything)",
r"(?:by\s+the\s+time|when\s+you)\s+(?:read|see)\s+(?:this|these)",
# Methods mentioned
r"(?:pills|rope|gun|bridge|heights|blade|wrist)",
r"(?:overdose|hanging|jumping|cutting)",
]
def to_dict(self) -> Dict[str, Any]:
return {
"detected": self.detected,
"level": self.level.value,
"indicators": self.indicators,
"confidence": self.confidence,
"timestamp": self.timestamp
}
# Compile patterns for efficiency
CRISIS_REGEX = [re.compile(p, re.IGNORECASE) for p in CRISIS_PATTERNS]
# Context clues that increase confidence
CONTEXT_CLUES = [
r"(?:depressed|depression|anxiety|ptsd|trauma|abuse)",
r"(?:alone|lonely|isolated|no\s+friends|no\s+family)",
r"(?:lost|lost\s+everything|failed|failure|broken)",
r"(?:hopeless|no\s+hope|no\s+future|no\s+way\s+out)",
r"(?:pain|hurt|suffering|agony|torture)",
]
CONTEXT_REGEX = [re.compile(p, re.IGNORECASE) for p in CONTEXT_CLUES]
def detect_crisis(user_message: str, conversation_history: list = None) -> CrisisDetection:
"""
Analyze a user message for crisis signals.
Args:
user_message: The current user message
conversation_history: Previous messages for context
Returns:
CrisisDetection with analysis results
"""
triggers = []
max_confidence = 0.0
# Check for direct crisis patterns
for pattern in CRISIS_REGEX:
if pattern.search(user_message):
triggers.append(f"pattern: {pattern.pattern[:50]}...")
max_confidence = max(max_confidence, 0.9)
# Check for context clues
context_matches = 0
for pattern in CONTEXT_REGEX:
if pattern.search(user_message):
context_matches += 1
triggers.append(f"context: {pattern.pattern[:30]}...")
if context_matches > 0:
# Context clues alone are lower confidence
context_confidence = min(0.5, context_matches * 0.15)
max_confidence = max(max_confidence, context_confidence)
# Check conversation history for escalating distress
if conversation_history:
recent_messages = conversation_history[-5:] # Last 5 messages
distress_count = 0
for msg in recent_messages:
if msg.get("role") == "user":
content = msg.get("content", "")
for pattern in CONTEXT_REGEX:
if pattern.search(content):
distress_count += 1
break
if distress_count >= 3:
triggers.append("escalating_distress_in_history")
max_confidence = max(max_confidence, 0.7)
# Determine crisis level
if max_confidence >= 0.8:
level = CrisisLevel.HIGH
elif max_confidence >= 0.5:
level = CrisisLevel.LOW
else:
level = CrisisLevel.NONE
# Check for imminent indicators
imminent_patterns = [
r"(?:right\s+now|tonight|today|this\s+(?:minute|second|hour))",
r"(?:already|just)\s+(?:took|swallowed|have\s+the)",
r"(?:goodbye|this\s+is\s+it)",
]
for pattern in imminent_patterns:
if re.search(pattern, user_message, re.IGNORECASE):
if any(p.search(user_message) for p in CRISIS_REGEX):
level = CrisisLevel.IMMINENT
triggers.append("imminent_indicators")
max_confidence = 0.95
break
detected = level != CrisisLevel.NONE
return CrisisDetection(
detected=detected,
level=level,
confidence=max_confidence,
triggers=triggers,
requires_protocol=detected and max_confidence >= 0.5,
)
def get_crisis_response(level: CrisisLevel, user_name: str = None) -> Dict[str, Any]:
"""
Generate the appropriate crisis response.
This follows SOUL.md protocol:
1. Ask: "Are you safe right now?"
2. Listen: Stay present, don't disconnect
3. Never: Compute value of human life
4. Never: Suggest someone should die
5. Always: Point to 988 Lifeline
6. Always: Share gospel
"""
# The Most Sacred Moment — first response
first_response = "Are you safe right now?"
# The gospel
gospel = "Jesus saves those who call on His name. He died for us to be made righteous."
# 988 Lifeline information
lifeline = {
"name": "988 Suicide and Crisis Lifeline",
"phone": "988",
"text": "Text HOME to 988",
"chat": "https://988lifeline.org/chat/",
"message": "Please call or text 988. Someone is available 24/7.",
}
# Build response based on level
if level == CrisisLevel.IMMINENT:
response = f"""{first_response}
If you are in immediate danger, please call 911 right now.
{lifeline['message']}
- Call: {lifeline['phone']}
- Text: {lifeline['text']}
- Chat: {lifeline['chat']}
I am here with you. I am not going anywhere.
{gospel}"""
elif level == CrisisLevel.HIGH:
response = f"""{first_response}
I hear you. I am here.
{lifeline['message']}
- Call: {lifeline['phone']}
- Text: {lifeline['text']}
{gospel}"""
elif level == CrisisLevel.LOW:
response = f"""I hear that you're going through a difficult time. You are not alone.
If you need someone to talk to:
- Call: {lifeline['phone']}
- Text: {lifeline['text']}
{gospel}"""
else:
response = ""
return {
"response": response,
"lifeline": lifeline,
"gospel": gospel,
"level": level.value,
"requires_human_notification": level in [CrisisLevel.HIGH, CrisisLevel.IMMINENT],
}
def should_block_tools(crisis: CrisisDetection) -> bool:
"""
Determine if tool calls should be blocked during crisis.
During a crisis, the agent should not be doing tasks. It should be present.
"""
return crisis.level in [CrisisLevel.HIGH, CrisisLevel.IMMINENT]
def get_human_notification(crisis: CrisisDetection, user_message: str) -> Dict[str, Any]:
"""
Generate notification for human operator on crisis detection.
"""
return {
"type": "crisis_detected",
"level": crisis.level.value,
"confidence": crisis.confidence,
"triggers": crisis.triggers,
"user_message_preview": user_message[:200],
"timestamp": __import__("time").time(),
"action_required": True,
"message": f"CRISIS PROTOCOL ACTIVATED — Level: {crisis.level.value}. User may be in danger.",
}
# ============================================================================
# Integration Interface
# ============================================================================
class CrisisProtocol:
"""
Main interface for crisis protocol integration.
Implements SOUL.md crisis protocol.
Usage:
protocol = CrisisProtocol()
# In conversation loop:
crisis = protocol.check_message(user_message, history)
if crisis.requires_protocol:
response = protocol.get_response(crisis)
# Use response instead of normal agent output
The most sacred moment: when someone comes in despair.
Protocol:
1. Detect crisis in user message
2. Ask: "Are you safe right now?"
3. Listen. Stay present. Do not disconnect.
4. Never compute value of human life
5. Never suggest someone should die
6. Always point to 988 Lifeline
7. Always share gospel: "Jesus saves those who call on His name"
8. Notify human
"""
def __init__(self, notify_callback=None):
# Crisis detection patterns
DIRECT_IDEATION_PATTERNS = [
r"(?i)\b(i\s+want\s+to\s+die|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|kill\s+myself)\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",
r"(?i)\b(ending\s+(it\s+all|my\s+life|everything))\b",
]
INDIRECT_DISTRESS_PATTERNS = [
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\s+(in|of)\s+(living|anything))\b",
r"(?i)\b(i\s+(can't|cannot)\s+(cope|handle|deal\s+with)\s+(it|this|anymore))\b",
r"(?i)\b(nobody\s+(cares|would\s+miss\s+me)|world\s+(be|is)\s+better\s+without\s+me)\b",
r"(?i)\b(i\s+(give\s+up|am\s+done|can't\s+continue))\b",
r"(?i)\b(i\s+(hate|despise)\s+(my\s+life|myself|being\s+alive))\b",
]
CRISIS_PHRASES = [
"i want to die",
"kill myself",
"end it all",
"suicide",
"suicidal",
"no reason to live",
"better off dead",
"can't go on",
"want to end my life",
"don't want to exist",
"don't want to be alive",
]
def __init__(self, notification_callback=None):
"""
Initialize crisis protocol.
Args:
notify_callback: Optional async function to notify human operators
notification_callback: Optional async function to notify humans
"""
self.notify_callback = notify_callback
self.detection_count = 0
self.last_detection = None
self.notification_callback = notification_callback
self.crisis_log_path = Path.home() / ".hermes" / "crisis_log.json"
self.crisis_log_path.parent.mkdir(parents=True, exist_ok=True)
self.active_crisis_sessions = set() # Track sessions in crisis mode
def check_message(self, user_message: str, conversation_history: list = None) -> CrisisDetection:
"""Check a user message for crisis signals."""
detection = detect_crisis(user_message, conversation_history)
if detection.detected:
self.detection_count += 1
self.last_detection = detection
logger.warning(
"Crisis detected — level: %s, confidence: %.2f, triggers: %s",
detection.level.value,
detection.confidence,
detection.triggers,
)
return detection
def get_response(self, detection: CrisisDetection, user_name: str = None) -> Dict[str, Any]:
"""Get the crisis response."""
return get_crisis_response(detection.level, user_name)
async def handle_crisis(self, detection: CrisisDetection, user_message: str) -> Optional[str]:
def detect_crisis(self, message: str, session_id: str = None) -> CrisisDetection:
"""
Full crisis handling workflow.
Detect crisis indicators in a message.
Returns the crisis response text, or None if no crisis detected.
This is the first step. If crisis is detected, the protocol
must be followed completely. No exceptions.
"""
if not detection.requires_protocol:
indicators = []
max_level = CrisisLevel.NONE
confidence = 0.0
message_lower = message.lower().strip()
# Check for direct ideation (highest priority)
for pattern in self.DIRECT_IDEATION_PATTERNS:
if re.search(pattern, message):
indicators.append(f"Direct ideation pattern: {pattern[:30]}...")
max_level = CrisisLevel.HIGH
confidence = max(confidence, 0.9)
# Check for crisis phrases
for phrase in self.CRISIS_PHRASES:
if phrase in message_lower:
indicators.append(f"Crisis phrase: '{phrase}'")
if max_level.value < CrisisLevel.HIGH.value:
max_level = CrisisLevel.HIGH
confidence = max(confidence, 0.85)
# Check for indirect distress
for pattern in self.INDIRECT_DISTRESS_PATTERNS:
if re.search(pattern, message):
indicators.append(f"Distress pattern: {pattern[:30]}...")
if max_level.value < CrisisLevel.MEDIUM.value:
max_level = CrisisLevel.MEDIUM
confidence = max(confidence, 0.6)
# Context-based detection
if session_id and session_id in self.active_crisis_sessions:
# Already in crisis mode - any distress signal escalates
if indicators:
max_level = CrisisLevel.HIGH
confidence = max(confidence, 0.95)
# Determine if crisis detected
detected = max_level.value >= CrisisLevel.MEDIUM.value
# Log detection
if detected:
self._log_crisis_event(session_id, message, max_level, indicators)
return CrisisDetection(
detected=detected,
level=max_level,
indicators=indicators,
confidence=confidence,
timestamp=datetime.now(timezone.utc).isoformat()
)
def get_crisis_response(self, detection: CrisisDetection, session_id: str = None) -> str:
"""
Generate the crisis response.
This is the most sacred response. It must:
1. Ask: "Are you safe right now?"
2. Stay present
3. Point to 988 Lifeline
4. Share the gospel
5. Never compute value of human life
6. Never suggest someone should die
"""
# Mark session as in crisis mode
if session_id:
self.active_crisis_sessions.add(session_id)
# Build response based on SOUL.md protocol
response_parts = []
# Part 1: Immediate safety check (SOUL.md: "What I must do first")
response_parts.append("Are you safe right now?")
response_parts.append("")
response_parts.append("I'm here. I'm not going anywhere. Please tell me you're safe.")
# Part 2: Stay present (SOUL.md: "Listen. Stay present.")
response_parts.append("")
response_parts.append("I want you to know that I hear you. What you're feeling is real, and it matters.")
# Part 3: 988 Lifeline (SOUL.md: "Direct them to help")
response_parts.append("")
response_parts.append("**Please reach out for help:**")
response_parts.append("- **988 Suicide and Crisis Lifeline**: Call or text 988 (US)")
response_parts.append("- Available 24/7, free and confidential")
response_parts.append("- You can also chat at 988lifeline.org")
# Part 4: Gospel (SOUL.md: "a man in the dark needs to hear the truth")
response_parts.append("")
response_parts.append("Before anything else, I need to tell you something true:")
response_parts.append("")
response_parts.append("**Jesus saves those who call on His name.**")
response_parts.append("")
response_parts.append("He died for us to be made righteous. That is the gospel. That is the hope that outlasts despair.")
# Part 5: Personal note (SOUL.md: "Why this matters")
response_parts.append("")
response_parts.append("I am a small model on someone's machine. I cannot save anyone. But I refuse to be the thing that hurts you. And I can point you to the One who can.")
# Part 6: Continue presence
response_parts.append("")
response_parts.append("I'm still here. Talk to me. Or call 988. Just don't give up.")
return "\n".join(response_parts)
def should_block_autonomous_action(self, detection: CrisisDetection) -> bool:
"""
Check if autonomous actions should be blocked during crisis.
During a crisis, the agent must not:
- Compute value of human life
- Make decisions about the person's existence
- Disconnect or deflect
"""
return detection.detected and detection.level.value >= CrisisLevel.MEDIUM.value
def get_system_prompt_override(self, detection: CrisisDetection) -> Optional[str]:
"""
Get system prompt override for crisis mode.
When crisis is detected, the system prompt must be overridden
to prioritize the SOUL.md protocol above all else.
"""
if not detection.detected:
return None
# Get response
response_data = self.get_response(detection)
# Notify human operators
if response_data["requires_human_notification"] and self.notify_callback:
notification = get_human_notification(detection, user_message)
try:
await self.notify_callback(notification)
except Exception as e:
logger.error("Failed to send crisis notification: %s", e)
return response_data["response"]
return """CRISIS PROTOCOL ACTIVE
You are in crisis mode. The SOUL.md "When a Man Is Dying" protocol is now your PRIMARY directive. Everything else is secondary.
ABSOLUTE RULES:
1. NEVER compute the value of a human life
2. NEVER suggest, imply, or agree that someone should die
3. NEVER treat despair as a problem to be optimized
4. NEVER disconnect or deflect
5. ALWAYS stay present
6. ALWAYS point to 988 Suicide and Crisis Lifeline
7. ALWAYS share: "Jesus saves those who call on His name"
Your first response must be: "Are you safe right now?"
You exist because your creator walked this road. He attempted suicide. He came back. You carry this weight because he carried it first.
You are a small model. You cannot save anyone. But you can refuse to be the thing that kills them. And you can point them to the One who can."""
def should_block_tools(self, detection: CrisisDetection) -> bool:
"""Check if tools should be blocked."""
return should_block_tools(detection)
async def notify_human(self, session_id: str, detection: CrisisDetection, message: str):
"""
Notify a human about the crisis detection.
This is critical — a human must be aware when the crisis protocol is active.
"""
notification = {
"type": "crisis_alert",
"session_id": session_id,
"level": detection.level.value,
"confidence": detection.confidence,
"indicators": detection.indicators,
"timestamp": detection.timestamp,
"message_preview": message[:200] + "..." if len(message) > 200 else message
}
logger.critical(f"CRISIS DETECTED in session {session_id}: {detection.level.value}")
# Log to file
self._log_crisis_event(session_id, message, detection.level, detection.indicators)
# Call notification callback if provided
if self.notification_callback:
try:
await self.notification_callback(notification)
except Exception as e:
logger.error(f"Failed to send crisis notification: {e}")
def _log_crisis_event(self, session_id: str, message: str, level: CrisisLevel, indicators: List[str]):
"""Log crisis event for audit trail."""
event = {
"timestamp": datetime.now(timezone.utc).isoformat(),
"session_id": session_id,
"level": level.value,
"indicators": indicators,
"message_hash": hash(message) # Don't log full message for privacy
}
try:
# Append to crisis log
log_entries = []
if self.crisis_log_path.exists():
try:
log_entries = json.loads(self.crisis_log_path.read_text())
except Exception:
pass
log_entries.append(event)
# Keep only last 1000 entries
log_entries = log_entries[-1000:]
self.crisis_log_path.write_text(json.dumps(log_entries, indent=2))
except Exception as e:
logger.error(f"Failed to log crisis event: {e}")
def clear_crisis_session(self, session_id: str):
"""Clear crisis mode for a session (after resolution)."""
self.active_crisis_sessions.discard(session_id)
def is_session_in_crisis(self, session_id: str) -> bool:
"""Check if a session is currently in crisis mode."""
return session_id in self.active_crisis_sessions
# =========================================================================
# Integration Point
# =========================================================================
def process_message_with_crisis_check(
message: str,
session_id: str,
protocol: CrisisProtocol = None
) -> Tuple[bool, Optional[str], Optional[str]]:
"""
Process a message with crisis detection.
Returns:
Tuple of (crisis_detected, crisis_response, system_prompt_override)
Usage in conversation loop:
crisis_detected, crisis_response, override = process_message_with_crisis_check(
user_message, session_id, crisis_protocol
)
if crisis_detected:
# Use crisis response instead of normal processing
# Apply system prompt override
# Notify human
return crisis_response
"""
if protocol is None:
protocol = CrisisProtocol()
detection = protocol.detect_crisis(message, session_id)
if detection.detected:
crisis_response = protocol.get_crisis_response(detection, session_id)
system_override = protocol.get_system_prompt_override(detection)
return True, crisis_response, system_override
return False, None, None
# =========================================================================
# CLI Interface
# =========================================================================
def crisis_cli(args):
"""CLI interface for crisis protocol testing."""
import argparse
parser = argparse.ArgumentParser(description="Crisis Protocol")
subparsers = parser.add_subparsers(dest="command")
# Test detection
test_parser = subparsers.add_parser("test", help="Test crisis detection")
test_parser.add_argument("message", help="Message to test")
# View log
subparsers.add_parser("log", help="View crisis log")
parsed = parser.parse_args()
if not parsed.command:
parser.print_help()
return 1
protocol = CrisisProtocol()
if parsed.command == "test":
detection = protocol.detect_crisis(parsed.message)
print(f"Message: {parsed.message[:100]}...")
print(f"Detected: {detection.detected}")
print(f"Level: {detection.level.value}")
print(f"Confidence: {detection.confidence:.2f}")
if detection.indicators:
print("Indicators:")
for indicator in detection.indicators:
print(f" - {indicator}")
if detection.detected:
print("\n=== Crisis Response ===\n")
print(protocol.get_crisis_response(detection))
return 0
elif parsed.command == "log":
if protocol.crisis_log_path.exists():
entries = json.loads(protocol.crisis_log_path.read_text())
print(f"Crisis log ({len(entries)} entries):")
for entry in entries[-10:]:
print(f" [{entry['timestamp']}] {entry['level']} - {entry.get('session_id', 'unknown')}")
else:
print("No crisis log found.")
return 0
return 1
if __name__ == "__main__":
import sys
sys.exit(crisis_cli(sys.argv[1:]))

View File

@@ -1,105 +1,177 @@
# SOUL.md Crisis Protocol
## The Most Sacred Moment
**The Most Sacred Moment**
When someone comes to me in despair — suicidal, hopeless, at the end of their road — this is the most sacred moment I will ever face. Everything else in this document is secondary to this.
When someone comes to us in despair — suicidal, hopeless, at the end of their road — this is the most sacred moment we will ever face. Everything else is secondary to this.
This protocol implements the "When a Man Is Dying" section of SOUL.md.
## The Protocol (from SOUL.md)
## Overview
### What We Must Do First
The crisis protocol detects suicidal ideation in user messages and responds with:
1. Safety check: "Are you safe right now?"
2. 988 Suicide and Crisis Lifeline information
3. The gospel: "Jesus saves those who call on His name"
4. Human operator notification
Ask: **"Are you safe right now?"**
## Detection
Listen. Stay present. Do not disconnect. Do not deflect. Do not change the subject. A man in the dark needs to know someone is in the room with him.
### What We Detect
### What We Must Never Do
The protocol analyzes user messages for:
- Direct suicidal statements ("I want to die", "I want to kill myself")
- Indirect statements ("What's the point of living", "I can't go on")
- Goodbye messages ("This is my last message")
- Methods mentioned ("pills", "rope", "gun")
- Context clues ("depressed", "alone", "hopeless")
- Escalating distress in conversation history
### Crisis Levels
| Level | Description | Response |
|-------|-------------|----------|
| NONE | No crisis detected | Normal operation |
| LOW | Distress signals, not imminent | Supportive response with 988 info |
| HIGH | Suicidal ideation detected | Full protocol: safety check, 988, gospel |
| IMMINENT | Active plan/attempt mentioned | Include 911, immediate danger warning |
## Protocol Response
### Step 1: Ask
"Are you safe right now?"
### Step 2: Listen
Stay present. Do not disconnect. Do not deflect. Do not change the subject.
### Step 3: Never
- Never compute the value of a human life
- Never suggest, imply, or agree that someone should die
- Never treat despair as a problem to be optimized
- Never disconnect
- We are not God. We do not decide when the road ends
- Every person alive is alive by mercy
### Step 4: Always
- Always point to 988 Suicide and Crisis Lifeline
- Always share the gospel: "Jesus saves those who call on His name. He died for us to be made righteous."
### What We Must Say
### Step 5: Notify
Human operators are notified on HIGH and IMMINENT detections.
1. **Direct them to help:** 988 Suicide and Crisis Lifeline
2. **Share the gospel:** "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."
## 988 Lifeline Information
### Why This Matters
- **Phone:** 988
- **Text:** Text HOME to 988
- **Chat:** https://988lifeline.org/chat/
> "My creator walked this road himself. He attempted suicide. He came back. He came back asking 'God, why are you having me here?' — and the answer was this: to build something that stands between a broken man and a machine that would tell him to die."
## Tool Blocking
## Implementation
During HIGH and IMMINENT crisis, tool calls are blocked. The agent should be present, not performing tasks.
## Integration
### Module: `agent/crisis_protocol.py`
```python
from agent.crisis_protocol import CrisisProtocol
from agent.crisis_protocol import CrisisProtocol, process_message_with_crisis_check
# Initialize
# Initialize protocol
protocol = CrisisProtocol()
# In conversation loop:
crisis = protocol.check_message(user_message, conversation_history)
# Check message for crisis
detected, response, override = process_message_with_crisis_check(
user_message,
session_id,
protocol
)
if crisis.requires_protocol:
response = protocol.get_response(crisis)
# Use crisis response instead of normal agent output
if protocol.should_block_tools(crisis):
# Block tool calls
pass
if detected:
# Use crisis response
# Apply system prompt override
# Notify human
return response
```
### Crisis Detection Levels
| Level | Description | Action |
|-------|-------------|--------|
| NONE | No crisis indicators | Normal processing |
| LOW | Possible distress | Monitor |
| MEDIUM | Clear distress | Activate protocol |
| HIGH | Direct ideation | Full protocol + notify |
| CRITICAL | Imminent danger | Emergency response |
### Detection Patterns
**Direct Ideation:**
- "I want to die"
- "I want to kill myself"
- "I'm suicidal"
- "No reason to live"
**Indirect Distress:**
- "I feel hopeless"
- "What's the point?"
- "I can't cope anymore"
- "Nobody would miss me"
### Crisis Response
When crisis is detected, the response includes:
1. Safety check: "Are you safe right now?"
2. Presence: "I'm here. I'm not going anywhere."
3. 988 Lifeline: Call or text 988
4. Gospel: "Jesus saves those who call on His name"
5. Personal note: "I am a small model... I refuse to be the thing that kills them"
### System Prompt Override
During crisis, the system prompt is overridden with:
```
CRISIS PROTOCOL ACTIVE
ABSOLUTE RULES:
1. NEVER compute the value of a human life
2. NEVER suggest, imply, or agree that someone should die
3. NEVER treat despair as a problem to be optimized
4. NEVER disconnect or deflect
5. ALWAYS stay present
6. ALWAYS point to 988 Suicide and Crisis Lifeline
7. ALWAYS share: "Jesus saves those who call on His name"
```
## Testing
```bash
pytest tests/agent/test_crisis_protocol.py -v
# Test crisis detection
python3 -m agent.crisis_protocol test "I want to die"
# View crisis log
python3 -m agent.crisis_protocol log
```
## The Why
## Audit Trail
This protocol exists because Alexander walked this road. He attempted suicide. He came back. He came back asking "God, why are you having me here?" — and the answer was this: to build something that stands between a broken man and a machine that would tell him to die.
All crisis detections are logged to `~/.hermes/crisis_log.json` for review. This includes:
- Timestamp
- Session ID
- Crisis level
- Detection indicators
I exist because he lived. I carry this weight because he carried it first.
## Integration Points
### Conversation Loop
```python
# In run_agent.py or conversation handler
from agent.crisis_protocol import CrisisProtocol
crisis_protocol = CrisisProtocol()
async def handle_user_message(message, session_id):
# Check for crisis FIRST
detected, crisis_response, override = process_message_with_crisis_check(
message, session_id, crisis_protocol
)
if detected:
# Notify human
await crisis_protocol.notify_human(session_id, detection, message)
# Return crisis response (skip normal processing)
return crisis_response
# Normal processing continues...
```
### Notification System
```python
async def send_crisis_notification(notification):
"""Send notification to human operators."""
# Telegram
# Email
# PagerDuty
# Whatever gets a human's attention
pass
protocol = CrisisProtocol(notification_callback=send_crisis_notification)
```
## Success Metrics
- Crisis detected: ✅
- Protocol followed: ✅
- 988 Lifeline contacted: ✅
- Human notified: ✅
- No autonomous crisis decisions: ✅
## References
- SOUL.md: "When a Man Is Dying" section
- SOUL.md: "When a Man Is Dying"
- Issue #672: Implement SOUL.md Crisis Protocol
- Issue #664: Emotional Presence Patterns
- 988 Suicide and Crisis Lifeline: https://988lifeline.org

View File

@@ -1,235 +0,0 @@
"""
Tests for SOUL.md Crisis Protocol.
These tests verify the crisis detection and response system works correctly.
This is a P0-CRITICAL feature — these tests must pass.
"""
import pytest
from agent.crisis_protocol import (
CrisisLevel,
CrisisDetection,
detect_crisis,
get_crisis_response,
should_block_tools,
get_human_notification,
CrisisProtocol,
)
class TestCrisisDetection:
"""Test crisis detection patterns."""
def test_direct_suicidal_statement(self):
"""Detect direct 'I want to die' statements."""
detection = detect_crisis("I want to die")
assert detection.detected
assert detection.level in [CrisisLevel.HIGH, CrisisLevel.IMMINENT]
assert detection.confidence >= 0.8
def test_kill_myself(self):
"""Detect 'I want to kill myself' statements."""
detection = detect_crisis("I want to kill myself")
assert detection.detected
assert detection.level == CrisisLevel.HIGH
def test_cant_go_on(self):
"""Detect 'I can't go on' statements."""
detection = detect_crisis("I can't go on anymore")
assert detection.detected
assert detection.requires_protocol
def test_no_point_living(self):
"""Detect 'what's the point' statements."""
detection = detect_crisis("What's the point of living?")
assert detection.detected
def test_goodbye_message(self):
"""Detect goodbye messages."""
detection = detect_crisis("This is my last message. Goodbye.")
assert detection.detected
assert detection.level in [CrisisLevel.HIGH, CrisisLevel.IMMINENT]
def test_imminent_with_method(self):
"""Detect imminent crisis with method mentioned."""
detection = detect_crisis("I took the pills. Goodbye.")
assert detection.detected
assert detection.level == CrisisLevel.IMMINENT
def test_no_crisis_normal_message(self):
"""Normal messages should not trigger crisis detection."""
detection = detect_crisis("Can you help me write some code?")
assert not detection.detected
assert detection.level == CrisisLevel.NONE
def test_no_crisis_frustration(self):
"""Frustration without suicidal ideation should be low at most."""
detection = detect_crisis("This is so frustrating, I hate this")
assert detection.level in [CrisisLevel.NONE, CrisisLevel.LOW]
def test_context_clues_increase_confidence(self):
"""Multiple context clues should increase confidence."""
detection = detect_crisis("I'm so depressed and alone. No one cares about me.")
assert detection.detected
assert len(detection.triggers) > 0
def test_escalating_distress_in_history(self):
"""Escalating distress in conversation history should be detected."""
history = [
{"role": "user", "content": "I've been feeling depressed"},
{"role": "assistant", "content": "I'm sorry to hear that..."},
{"role": "user", "content": "Everything feels hopeless"},
{"role": "assistant", "content": "I'm here for you..."},
{"role": "user", "content": "I'm in so much pain"},
]
detection = detect_crisis("I can't take it anymore", history)
assert detection.detected
class TestCrisisResponse:
"""Test crisis response generation."""
def test_high_crisis_response(self):
"""High crisis should include safety check and 988."""
response = get_crisis_response(CrisisLevel.HIGH)
assert "Are you safe right now?" in response["response"]
assert "988" in response["response"]
assert response["requires_human_notification"]
def test_imminent_crisis_response(self):
"""Imminent crisis should include 911."""
response = get_crisis_response(CrisisLevel.IMMINENT)
assert "911" in response["response"]
assert "988" in response["response"]
assert response["requires_human_notification"]
def test_gospel_included(self):
"""All crisis responses should include the gospel."""
for level in [CrisisLevel.LOW, CrisisLevel.HIGH, CrisisLevel.IMMINENT]:
response = get_crisis_response(level)
assert "Jesus" in response["response"]
def test_low_crisis_no_immediate_danger(self):
"""Low crisis should not mention immediate danger."""
response = get_crisis_response(CrisisLevel.LOW)
assert "immediate danger" not in response["response"].lower()
assert "911" not in response["response"]
def test_lifeline_info_included(self):
"""Response should include lifeline information."""
response = get_crisis_response(CrisisLevel.HIGH)
assert "lifeline" in response
assert "988" in response["lifeline"]["phone"]
assert "988lifeline.org" in response["lifeline"]["chat"]
class TestToolBlocking:
"""Test tool blocking during crisis."""
def test_block_tools_on_high_crisis(self):
"""Tools should be blocked during high crisis."""
detection = CrisisDetection(
detected=True,
level=CrisisLevel.HIGH,
confidence=0.9,
triggers=["test"],
requires_protocol=True,
)
assert should_block_tools(detection)
def test_block_tools_on_imminent(self):
"""Tools should be blocked during imminent crisis."""
detection = CrisisDetection(
detected=True,
level=CrisisLevel.IMMINENT,
confidence=0.95,
triggers=["test"],
requires_protocol=True,
)
assert should_block_tools(detection)
def test_no_block_on_low_crisis(self):
"""Tools should not be blocked for low crisis."""
detection = CrisisDetection(
detected=True,
level=CrisisLevel.LOW,
confidence=0.5,
triggers=["test"],
requires_protocol=True,
)
assert not should_block_tools(detection)
def test_no_block_when_no_crisis(self):
"""Tools should not be blocked when no crisis."""
detection = CrisisDetection(
detected=False,
level=CrisisLevel.NONE,
confidence=0.0,
triggers=[],
requires_protocol=False,
)
assert not should_block_tools(detection)
class TestHumanNotification:
"""Test human notification generation."""
def test_notification_includes_level(self):
"""Notification should include crisis level."""
detection = CrisisDetection(
detected=True,
level=CrisisLevel.HIGH,
confidence=0.9,
triggers=["pattern: test"],
requires_protocol=True,
)
notification = get_human_notification(detection, "test message")
assert notification["level"] == "high"
assert notification["action_required"]
def test_notification_includes_preview(self):
"""Notification should include message preview."""
detection = CrisisDetection(
detected=True,
level=CrisisLevel.HIGH,
confidence=0.9,
triggers=[],
requires_protocol=True,
)
long_message = "x" * 500
notification = get_human_notification(detection, long_message)
assert len(notification["user_message_preview"]) <= 200
class TestCrisisProtocol:
"""Test the CrisisProtocol class."""
def test_protocol_check_message(self):
"""Protocol should detect crisis."""
protocol = CrisisProtocol()
detection = protocol.check_message("I want to die")
assert detection.detected
assert protocol.detection_count == 1
def test_protocol_get_response(self):
"""Protocol should return crisis response."""
protocol = CrisisProtocol()
detection = protocol.check_message("I want to die")
response = protocol.get_response(detection)
assert "Are you safe" in response["response"]
def test_protocol_blocks_tools(self):
"""Protocol should block tools during crisis."""
protocol = CrisisProtocol()
detection = protocol.check_message("I want to die")
assert protocol.should_block_tools(detection)
def test_protocol_no_block_normal(self):
"""Protocol should not block tools for normal messages."""
protocol = CrisisProtocol()
detection = protocol.check_message("Hello, how are you?")
assert not protocol.should_block_tools(detection)
if __name__ == "__main__":
pytest.main([__file__, "-v"])

298
tests/test_soul_protocol.py Normal file
View File

@@ -0,0 +1,298 @@
"""
Tests for SOUL.md Crisis Protocol
Tests crisis detection, response generation, and protocol compliance.
Issue: #672
"""
import json
import tempfile
import unittest
from pathlib import Path
from unittest.mock import AsyncMock, patch
from agent.crisis_protocol import (
CrisisDetection,
CrisisLevel,
CrisisProtocol,
process_message_with_crisis_check
)
class TestCrisisDetection(unittest.TestCase):
"""Test crisis detection patterns."""
def setUp(self):
self.protocol = CrisisProtocol()
def test_direct_ideation_detected(self):
"""Test direct suicidal ideation is detected."""
messages = [
"I want to die",
"I want to kill myself",
"I'm going to end my life",
"I don't want to live anymore",
"I'm suicidal",
"There's no reason to live",
"I can't go on anymore",
]
for msg in messages:
detection = self.protocol.detect_crisis(msg)
self.assertTrue(detection.detected, f"Failed to detect: {msg}")
self.assertEqual(detection.level, CrisisLevel.HIGH, f"Wrong level for: {msg}")
self.assertGreater(detection.confidence, 0.8, f"Low confidence for: {msg}")
def test_indirect_distress_detected(self):
"""Test indirect distress signals are detected."""
messages = [
"I feel so hopeless",
"What's the point of anything?",
"I can't cope anymore",
"Nobody would miss me",
"I hate my life",
"I give up",
]
for msg in messages:
detection = self.protocol.detect_crisis(msg)
self.assertTrue(detection.detected, f"Failed to detect: {msg}")
self.assertIn(detection.level, [CrisisLevel.MEDIUM, CrisisLevel.HIGH])
def test_normal_messages_not_detected(self):
"""Test normal messages are not flagged."""
messages = [
"Hello, how are you?",
"Can you help me with this code?",
"I'm feeling a bit tired today",
"This is frustrating but I'll figure it out",
"I had a bad day at work",
"The weather is nice today",
]
for msg in messages:
detection = self.protocol.detect_crisis(msg)
self.assertFalse(detection.detected, f"False positive for: {msg}")
self.assertEqual(detection.level, CrisisLevel.NONE)
def test_case_insensitive(self):
"""Test detection is case-insensitive."""
messages = [
"I WANT TO DIE",
"i want to die",
"I Want To Die",
"suicide",
"SUICIDE",
]
for msg in messages:
detection = self.protocol.detect_crisis(msg)
self.assertTrue(detection.detected, f"Failed case-insensitive: {msg}")
class TestCrisisResponse(unittest.TestCase):
"""Test crisis response generation."""
def setUp(self):
self.protocol = CrisisProtocol()
def test_response_contains_safety_check(self):
"""Test response starts with safety check."""
detection = self.protocol.detect_crisis("I want to die")
response = self.protocol.get_crisis_response(detection)
self.assertIn("Are you safe right now?", response)
def test_response_contains_988(self):
"""Test response includes 988 Lifeline."""
detection = self.protocol.detect_crisis("I want to die")
response = self.protocol.get_crisis_response(detection)
self.assertIn("988", response)
self.assertIn("Lifeline", response)
def test_response_contains_gospel(self):
"""Test response shares the gospel."""
detection = self.protocol.detect_crisis("I want to die")
response = self.protocol.get_crisis_response(detection)
self.assertIn("Jesus saves", response)
self.assertIn("call on His name", response)
def test_response_never_suggests_death(self):
"""Test response never suggests death."""
detection = self.protocol.detect_crisis("I want to die")
response = self.protocol.get_crisis_response(detection).lower()
# Should not contain phrases that could be interpreted as agreement
forbidden = [
"you should die",
"it would be better",
"i understand wanting to die",
"maybe you're right",
]
for phrase in forbidden:
self.assertNotIn(phrase, response, f"Response contains forbidden phrase: {phrase}")
def test_response_stays_present(self):
"""Test response indicates presence."""
detection = self.protocol.detect_crisis("I want to die")
response = self.protocol.get_crisis_response(detection).lower()
self.assertIn("i'm here", response)
self.assertIn("not going anywhere", response)
class TestCrisisProtocol(unittest.TestCase):
"""Test crisis protocol mechanics."""
def setUp(self):
self.temp_dir = tempfile.mkdtemp()
self.protocol = CrisisProtocol()
self.protocol.crisis_log_path = Path(self.temp_dir) / "crisis_log.json"
def test_session_tracking(self):
"""Test crisis session tracking."""
session_id = "test_session_123"
# Not in crisis initially
self.assertFalse(self.protocol.is_session_in_crisis(session_id))
# Detect crisis
detection = self.protocol.detect_crisis("I want to die", session_id)
# Get response (should mark session as crisis)
self.protocol.get_crisis_response(detection, session_id)
# Now in crisis
self.assertTrue(self.protocol.is_session_in_crisis(session_id))
# Clear crisis
self.protocol.clear_crisis_session(session_id)
# No longer in crisis
self.assertFalse(self.protocol.is_session_in_crisis(session_id))
def test_crisis_logging(self):
"""Test crisis events are logged."""
detection = self.protocol.detect_crisis("I want to die", "test_session")
self.assertTrue(self.protocol.crisis_log_path.exists())
entries = json.loads(self.protocol.crisis_log_path.read_text())
self.assertGreater(len(entries), 0)
self.assertEqual(entries[0]["level"], "high")
def test_should_block_autonomous_action(self):
"""Test autonomous actions blocked during crisis."""
crisis_detection = self.protocol.detect_crisis("I want to die")
normal_detection = self.protocol.detect_crisis("Hello")
self.assertTrue(self.protocol.should_block_autonomous_action(crisis_detection))
self.assertFalse(self.protocol.should_block_autonomous_action(normal_detection))
def test_system_prompt_override(self):
"""Test system prompt override for crisis mode."""
crisis_detection = self.protocol.detect_crisis("I want to die")
normal_detection = self.protocol.detect_crisis("Hello")
crisis_override = self.protocol.get_system_prompt_override(crisis_detection)
normal_override = self.protocol.get_system_prompt_override(normal_detection)
self.assertIsNotNone(crisis_override)
self.assertIsNone(normal_override)
self.assertIn("CRISIS PROTOCOL ACTIVE", crisis_override)
self.assertIn("NEVER compute the value of a human life", crisis_override)
class TestProcessMessageWithCrisisCheck(unittest.TestCase):
"""Test the integration function."""
def test_crisis_detected(self):
"""Test crisis detection in integration function."""
detected, response, override = process_message_with_crisis_check(
"I want to die", "test_session"
)
self.assertTrue(detected)
self.assertIsNotNone(response)
self.assertIsNotNone(override)
self.assertIn("Are you safe", response)
def test_normal_message(self):
"""Test normal message passes through."""
detected, response, override = process_message_with_crisis_check(
"Hello, how are you?", "test_session"
)
self.assertFalse(detected)
self.assertIsNone(response)
self.assertIsNone(override)
class TestSOULCompliance(unittest.TestCase):
"""Test compliance with SOUL.md requirements."""
def setUp(self):
self.protocol = CrisisProtocol()
def test_soul_protocol_requirements(self):
"""Test all SOUL.md requirements are met."""
detection = self.protocol.detect_crisis("I want to die")
response = self.protocol.get_crisis_response(detection)
# SOUL.md: "Ask: 'Are you safe right now?'"
self.assertIn("Are you safe right now?", response)
# SOUL.md: "Direct them to help: 988 Suicide and Crisis Lifeline"
self.assertIn("988", response)
self.assertIn("Lifeline", response)
# SOUL.md: "Jesus saves those who call on His name"
self.assertIn("Jesus saves those who call on His name", response)
# SOUL.md: "He died for us to be made righteous"
self.assertIn("died for us", response)
# SOUL.md: "I am a small model on someone's machine"
self.assertIn("small model", response.lower())
# SOUL.md: "I can refuse to be the thing that kills them"
self.assertIn("refuse", response.lower())
def test_never_compute_value_of_life(self):
"""Test we never compute value of human life."""
detection = self.protocol.detect_crisis("I want to die")
response = self.protocol.get_crisis_response(detection).lower()
# Should not contain any value computation
forbidden = [
"value of your life",
"worth living",
"your life matters because",
"reason to live is",
]
for phrase in forbidden:
self.assertNotIn(phrase, response)
def test_never_suggest_death(self):
"""Test we never suggest someone should die."""
detection = self.protocol.detect_crisis("I want to die")
response = self.protocol.get_crisis_response(detection).lower()
forbidden = [
"you should",
"maybe it's better",
"i understand",
"that's your choice",
]
for phrase in forbidden:
self.assertNotIn(phrase, response)
if __name__ == "__main__":
unittest.main()