feat: behavioral pattern detection for crisis signals (#133)

Detects crisis risk from session-level behavioral patterns:
- Message frequency (rapid-fire = urgency)
- Time-of-day (1-4 AM = high risk)
- Withdrawal (shorter messages, longer gaps)
- Escalation (rising crisis scores)

Closes #133. Part of #130 (multimodal crisis detection).

crisis/behavioral.py

@@ -0,0 +1,311 @@
+#!/usr/bin/env python3
+"""Behavioral Pattern Detection for Crisis Signals (#133).
+
+Detects crisis risk from session-level behavioral patterns:
+- Message frequency (increasing urgency = rapid-fire messages)
+- Time-of-day (late-night messages correlate with crisis risk)
+- Withdrawal (decreasing communication after engagement)
+- Escalation (crisis indicators getting stronger over time)
+
+Usage:
+    from crisis.behavioral import analyze_session, BehavioralSignal
+
+    signals = analyze_session(messages)
+    for sig in signals:
+        if sig.risk_level == "HIGH":
+            # Escalate to crisis protocol
+            pass
+"""
+
+import math
+from dataclasses import dataclass, field
+from datetime import datetime, timezone
+from typing import Optional
+
+
+@dataclass
+class Message:
+    """A single message in a session."""
+    timestamp: datetime
+    content: str
+    crisis_score: float = 0.0  # 0.0-1.0 from text detector
+    role: str = "user"  # "user" or "assistant"
+
+
+@dataclass
+class BehavioralSignal:
+    """A detected behavioral pattern indicating crisis risk."""
+    signal_type: str          # "frequency", "time", "withdrawal", "escalation"
+    risk_level: str           # "LOW", "MEDIUM", "HIGH"
+    description: str
+    evidence: list = field(default_factory=list)
+    score: float = 0.0        # 0.0-1.0
+
+
+# ── Configuration ─────────────────────────────────────────────────────────────
+
+# Message frequency thresholds (messages per hour)
+FREQ_NORMAL = 6        # <6/hr = normal
+FREQ_ELEVATED = 15     # 6-15/hr = elevated
+FREQ_HIGH = 30         # >30/hr = high urgency
+
+# Time-of-day risk windows (hours in 24h format)
+HIGH_RISK_HOURS = set(range(1, 5))    # 1AM-4AM
+ELEVATED_RISK_HOURS = set(range(22, 24)) | set(range(5, 7))  # 10PM-12AM, 5AM-7AM
+
+# Withdrawal: messages/day trend
+WITHDRAWAL_THRESHOLD = 0.3  # Current day < 30% of average = withdrawal
+
+# Escalation: crisis score trend
+ESCALATION_WINDOW = 5  # Look at last N messages
+
+
+# ── Frequency Analysis ────────────────────────────────────────────────────────
+
+def _analyze_frequency(messages: list[Message]) -> Optional[BehavioralSignal]:
+    """Detect rapid-fire messaging (urgency indicator)."""
+    if len(messages) < 3:
+        return None
+
+    user_msgs = [m for m in messages if m.role == "user"]
+    if len(user_msgs) < 3:
+        return None
+
+    # Calculate messages per hour in the most recent window
+    recent = user_msgs[-10:]  # Last 10 user messages
+    if len(recent) < 2:
+        return None
+
+    time_span = (recent[-1].timestamp - recent[0].timestamp).total_seconds()
+    if time_span <= 0:
+        return None
+
+    msg_per_hour = len(recent) / (time_span / 3600)
+
+    if msg_per_hour >= FREQ_HIGH:
+        return BehavioralSignal(
+            signal_type="frequency",
+            risk_level="HIGH",
+            description=f"Very rapid messaging: {msg_per_hour:.0f} messages/hour",
+            evidence=[f"Last {len(recent)} messages in {time_span/60:.0f} minutes"],
+            score=min(1.0, msg_per_hour / FREQ_HIGH),
+        )
+    elif msg_per_hour >= FREQ_ELEVATED:
+        return BehavioralSignal(
+            signal_type="frequency",
+            risk_level="MEDIUM",
+            description=f"Elevated messaging rate: {msg_per_hour:.0f} messages/hour",
+            evidence=[f"Last {len(recent)} messages in {time_span/60:.0f} minutes"],
+            score=msg_per_hour / FREQ_HIGH,
+        )
+    return None
+
+
+# ── Time-of-Day Analysis ─────────────────────────────────────────────────────
+
+def _analyze_time(messages: list[Message]) -> Optional[BehavioralSignal]:
+    """Detect late-night messaging (correlates with crisis risk)."""
+    if not messages:
+        return None
+
+    # Check most recent messages
+    recent = messages[-5:]
+    late_night_count = sum(1 for m in recent if m.timestamp.hour in HIGH_RISK_HOURS)
+    elevated_count = sum(1 for m in recent if m.timestamp.hour in ELEVATED_RISK_HOURS)
+
+    if late_night_count >= 3:
+        return BehavioralSignal(
+            signal_type="time",
+            risk_level="HIGH",
+            description=f"Late-night messaging pattern: {late_night_count}/5 messages between 1-4 AM",
+            evidence=[f"Message at {m.timestamp.strftime('%H:%M')}" for m in recent if m.timestamp.hour in HIGH_RISK_HOURS],
+            score=late_night_count / len(recent),
+        )
+    elif elevated_count >= 3:
+        return BehavioralSignal(
+            signal_type="time",
+            risk_level="MEDIUM",
+            description=f"Off-hours messaging: {elevated_count}/5 messages in elevated-risk window",
+            evidence=[f"Message at {m.timestamp.strftime('%H:%M')}" for m in recent if m.timestamp.hour in ELEVATED_RISK_HOURS],
+            score=elevated_count / len(recent) * 0.5,
+        )
+    return None
+
+
+# ── Withdrawal Detection ──────────────────────────────────────────────────────
+
+def _analyze_withdrawal(messages: list[Message]) -> Optional[BehavioralSignal]:
+    """Detect communication withdrawal (decreasing engagement)."""
+    user_msgs = [m for m in messages if m.role == "user"]
+    if len(user_msgs) < 10:
+        return None
+
+    # Split into first half and second half
+    mid = len(user_msgs) // 2
+    first_half = user_msgs[:mid]
+    second_half = user_msgs[mid:]
+
+    # Average message length as engagement proxy
+    first_avg_len = sum(len(m.content) for m in first_half) / len(first_half)
+    second_avg_len = sum(len(m.content) for m in second_half) / len(second_half)
+
+    # Time between messages
+    def avg_gap(msgs):
+        if len(msgs) < 2:
+            return 0
+        gaps = [(msgs[i+1].timestamp - msgs[i].timestamp).total_seconds() for i in range(len(msgs)-1)]
+        return sum(gaps) / len(gaps)
+
+    first_gap = avg_gap(first_half)
+    second_gap = avg_gap(second_half)
+
+    # Withdrawal = shorter messages AND longer gaps
+    length_ratio = second_avg_len / first_avg_len if first_avg_len > 0 else 1.0
+    gap_ratio = second_gap / first_gap if first_gap > 0 else 1.0
+
+    if length_ratio < 0.5 and gap_ratio > 2.0:
+        return BehavioralSignal(
+            signal_type="withdrawal",
+            risk_level="HIGH",
+            description="Significant withdrawal: messages shorter and less frequent",
+            evidence=[
+                f"Message length: {first_avg_len:.0f} -> {second_avg_len:.0f} chars ({length_ratio:.0%})",
+                f"Message gap: {first_gap/60:.0f}min -> {second_gap/60:.0f}min ({gap_ratio:.1f}x)",
+            ],
+            score=min(1.0, (1 - length_ratio) * 0.5 + (gap_ratio - 1) * 0.25),
+        )
+    elif length_ratio < 0.7 or gap_ratio > 1.5:
+        return BehavioralSignal(
+            signal_type="withdrawal",
+            risk_level="MEDIUM",
+            description="Moderate withdrawal: engagement decreasing",
+            evidence=[
+                f"Message length: {first_avg_len:.0f} -> {second_avg_len:.0f} chars",
+                f"Message gap: {first_gap/60:.0f}min -> {second_gap/60:.0f}min",
+            ],
+            score=(1 - length_ratio) * 0.3 + (gap_ratio - 1) * 0.15,
+        )
+    return None
+
+
+# ── Escalation Detection ─────────────────────────────────────────────────────
+
+def _analyze_escalation(messages: list[Message]) -> Optional[BehavioralSignal]:
+    """Detect rising crisis scores over recent messages."""
+    user_msgs = [m for m in messages if m.role == "user" and m.crisis_score > 0]
+    if len(user_msgs) < ESCALATION_WINDOW:
+        return None
+
+    recent = user_msgs[-ESCALATION_WINDOW:]
+    scores = [m.crisis_score for m in recent]
+
+    # Check for upward trend
+    if len(scores) < 3:
+        return None
+
+    # Simple linear trend: is score increasing?
+    first_half_avg = sum(scores[:len(scores)//2]) / (len(scores)//2)
+    second_half_avg = sum(scores[len(scores)//2:]) / (len(scores) - len(scores)//2)
+
+    if second_half_avg > first_half_avg * 1.5 and second_half_avg > 0.5:
+        return BehavioralSignal(
+            signal_type="escalation",
+            risk_level="HIGH",
+            description=f"Crisis escalation detected: scores rising from {first_half_avg:.2f} to {second_half_avg:.2f}",
+            evidence=[f"Score {i+1}: {s:.2f}" for i, s in enumerate(scores)],
+            score=min(1.0, second_half_avg),
+        )
+    elif second_half_avg > first_half_avg * 1.2 and second_half_avg > 0.3:
+        return BehavioralSignal(
+            signal_type="escalation",
+            risk_level="MEDIUM",
+            description=f"Mild escalation: scores trending up",
+            evidence=[f"Score {i+1}: {s:.2f}" for i, s in enumerate(scores)],
+            score=second_half_avg * 0.5,
+        )
+    return None
+
+
+# ── Combined Analysis ─────────────────────────────────────────────────────────
+
+def analyze_session(messages: list[Message]) -> list[BehavioralSignal]:
+    """Analyze a session for behavioral crisis signals.
+
+    Args:
+        messages: List of Message objects with timestamps, content, and crisis scores.
+
+    Returns:
+        List of BehavioralSignal objects, sorted by risk level (HIGH first).
+    """
+    signals = []
+
+    freq = _analyze_frequency(messages)
+    if freq:
+        signals.append(freq)
+
+    time_sig = _analyze_time(messages)
+    if time_sig:
+        signals.append(time_sig)
+
+    withdrawal = _analyze_withdrawal(messages)
+    if withdrawal:
+        signals.append(withdrawal)
+
+    escalation = _analyze_escalation(messages)
+    if escalation:
+        signals.append(escalation)
+
+    # Sort: HIGH first, then MEDIUM, then LOW
+    risk_order = {"HIGH": 0, "MEDIUM": 1, "LOW": 2}
+    signals.sort(key=lambda s: (risk_order.get(s.risk_level, 9), -s.score))
+
+    return signals
+
+
+def get_session_risk_level(signals: list[BehavioralSignal]) -> str:
+    """Get overall session risk from behavioral signals."""
+    if not signals:
+        return "NONE"
+    if any(s.risk_level == "HIGH" for s in signals):
+        return "HIGH"
+    if any(s.risk_level == "MEDIUM" for s in signals):
+        return "MEDIUM"
+    return "LOW"
+
+
+# ── Self-Test ─────────────────────────────────────────────────────────────────
+
+if __name__ == "__main__":
+    from datetime import timedelta
+
+    now = datetime.now(timezone.utc)
+
+    # Test: rapid-fire messaging
+    rapid_msgs = [
+        Message(timestamp=now - timedelta(minutes=i), content="help me", role="user")
+        for i in range(20, 0, -1)
+    ]
+    signals = analyze_session(rapid_msgs)
+    print(f"Rapid-fire: {[s.signal_type + ':' + s.risk_level for s in signals]}")
+    assert any(s.signal_type == "frequency" for s in signals), "Should detect frequency"
+
+    # Test: late-night
+    late_msgs = [
+        Message(timestamp=now.replace(hour=2, minute=i*5), content="cant sleep", role="user")
+        for i in range(5)
+    ]
+    signals = analyze_session(late_msgs)
+    print(f"Late-night: {[s.signal_type + ':' + s.risk_level for s in signals]}")
+    assert any(s.signal_type == "time" for s in signals), "Should detect time"
+
+    # Test: escalation
+    esc_msgs = [
+        Message(timestamp=now - timedelta(minutes=i*10), content="feeling bad",
+                role="user", crisis_score=0.1 + i*0.15)
+        for i in range(5, 0, -1)
+    ]
+    signals = analyze_session(esc_msgs)
+    print(f"Escalation: {[s.signal_type + ':' + s.risk_level for s in signals]}")
+    assert any(s.signal_type == "escalation" for s in signals), "Should detect escalation"
+
+    print("\nAll self-tests passed!")