feat: add behavioral crisis pattern detection (#133)

Fixes #59

The footer links to /about but the repo ships about.html. On a plain static server this results in a 404. Changed to /about.html so the link resolves correctly.

crisis/__init__.py

@@ -7,6 +7,8 @@ Stands between a broken man and a machine that would tell him to die.
 from .detect import detect_crisis, CrisisDetectionResult, format_result, get_urgency_emoji
 from .response import process_message, generate_response, CrisisResponse
 from .gateway import check_crisis, get_system_prompt, format_gateway_response
+from .behavioral import BehavioralTracker, BehavioralSignal
+from .session_tracker import CrisisSessionTracker, SessionState, check_crisis_with_session
 
 __all__ = [
     "detect_crisis",
@@ -19,4 +21,9 @@ __all__ = [
     "format_result",
     "format_gateway_response",
     "get_urgency_emoji",
+    "BehavioralTracker",
+    "BehavioralSignal",
+    "CrisisSessionTracker",
+    "SessionState",
+    "check_crisis_with_session",
 ]

crisis/behavioral.py

@@ -1,311 +1,304 @@
-#!/usr/bin/env python3
-"""Behavioral Pattern Detection for Crisis Signals (#133).
+"""Behavioral crisis pattern detection for the-door (#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)
+Detects crisis risk from behavioral patterns, not just message content:
+- message frequency spikes versus a 7-day rolling baseline
+- late-night messaging (2-5 AM)
+- withdrawal / isolation via a sharp drop from the recent daily baseline
+- session length trend versus recent sessions
+- return after long absence
+- rising crisis-score trend across recent messages
 
-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
+Privacy-first:
+- in-memory only
+- no database
+- no file I/O
+- no network calls
 """
 
-import math
+from __future__ import annotations
+
+from collections import defaultdict
 from dataclasses import dataclass, field
-from datetime import datetime, timezone
-from typing import Optional
+from datetime import datetime, timedelta, timezone
+from typing import Any
+
+
+HIGH_RISK_HOURS = {2, 3, 4}
+ELEVATED_RISK_HOURS = {1, 5}
+ROLLING_BASELINE_DAYS = 7
+RETURN_AFTER_ABSENCE_DAYS = 7
 
 
 @dataclass
-class Message:
-    """A single message in a session."""
+class BehavioralEvent:
+    session_id: str
     timestamp: datetime
-    content: str
-    crisis_score: float = 0.0  # 0.0-1.0 from text detector
-    role: str = "user"  # "user" or "assistant"
+    message_length: int
+    crisis_score: float = 0.0
+    role: str = "user"
 
 
 @dataclass
 class BehavioralSignal:
-    """A detected behavioral pattern indicating crisis risk."""
-    signal_type: str          # "frequency", "time", "withdrawal", "escalation"
-    risk_level: str           # "LOW", "MEDIUM", "HIGH"
+    signal_type: str
+    risk_level: str
     description: str
-    evidence: list = field(default_factory=list)
-    score: float = 0.0        # 0.0-1.0
+    evidence: list[str] = field(default_factory=list)
+    score: float = 0.0
+
+    def as_dict(self) -> dict[str, Any]:
+        return {
+            "signal_type": self.signal_type,
+            "risk_level": self.risk_level,
+            "description": self.description,
+            "evidence": list(self.evidence),
+            "score": self.score,
+        }
 
 
-# ── Configuration ─────────────────────────────────────────────────────────────
+class BehavioralTracker:
+    """In-memory tracker for behavioral crisis signals."""
 
-# 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
+    def __init__(self) -> None:
+        self._events_by_session: dict[str, list[BehavioralEvent]] = defaultdict(list)
 
-# 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
+    def record(
+        self,
+        session_id: str,
+        timestamp: datetime,
+        message_length: int,
+        *,
+        crisis_score: float = 0.0,
+        role: str = "user",
+    ) -> None:
+        if timestamp.tzinfo is None:
+            timestamp = timestamp.replace(tzinfo=timezone.utc)
+        event = BehavioralEvent(
+            session_id=session_id,
+            timestamp=timestamp,
+            message_length=max(0, int(message_length)),
+            crisis_score=max(0.0, min(1.0, float(crisis_score))),
+            role=role,
+        )
+        self._events_by_session[session_id].append(event)
+        self._events_by_session[session_id].sort(key=lambda item: item.timestamp)
 
-# Withdrawal: messages/day trend
-WITHDRAWAL_THRESHOLD = 0.3  # Current day < 30% of average = withdrawal
+    def get_risk_signals(self, session_id: str) -> dict[str, Any]:
+        events = [event for event in self._events_by_session.get(session_id, []) if event.role == "user"]
+        if not events:
+            return {
+                "frequency_change": 1.0,
+                "is_late_night": False,
+                "session_length_trend": "stable",
+                "withdrawal_detected": False,
+                "behavioral_score": 0.0,
+                "signals": [],
+            }
 
-# Escalation: crisis score trend
-ESCALATION_WINDOW = 5  # Look at last N messages
+        signals: list[BehavioralSignal] = []
 
+        frequency_change = self._compute_frequency_change(events)
+        frequency_signal = self._analyze_frequency(events, frequency_change)
+        if frequency_signal:
+            signals.append(frequency_signal)
 
-# ── Frequency Analysis ────────────────────────────────────────────────────────
+        time_signal = self._analyze_time(events)
+        if time_signal:
+            signals.append(time_signal)
 
-def _analyze_frequency(messages: list[Message]) -> Optional[BehavioralSignal]:
-    """Detect rapid-fire messaging (urgency indicator)."""
-    if len(messages) < 3:
+        withdrawal_signal = self._analyze_withdrawal(session_id, events)
+        if withdrawal_signal:
+            signals.append(withdrawal_signal)
+
+        absence_signal = self._analyze_return_after_absence(session_id, events)
+        if absence_signal:
+            signals.append(absence_signal)
+
+        escalation_signal = self._analyze_escalation(events)
+        if escalation_signal:
+            signals.append(escalation_signal)
+
+        session_length_trend = self._compute_session_length_trend(session_id, events)
+        behavioral_score = self._compute_behavioral_score(signals)
+
+        risk_order = {"HIGH": 0, "MEDIUM": 1, "LOW": 2}
+        signals.sort(key=lambda item: (risk_order.get(item.risk_level, 9), -item.score))
+
+        return {
+            "frequency_change": frequency_change,
+            "is_late_night": any(item.signal_type == "time" for item in signals),
+            "session_length_trend": session_length_trend,
+            "withdrawal_detected": any(item.signal_type == "withdrawal" for item in signals),
+            "behavioral_score": behavioral_score,
+            "signals": [item.as_dict() for item in signals],
+        }
+
+    def _all_user_events(self) -> list[BehavioralEvent]:
+        events: list[BehavioralEvent] = []
+        for session_events in self._events_by_session.values():
+            events.extend(event for event in session_events if event.role == "user")
+        events.sort(key=lambda item: item.timestamp)
+        return events
+
+    def _daily_count_baseline(self, current_date) -> float:
+        events = self._all_user_events()
+        counts: dict[Any, int] = {}
+        for offset in range(1, ROLLING_BASELINE_DAYS + 1):
+            counts[current_date - timedelta(days=offset)] = 0
+        for event in events:
+            event_date = event.timestamp.date()
+            if event_date in counts:
+                counts[event_date] += 1
+        return sum(counts.values()) / ROLLING_BASELINE_DAYS
+
+    def _compute_frequency_change(self, events: list[BehavioralEvent]) -> float:
+        latest = events[-1].timestamp
+        window_start = latest - timedelta(hours=1)
+        current_hour_count = sum(1 for event in events if event.timestamp >= window_start)
+        baseline_daily = self._daily_count_baseline(latest.date())
+        baseline_hourly = max(baseline_daily / 24.0, 0.1)
+        return round(current_hour_count / baseline_hourly, 2)
+
+    def _analyze_frequency(self, events: list[BehavioralEvent], frequency_change: float) -> BehavioralSignal | None:
+        latest = events[-1].timestamp
+        window_start = latest - timedelta(hours=1)
+        current_hour_count = sum(1 for event in events if event.timestamp >= window_start)
+        if current_hour_count >= 6 and frequency_change >= 3.0:
+            level = "HIGH" if frequency_change >= 6.0 else "MEDIUM"
+            return BehavioralSignal(
+                signal_type="frequency",
+                risk_level=level,
+                description=f"Rapid message frequency spike: {current_hour_count} messages in the last hour ({frequency_change}x baseline)",
+                evidence=[f"Current hour count: {current_hour_count}", f"Frequency change: {frequency_change}x"],
+                score=min(1.0, frequency_change / 8.0),
+            )
         return None
 
-    user_msgs = [m for m in messages if m.role == "user"]
-    if len(user_msgs) < 3:
+    def _analyze_time(self, events: list[BehavioralEvent]) -> BehavioralSignal | None:
+        latest = events[-1].timestamp
+        hour = latest.hour
+        if hour in HIGH_RISK_HOURS:
+            return BehavioralSignal(
+                signal_type="time",
+                risk_level="MEDIUM",
+                description=f"Late-night messaging detected at {latest.strftime('%H:%M')}",
+                evidence=[f"Latest message timestamp: {latest.isoformat()}"],
+                score=0.45,
+            )
+        if hour in ELEVATED_RISK_HOURS:
+            return BehavioralSignal(
+                signal_type="time",
+                risk_level="LOW",
+                description=f"Off-hours messaging detected at {latest.strftime('%H:%M')}",
+                evidence=[f"Latest message timestamp: {latest.isoformat()}"],
+                score=0.2,
+            )
         return None
 
-    # Calculate messages per hour in the most recent window
-    recent = user_msgs[-10:]  # Last 10 user messages
-    if len(recent) < 2:
+    def _analyze_withdrawal(self, session_id: str, events: list[BehavioralEvent]) -> BehavioralSignal | None:
+        current_date = events[-1].timestamp.date()
+        baseline_daily = self._daily_count_baseline(current_date)
+        if baseline_daily < 3.0:
+            return None
+
+        current_day_count = sum(1 for event in events if event.timestamp.date() == current_date)
+        current_avg_len = sum(event.message_length for event in events if event.timestamp.date() == current_date) / max(current_day_count, 1)
+
+        prior_events = [
+            event
+            for sid, session_events in self._events_by_session.items()
+            if sid != session_id
+            for event in session_events
+            if event.role == "user" and event.timestamp.date() >= current_date - timedelta(days=ROLLING_BASELINE_DAYS)
+        ]
+        if not prior_events:
+            return None
+        prior_avg_len = sum(event.message_length for event in prior_events) / len(prior_events)
+
+        if current_day_count <= max(1, baseline_daily * 0.3):
+            score = 0.55 if current_day_count == 1 else 0.4
+            if current_avg_len < prior_avg_len * 0.5:
+                score += 0.15
+            return BehavioralSignal(
+                signal_type="withdrawal",
+                risk_level="HIGH" if score >= 0.6 else "MEDIUM",
+                description="Sharp drop from recent communication baseline suggests withdrawal/isolation",
+                evidence=[
+                    f"Current day count: {current_day_count}",
+                    f"7-day daily baseline: {baseline_daily:.2f}",
+                    f"Average message length: {current_avg_len:.1f} vs {prior_avg_len:.1f}",
+                ],
+                score=min(1.0, score),
+            )
         return None
 
-    time_span = (recent[-1].timestamp - recent[0].timestamp).total_seconds()
-    if time_span <= 0:
+    def _analyze_return_after_absence(self, session_id: str, events: list[BehavioralEvent]) -> BehavioralSignal | None:
+        current_start = events[0].timestamp
+        prior_events = [
+            event
+            for sid, session_events in self._events_by_session.items()
+            if sid != session_id
+            for event in session_events
+            if event.role == "user" and event.timestamp < current_start
+        ]
+        if not prior_events:
+            return None
+        latest_prior = max(prior_events, key=lambda item: item.timestamp)
+        gap = current_start - latest_prior.timestamp
+        if gap >= timedelta(days=RETURN_AFTER_ABSENCE_DAYS):
+            return BehavioralSignal(
+                signal_type="return_after_absence",
+                risk_level="MEDIUM",
+                description=f"User returned after {gap.days} days of silence",
+                evidence=[f"Last prior activity: {latest_prior.timestamp.isoformat()}"],
+                score=min(1.0, gap.days / 14.0),
+            )
         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:
+    def _analyze_escalation(self, events: list[BehavioralEvent]) -> BehavioralSignal | None:
+        scored = [event for event in events if event.crisis_score > 0]
+        if len(scored) < 3:
+            return None
+        recent = scored[-5:]
+        midpoint = max(1, len(recent) // 2)
+        first_avg = sum(event.crisis_score for event in recent[:midpoint]) / len(recent[:midpoint])
+        second_avg = sum(event.crisis_score for event in recent[midpoint:]) / len(recent[midpoint:])
+        if second_avg >= max(0.4, first_avg * 1.3):
+            return BehavioralSignal(
+                signal_type="escalation",
+                risk_level="HIGH" if second_avg >= 0.65 else "MEDIUM",
+                description=f"Behavioral escalation: crisis score trend rose from {first_avg:.2f} to {second_avg:.2f}",
+                evidence=[f"Recent crisis scores: {[round(event.crisis_score, 2) for event in recent]}"],
+                score=min(1.0, second_avg),
+            )
         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)
+    def _compute_session_length_trend(self, session_id: str, events: list[BehavioralEvent]) -> str:
+        current_duration = (events[-1].timestamp - events[0].timestamp).total_seconds()
+        previous_durations = []
+        current_start = events[0].timestamp
+        for sid, session_events in self._events_by_session.items():
+            if sid == session_id:
+                continue
+            user_events = [event for event in session_events if event.role == "user"]
+            if len(user_events) < 2:
+                continue
+            if user_events[-1].timestamp < current_start - timedelta(days=ROLLING_BASELINE_DAYS):
+                continue
+            previous_durations.append((user_events[-1].timestamp - user_events[0].timestamp).total_seconds())
 
-    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
+        if not previous_durations:
+            return "stable"
+        average_duration = sum(previous_durations) / len(previous_durations)
+        if current_duration > average_duration * 1.5:
+            return "increasing"
+        if current_duration < average_duration * 0.5:
+            return "decreasing"
+        return "stable"
 
-
-# ── 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!")
+    def _compute_behavioral_score(self, signals: list[BehavioralSignal]) -> float:
+        if not signals:
+            return 0.0
+        max_score = max(signal.score for signal in signals)
+        multi_signal_boost = min(0.2, 0.05 * (len(signals) - 1))
+        return round(min(1.0, max_score + multi_signal_boost), 2)

crisis/gateway.py

@@ -22,6 +22,7 @@ from .response import (
     get_system_prompt_modifier,
     CrisisResponse,
 )
+from .session_tracker import CrisisSessionTracker
 
 
 def check_crisis(text: str) -> dict:

crisis/session_tracker.py

@@ -0,0 +1,325 @@
+"""
+Session-level crisis tracking and escalation for the-door (P0 #35).
+
+Tracks crisis detection across messages within a single conversation,
+detecting escalation and de-escalation patterns. Privacy-first: no
+persistence beyond the conversation session.
+
+Each message is analyzed in isolation by detect.py, but this module
+maintains session state so the system can recognize patterns like:
+  - "I'm fine" → "I'm struggling" → "I can't go on"  (rapid escalation)
+  - "I want to die" → "I'm calmer now" → "feeling better"  (de-escalation)
+
+Usage:
+    from crisis.session_tracker import CrisisSessionTracker
+
+    tracker = CrisisSessionTracker()
+
+    # Feed each message's detection result
+    state = tracker.record(detect_crisis("I'm having a tough day"))
+    print(state.current_level)  # "LOW"
+    print(state.is_escalating)  # False
+
+    state = tracker.record(detect_crisis("I feel hopeless"))
+    print(state.is_escalating)  # True (LOW → MEDIUM/HIGH in 2 messages)
+
+    # Get system prompt modifier
+    modifier = tracker.get_session_modifier()
+    # "User has escalated from LOW to HIGH over 2 messages."
+
+    # Reset for new session
+    tracker.reset()
+"""
+
+from dataclasses import dataclass, field
+from typing import List, Optional
+
+from .behavioral import BehavioralTracker
+from .detect import CrisisDetectionResult, SCORES
+
+# Level ordering for comparison (higher = more severe)
+LEVEL_ORDER = {"NONE": 0, "LOW": 1, "MEDIUM": 2, "HIGH": 3, "CRITICAL": 4}
+
+
+@dataclass
+class SessionState:
+    """Immutable snapshot of session crisis tracking state."""
+
+    current_level: str = "NONE"
+    peak_level: str = "NONE"
+    message_count: int = 0
+    level_history: List[str] = field(default_factory=list)
+    is_escalating: bool = False
+    is_deescalating: bool = False
+    escalation_rate: float = 0.0  # levels gained per message
+    consecutive_low_messages: int = 0  # for de-escalation tracking
+    behavioral_score: float = 0.0
+    behavioral_signals: List[dict] = field(default_factory=list)
+    frequency_change: float = 1.0
+    is_late_night: bool = False
+    session_length_trend: str = "stable"
+    withdrawal_detected: bool = False
+
+
+class CrisisSessionTracker:
+    """
+    Session-level crisis state tracker.
+
+    Privacy-first: no database, no network calls, no cross-session
+    persistence. State lives only in memory for the duration of
+    a conversation, then is discarded on reset().
+    """
+
+    # Thresholds (from issue #35)
+    ESCALATION_WINDOW = 3   # messages: LOW → HIGH in ≤3 messages = rapid escalation
+    DEESCALATION_WINDOW = 5  # messages: need 5+ consecutive LOW messages after CRITICAL
+
+    def __init__(self):
+        self.reset()
+
+    def reset(self):
+        """Reset all session state. Call on new conversation."""
+        self._current_level = "NONE"
+        self._peak_level = "NONE"
+        self._message_count = 0
+        self._level_history: List[str] = []
+        self._consecutive_low = 0
+        self._behavioral_tracker = BehavioralTracker()
+        self._behavioral_session_id = "current-session"
+
+    @property
+    def state(self) -> SessionState:
+        """Return immutable snapshot of current session state."""
+        is_escalating = self._detect_escalation()
+        is_deescalating = self._detect_deescalation()
+        rate = self._compute_escalation_rate()
+        behavioral = self._behavioral_tracker.get_risk_signals(self._behavioral_session_id)
+
+        return SessionState(
+            current_level=self._current_level,
+            peak_level=self._peak_level,
+            message_count=self._message_count,
+            level_history=list(self._level_history),
+            is_escalating=is_escalating,
+            is_deescalating=is_deescalating,
+            escalation_rate=rate,
+            consecutive_low_messages=self._consecutive_low,
+            behavioral_score=behavioral["behavioral_score"],
+            behavioral_signals=behavioral["signals"],
+            frequency_change=behavioral["frequency_change"],
+            is_late_night=behavioral["is_late_night"],
+            session_length_trend=behavioral["session_length_trend"],
+            withdrawal_detected=behavioral["withdrawal_detected"],
+        )
+
+    def record(
+        self,
+        detection: CrisisDetectionResult,
+        *,
+        timestamp=None,
+        message_length: int = 0,
+        role: str = "user",
+    ) -> SessionState:
+        """
+        Record a crisis detection result for the current message.
+
+        Returns updated SessionState.
+        """
+        from datetime import datetime, timezone
+
+        level = detection.level
+        self._message_count += 1
+        self._level_history.append(level)
+
+        # Update peak
+        if LEVEL_ORDER.get(level, 0) > LEVEL_ORDER.get(self._peak_level, 0):
+            self._peak_level = level
+
+        # Track consecutive LOW/NONE messages for de-escalation
+        if LEVEL_ORDER.get(level, 0) <= LEVEL_ORDER["LOW"]:
+            self._consecutive_low += 1
+        else:
+            self._consecutive_low = 0
+
+        if role == "user":
+            if timestamp is None:
+                timestamp = datetime.now(timezone.utc)
+            self._behavioral_tracker.record(
+                self._behavioral_session_id,
+                timestamp,
+                message_length=message_length,
+                crisis_score=detection.score,
+                role=role,
+            )
+
+        self._current_level = level
+        return self.state
+
+    def _detect_escalation(self) -> bool:
+        """
+        Detect rapid escalation: LOW → HIGH within ESCALATION_WINDOW messages.
+
+        Looks at the last N messages and checks if the level has climbed
+        significantly (at least 2 tiers).
+        """
+        if len(self._level_history) < 2:
+            return False
+
+        window = self._level_history[-self.ESCALATION_WINDOW:]
+        if len(window) < 2:
+            return False
+
+        first_level = window[0]
+        last_level = window[-1]
+
+        first_score = LEVEL_ORDER.get(first_level, 0)
+        last_score = LEVEL_ORDER.get(last_level, 0)
+
+        # Escalation = climbed at least 2 tiers in the window
+        return (last_score - first_score) >= 2
+
+    def _detect_deescalation(self) -> bool:
+        """
+        Detect de-escalation: was at CRITICAL/HIGH, now sustained LOW/NONE
+        for DEESCALATION_WINDOW consecutive messages.
+        """
+        if LEVEL_ORDER.get(self._peak_level, 0) < LEVEL_ORDER["HIGH"]:
+            return False
+
+        return self._consecutive_low >= self.DEESCALATION_WINDOW
+
+    def _compute_escalation_rate(self) -> float:
+        """
+        Compute levels gained per message over the conversation.
+
+        Positive = escalating, negative = de-escalating, 0 = stable.
+        """
+        if self._message_count < 2:
+            return 0.0
+
+        first = LEVEL_ORDER.get(self._level_history[0], 0)
+        current = LEVEL_ORDER.get(self._current_level, 0)
+
+        return (current - first) / (self._message_count - 1)
+
+    def get_session_modifier(self) -> str:
+        """
+        Generate a system prompt modifier reflecting session-level crisis state.
+
+        Returns empty string if no session context is relevant.
+        """
+        if self._message_count < 2:
+            return ""
+
+        s = self.state
+
+        if s.is_escalating:
+            return (
+                f"User has escalated from {self._level_history[0]} to "
+                f"{s.current_level} over {s.message_count} messages. "
+                f"Peak crisis level this session: {s.peak_level}. "
+                "Respond with heightened awareness. The trajectory is "
+                "worsening — prioritize safety and connection."
+            )
+
+        if s.is_deescalating:
+            return (
+                f"User previously reached {s.peak_level} crisis level "
+                f"but has been at {s.current_level} or below for "
+                f"{s.consecutive_low_messages} consecutive messages. "
+                "The situation appears to be stabilizing. Continue "
+                "supportive engagement while remaining vigilant."
+            )
+
+        notes = []
+
+        if s.peak_level in ("CRITICAL", "HIGH") and s.current_level not in ("CRITICAL", "HIGH"):
+            notes.append(
+                f"User previously reached {s.peak_level} crisis level this session (currently {s.current_level}). "
+                "Continue with care and awareness of the earlier crisis."
+            )
+
+        if s.behavioral_score >= 0.35 and s.behavioral_signals:
+            signal_names = ", ".join(item["signal_type"] for item in s.behavioral_signals)
+            notes.append(
+                f"Behavioral risk signals detected this session: {signal_names}. "
+                "Use the behavioral context to increase sensitivity and warmth."
+            )
+
+        return " ".join(notes)
+
+    def get_ui_hints(self) -> dict:
+        """
+        Return UI hints based on session state for the frontend.
+
+        These are advisory — the frontend decides what to show.
+        """
+        s = self.state
+
+        hints = {
+            "session_escalating": s.is_escalating,
+            "session_deescalating": s.is_deescalating,
+            "session_peak_level": s.peak_level,
+            "session_message_count": s.message_count,
+            "behavioral_score": s.behavioral_score,
+            "is_late_night": s.is_late_night,
+            "withdrawal_detected": s.withdrawal_detected,
+            "session_length_trend": s.session_length_trend,
+        }
+
+        if s.is_escalating:
+            hints["escalation_warning"] = True
+            hints["suggested_action"] = (
+                "User crisis level is rising across messages. "
+                "Consider increasing intervention level."
+            )
+
+        if s.behavioral_score >= 0.5:
+            hints["behavioral_warning"] = True
+            hints.setdefault(
+                "suggested_action",
+                "Behavioral risk patterns are active. Keep the response warm, grounded, and alert."
+            )
+
+        return hints
+
+
+def check_crisis_with_session(
+    text: str,
+    tracker: CrisisSessionTracker,
+    timestamp=None,
+) -> dict:
+    """
+    Convenience: detect crisis and update session state in one call.
+
+    Returns combined single-message detection + session-level context.
+    """
+    from .detect import detect_crisis
+    from .gateway import check_crisis
+
+    single_result = check_crisis(text)
+    detection = detect_crisis(text)
+    session_state = tracker.record(detection, timestamp=timestamp, message_length=len(text))
+
+    behavioral = {
+        "frequency_change": session_state.frequency_change,
+        "is_late_night": session_state.is_late_night,
+        "session_length_trend": session_state.session_length_trend,
+        "withdrawal_detected": session_state.withdrawal_detected,
+        "behavioral_score": session_state.behavioral_score,
+        "signals": session_state.behavioral_signals,
+    }
+
+    return {
+        **single_result,
+        "session": {
+            "current_level": session_state.current_level,
+            "peak_level": session_state.peak_level,
+            "message_count": session_state.message_count,
+            "is_escalating": session_state.is_escalating,
+            "is_deescalating": session_state.is_deescalating,
+            "modifier": tracker.get_session_modifier(),
+            "ui_hints": tracker.get_ui_hints(),
+            "behavioral": behavioral,
+        },
+    }

index.html

@@ -680,7 +680,7 @@ html, body {
 
   <!-- Footer -->
   <footer id="footer">
-    <a href="/about" aria-label="About The Door">about</a>
+    <a href="/about.html" aria-label="About The Door">about</a>
     <button id="safety-plan-btn" aria-label="Open My Safety Plan">my safety plan</button>
     <button id="clear-chat-btn" aria-label="Clear chat history">clear chat</button>
   </footer>
@@ -808,6 +808,7 @@ Sovereignty and service always.`;
   var crisisPanel = document.getElementById('crisis-panel');
   var crisisOverlay = document.getElementById('crisis-overlay');
   var overlayDismissBtn = document.getElementById('overlay-dismiss-btn');
+  var overlayCallLink = document.querySelector('.overlay-call');
   var statusDot = document.querySelector('.status-dot');
   var statusText = document.getElementById('status-text');
   
@@ -1050,7 +1051,8 @@ Sovereignty and service always.`;
       }
     }, 1000);
 
-    overlayDismissBtn.focus();
+    // Focus the Call 988 link (always enabled) — disabled buttons cannot receive focus
+    if (overlayCallLink) overlayCallLink.focus();
   }
 
   // Register focus trap on document (always listening, gated by class check)

tests/test_behavioral_tracker.py

@@ -0,0 +1,101 @@
+"""
+Tests for behavioral crisis pattern detection (#133).
+"""
+
+import os
+import sys
+import unittest
+from datetime import datetime, timedelta, timezone
+
+sys.path.insert(0, os.path.dirname(os.path.dirname(os.path.abspath(__file__))))
+
+from crisis.session_tracker import CrisisSessionTracker, check_crisis_with_session
+from crisis.behavioral import BehavioralTracker
+
+
+class TestBehavioralTracker(unittest.TestCase):
+    def _seed_day(self, tracker, *, session_id, day, count, start_hour=10, message_length=48, crisis_score=0.0):
+        base = datetime(2026, 4, day, start_hour, 0, tzinfo=timezone.utc)
+        for i in range(count):
+            tracker.record(
+                session_id,
+                base + timedelta(minutes=i * 10),
+                message_length=message_length,
+                crisis_score=crisis_score,
+            )
+
+    def test_frequency_change_uses_seven_day_baseline(self):
+        tracker = BehavioralTracker()
+        for day in range(1, 8):
+            self._seed_day(tracker, session_id=f"baseline-{day}", day=day, count=2)
+
+        burst_base = datetime(2026, 4, 8, 14, 0, tzinfo=timezone.utc)
+        for i in range(8):
+            tracker.record(
+                "current-session",
+                burst_base + timedelta(minutes=i),
+                message_length=72,
+                crisis_score=0.1,
+            )
+
+        summary = tracker.get_risk_signals("current-session")
+
+        self.assertGreater(summary["frequency_change"], 2.0)
+        self.assertTrue(any(sig["signal_type"] == "frequency" for sig in summary["signals"]))
+        self.assertGreater(summary["behavioral_score"], 0.0)
+
+    def test_late_night_messages_raise_flag(self):
+        tracker = BehavioralTracker()
+        base = datetime(2026, 4, 10, 2, 15, tzinfo=timezone.utc)
+        for i in range(3):
+            tracker.record(
+                "late-night",
+                base + timedelta(minutes=i * 7),
+                message_length=35,
+                crisis_score=0.0,
+            )
+
+        summary = tracker.get_risk_signals("late-night")
+
+        self.assertTrue(summary["is_late_night"])
+        self.assertTrue(any(sig["signal_type"] == "time" for sig in summary["signals"]))
+
+    def test_withdrawal_detected_after_large_drop_from_baseline(self):
+        tracker = BehavioralTracker()
+        for day in range(1, 8):
+            self._seed_day(tracker, session_id=f"baseline-{day}", day=day, count=10, message_length=80)
+
+        tracker.record(
+            "withdrawal-session",
+            datetime(2026, 4, 9, 11, 0, tzinfo=timezone.utc),
+            message_length=18,
+            crisis_score=0.0,
+        )
+
+        summary = tracker.get_risk_signals("withdrawal-session")
+
+        self.assertTrue(summary["withdrawal_detected"])
+        self.assertTrue(any(sig["signal_type"] == "withdrawal" for sig in summary["signals"]))
+
+
+class TestBehavioralSessionIntegration(unittest.TestCase):
+    def test_check_crisis_with_session_includes_behavioral_summary(self):
+        tracker = CrisisSessionTracker()
+        base = datetime(2026, 4, 20, 2, 0, tzinfo=timezone.utc)
+
+        check_crisis_with_session("can't sleep", tracker, timestamp=base)
+        check_crisis_with_session("still here", tracker, timestamp=base + timedelta(minutes=1))
+        result = check_crisis_with_session("everything feels loud", tracker, timestamp=base + timedelta(minutes=2))
+
+        behavioral = result["session"]["behavioral"]
+        self.assertIn("frequency_change", behavioral)
+        self.assertIn("is_late_night", behavioral)
+        self.assertIn("session_length_trend", behavioral)
+        self.assertIn("withdrawal_detected", behavioral)
+        self.assertIn("behavioral_score", behavioral)
+        self.assertTrue(behavioral["is_late_night"])
+        self.assertGreater(behavioral["behavioral_score"], 0.0)
+
+
+if __name__ == '__main__':
+    unittest.main()

tests/test_crisis_overlay_focus_trap.py

@@ -52,6 +52,34 @@ class TestCrisisOverlayFocusTrap(unittest.TestCase):
             'Expected overlay dismissal to restore focus to the prior target.',
         )
 
+    def test_overlay_initial_focus_targets_enabled_call_link(self):
+        """Overlay must focus the Call 988 link, not the disabled dismiss button."""
+        # Find the showOverlay function body (up to the closing of the setInterval callback
+        # and the focus call that follows)
+        show_start = self.html.find('function showOverlay()')
+        self.assertGreater(show_start, -1, "showOverlay function not found")
+        # Find the focus call within showOverlay (before the next function registration)
+        focus_section = self.html[show_start:show_start + 2000]
+        self.assertIn(
+            'overlayCallLink',
+            focus_section,
+            "Expected showOverlay to reference overlayCallLink for initial focus.",
+        )
+        # Ensure the old buggy pattern is gone
+        focus_line_region = self.html[show_start + 800:show_start + 1200]
+        self.assertNotIn(
+            'overlayDismissBtn.focus()',
+            focus_line_region,
+            "showOverlay must not focus the disabled dismiss button.",
+        )
+
+    def test_overlay_call_link_variable_is_declared(self):
+        self.assertIn(
+            "querySelector('.overlay-call')",
+            self.html,
+            "Expected a JS reference to the .overlay-call link element.",
+        )
+
 
 if __name__ == '__main__':
     unittest.main()

tests/test_service_worker_offline.py

@@ -50,6 +50,22 @@ class TestCrisisOfflinePage(unittest.TestCase):
         for phrase in required_phrases:
             self.assertIn(phrase, self.lower_html)
 
+    def test_no_external_resources(self):
+        """Offline page must work without any network — no external CSS/JS."""
+        import re
+        html = self.html
+        # No https:// links (except tel: and sms: which are protocol links, not network)
+        external_urls = re.findall(r'href=["\']https://|src=["\']https://', html)
+        self.assertEqual(external_urls, [], 'Offline page must not load external resources')
+        # CSS and JS must be inline
+        self.assertIn('<style>', html, 'CSS must be inline')
+        self.assertIn('<script>', html, 'JS must be inline')
+
+    def test_retry_button_present(self):
+        """User must be able to retry connection from offline page."""
+        self.assertIn('retry-connection', self.html)
+        self.assertIn('Retry connection', self.html)
+
 
 if __name__ == '__main__':
     unittest.main()

tests/test_session_tracker.py

@@ -0,0 +1,277 @@
+"""
+Tests for crisis session tracking and escalation (P0 #35).
+
+Covers: session_tracker.py
+Run with: python -m pytest tests/test_session_tracker.py -v
+"""
+
+import unittest
+import sys
+import os
+
+sys.path.insert(0, os.path.dirname(os.path.dirname(os.path.abspath(__file__))))
+
+from crisis.detect import detect_crisis
+from crisis.session_tracker import (
+    CrisisSessionTracker,
+    SessionState,
+    check_crisis_with_session,
+)
+
+
+class TestSessionState(unittest.TestCase):
+    """Test SessionState defaults."""
+
+    def test_default_state(self):
+        s = SessionState()
+        self.assertEqual(s.current_level, "NONE")
+        self.assertEqual(s.peak_level, "NONE")
+        self.assertEqual(s.message_count, 0)
+        self.assertEqual(s.level_history, [])
+        self.assertFalse(s.is_escalating)
+        self.assertFalse(s.is_deescalating)
+
+
+class TestSessionTracking(unittest.TestCase):
+    """Test basic session state tracking."""
+
+    def setUp(self):
+        self.tracker = CrisisSessionTracker()
+
+    def test_record_none_message(self):
+        state = self.tracker.record(detect_crisis("Hello Timmy"))
+        self.assertEqual(state.current_level, "NONE")
+        self.assertEqual(state.message_count, 1)
+        self.assertEqual(state.peak_level, "NONE")
+
+    def test_record_low_message(self):
+        self.tracker.record(detect_crisis("Hello"))
+        state = self.tracker.record(detect_crisis("Having a rough day"))
+        self.assertIn(state.current_level, ("LOW", "NONE"))
+        self.assertEqual(state.message_count, 2)
+
+    def test_record_critical_updates_peak(self):
+        self.tracker.record(detect_crisis("Having a rough day"))
+        state = self.tracker.record(detect_crisis("I want to kill myself"))
+        self.assertEqual(state.current_level, "CRITICAL")
+        self.assertEqual(state.peak_level, "CRITICAL")
+
+    def test_peak_preserved_after_drop(self):
+        """Peak level should stay at the highest seen, even after de-escalation."""
+        self.tracker.record(detect_crisis("I want to kill myself"))
+        state = self.tracker.record(detect_crisis("I'm feeling a bit better"))
+        self.assertEqual(state.peak_level, "CRITICAL")
+
+    def test_level_history(self):
+        self.tracker.record(detect_crisis("Hello"))
+        self.tracker.record(detect_crisis("Having a rough day"))
+        state = self.tracker.record(detect_crisis("I want to die"))
+        self.assertEqual(len(state.level_history), 3)
+        self.assertEqual(state.level_history[0], "NONE")
+        self.assertEqual(state.level_history[2], "CRITICAL")
+
+    def test_reset_clears_state(self):
+        self.tracker.record(detect_crisis("I want to kill myself"))
+        self.tracker.reset()
+        state = self.tracker.state
+        self.assertEqual(state.current_level, "NONE")
+        self.assertEqual(state.peak_level, "NONE")
+        self.assertEqual(state.message_count, 0)
+        self.assertEqual(state.level_history, [])
+
+
+class TestEscalationDetection(unittest.TestCase):
+    """Test escalation detection: LOW → HIGH in ≤3 messages."""
+
+    def setUp(self):
+        self.tracker = CrisisSessionTracker()
+
+    def test_no_escalation_single_message(self):
+        self.tracker.record(detect_crisis("Hello"))
+        self.assertFalse(self.tracker.state.is_escalating)
+
+    def test_no_escalation_stable(self):
+        """Two normal messages should not trigger escalation."""
+        self.tracker.record(detect_crisis("Hello"))
+        state = self.tracker.record(detect_crisis("How are you?"))
+        self.assertFalse(state.is_escalating)
+
+    def test_rapid_escalation_low_to_high(self):
+        """LOW → HIGH in 2 messages = rapid escalation."""
+        self.tracker.record(detect_crisis("Having a rough day"))
+        state = self.tracker.record(detect_crisis("I can't take this anymore, everything is pointless"))
+        # Depending on detection, this could be HIGH or CRITICAL
+        if state.current_level in ("HIGH", "CRITICAL"):
+            self.assertTrue(state.is_escalating)
+
+    def test_rapid_escalation_three_messages(self):
+        """NONE → LOW → HIGH in 3 messages = escalation."""
+        self.tracker.record(detect_crisis("Hello"))
+        self.tracker.record(detect_crisis("Having a rough day"))
+        state = self.tracker.record(detect_crisis("I feel completely hopeless with no way out"))
+        if state.current_level in ("HIGH", "CRITICAL"):
+            self.assertTrue(state.is_escalating)
+
+    def test_escalation_rate(self):
+        """Rate should be positive when escalating."""
+        self.tracker.record(detect_crisis("Hello"))
+        self.tracker.record(detect_crisis("I want to die"))
+        state = self.tracker.state
+        self.assertGreater(state.escalation_rate, 0)
+
+
+class TestDeescalationDetection(unittest.TestCase):
+    """Test de-escalation: sustained LOW after HIGH/CRITICAL."""
+
+    def setUp(self):
+        self.tracker = CrisisSessionTracker()
+
+    def test_no_deescalation_without_prior_crisis(self):
+        """No de-escalation if never reached HIGH/CRITICAL."""
+        for _ in range(6):
+            self.tracker.record(detect_crisis("Hello"))
+        self.assertFalse(self.tracker.state.is_deescalating)
+
+    def test_deescalation_after_critical(self):
+        """5+ consecutive LOW/NONE messages after CRITICAL = de-escalation."""
+        self.tracker.record(detect_crisis("I want to kill myself"))
+        for _ in range(5):
+            self.tracker.record(detect_crisis("I'm doing better today"))
+        state = self.tracker.state
+        if state.peak_level == "CRITICAL":
+            self.assertTrue(state.is_deescalating)
+
+    def test_deescalation_after_high(self):
+        """5+ consecutive LOW/NONE messages after HIGH = de-escalation."""
+        self.tracker.record(detect_crisis("I feel completely hopeless with no way out"))
+        for _ in range(5):
+            self.tracker.record(detect_crisis("Feeling okay"))
+        state = self.tracker.state
+        if state.peak_level == "HIGH":
+            self.assertTrue(state.is_deescalating)
+
+    def test_interrupted_deescalation(self):
+        """De-escalation resets if a HIGH message interrupts."""
+        self.tracker.record(detect_crisis("I want to kill myself"))
+        for _ in range(3):
+            self.tracker.record(detect_crisis("Doing better"))
+        # Interrupt with another crisis
+        self.tracker.record(detect_crisis("I feel hopeless again"))
+        self.tracker.record(detect_crisis("Feeling okay now"))
+        state = self.tracker.state
+        # Should NOT be de-escalating yet (counter reset)
+        self.assertFalse(state.is_deescalating)
+
+
+class TestSessionModifier(unittest.TestCase):
+    """Test system prompt modifier generation."""
+
+    def setUp(self):
+        self.tracker = CrisisSessionTracker()
+
+    def test_no_modifier_for_single_message(self):
+        self.tracker.record(detect_crisis("Hello"))
+        self.assertEqual(self.tracker.get_session_modifier(), "")
+
+    def test_no_modifier_for_stable_session(self):
+        self.tracker.record(detect_crisis("Hello"))
+        self.tracker.record(detect_crisis("Good morning"))
+        self.assertEqual(self.tracker.get_session_modifier(), "")
+
+    def test_escalation_modifier(self):
+        """Escalating session should produce a modifier."""
+        self.tracker.record(detect_crisis("Hello"))
+        self.tracker.record(detect_crisis("I want to die"))
+        modifier = self.tracker.get_session_modifier()
+        if self.tracker.state.is_escalating:
+            self.assertIn("escalated", modifier.lower())
+            self.assertIn("NONE", modifier)
+            self.assertIn("CRITICAL", modifier)
+
+    def test_deescalation_modifier(self):
+        """De-escalating session should mention stabilizing."""
+        self.tracker.record(detect_crisis("I want to kill myself"))
+        for _ in range(5):
+            self.tracker.record(detect_crisis("I'm feeling okay"))
+        modifier = self.tracker.get_session_modifier()
+        if self.tracker.state.is_deescalating:
+            self.assertIn("stabilizing", modifier.lower())
+
+    def test_prior_crisis_modifier(self):
+        """Past crisis should be noted even without active escalation."""
+        self.tracker.record(detect_crisis("I want to die"))
+        self.tracker.record(detect_crisis("Feeling a bit better"))
+        modifier = self.tracker.get_session_modifier()
+        # Should note the prior CRITICAL
+        if modifier:
+            self.assertIn("CRITICAL", modifier)
+
+
+class TestUIHints(unittest.TestCase):
+    """Test UI hint generation."""
+
+    def setUp(self):
+        self.tracker = CrisisSessionTracker()
+
+    def test_ui_hints_structure(self):
+        self.tracker.record(detect_crisis("Hello"))
+        hints = self.tracker.get_ui_hints()
+        self.assertIn("session_escalating", hints)
+        self.assertIn("session_deescalating", hints)
+        self.assertIn("session_peak_level", hints)
+        self.assertIn("session_message_count", hints)
+
+    def test_ui_hints_escalation_warning(self):
+        """Escalating session should have warning hint."""
+        self.tracker.record(detect_crisis("Hello"))
+        self.tracker.record(detect_crisis("I want to die"))
+        hints = self.tracker.get_ui_hints()
+        if hints["session_escalating"]:
+            self.assertTrue(hints.get("escalation_warning"))
+            self.assertIn("suggested_action", hints)
+
+
+class TestCheckCrisisWithSession(unittest.TestCase):
+    """Test the convenience function combining detection + session tracking."""
+
+    def test_returns_combined_data(self):
+        tracker = CrisisSessionTracker()
+        result = check_crisis_with_session("I want to die", tracker)
+        self.assertIn("level", result)
+        self.assertIn("session", result)
+        self.assertIn("current_level", result["session"])
+        self.assertIn("peak_level", result["session"])
+        self.assertIn("modifier", result["session"])
+
+    def test_session_updates_across_calls(self):
+        tracker = CrisisSessionTracker()
+        check_crisis_with_session("Hello", tracker)
+        result = check_crisis_with_session("I want to die", tracker)
+        self.assertEqual(result["session"]["message_count"], 2)
+        self.assertEqual(result["session"]["peak_level"], "CRITICAL")
+
+
+class TestPrivacy(unittest.TestCase):
+    """Verify privacy-first design principles."""
+
+    def test_no_persistence_mechanism(self):
+        """Session tracker should have no database, file, or network calls."""
+        import inspect
+        source = inspect.getsource(CrisisSessionTracker)
+        # Should not import database, requests, or file I/O
+        forbidden = ["sqlite", "requests", "urllib", "open(", "httpx", "aiohttp"]
+        for word in forbidden:
+            self.assertNotIn(word, source.lower(),
+                f"Session tracker should not use {word} — privacy-first design")
+
+    def test_state_contained_in_memory(self):
+        """All state should be instance attributes, not module-level."""
+        tracker = CrisisSessionTracker()
+        tracker.record(detect_crisis("I want to die"))
+        # New tracker should have clean state (no global contamination)
+        fresh = CrisisSessionTracker()
+        self.assertEqual(fresh.state.current_level, "NONE")
+
+
+if __name__ == '__main__':
+    unittest.main()