feat: add unified multimodal scorer for #134

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

@@ -6,7 +6,8 @@ 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 .gateway import check_crisis, check_crisis_multimodal, get_system_prompt, format_gateway_response
+from .session_tracker import CrisisSessionTracker, SessionState, check_crisis_with_session
 
 __all__ = [
     "detect_crisis",
@@ -15,8 +16,12 @@ __all__ = [
     "generate_response",
     "CrisisResponse",
     "check_crisis",
+    "check_crisis_multimodal",
     "get_system_prompt",
     "format_result",
     "format_gateway_response",
     "get_urgency_emoji",
+    "CrisisSessionTracker",
+    "SessionState",
+    "check_crisis_with_session",
 ]

crisis/behavioral.py

@@ -1,311 +0,0 @@
-#!/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!")

crisis/gateway.py

@@ -2,18 +2,21 @@
 Crisis Gateway Module for the-door.
 
 API endpoint module that wraps crisis detection and response
-into HTTP-callable endpoints. Integrates detect.py and response.py.
+into HTTP-callable endpoints. Integrates detect.py, unified_scorer.py, and response.py.
 
 Usage:
     from crisis.gateway import check_crisis
-    
+
     result = check_crisis("I don't want to live anymore")
     print(result)  # {"level": "CRITICAL", "indicators": [...], "response": {...}}
 """
 
 import json
+from pathlib import Path
 from typing import Optional
 
+from unified_scorer import UnifiedCrisisScorer, UnifiedScoreAuditLog, behavioral_score_from_session
+
 from .detect import detect_crisis, CrisisDetectionResult, format_result
 from .compassion_router import router
 from .response import (
@@ -22,6 +25,7 @@ from .response import (
     get_system_prompt_modifier,
     CrisisResponse,
 )
+from .session_tracker import CrisisSessionTracker
 
 
 def check_crisis(text: str) -> dict:
@@ -49,6 +53,74 @@ def check_crisis(text: str) -> dict:
     }
 
 
+def check_crisis_multimodal(
+    text: str,
+    *,
+    tracker: Optional[CrisisSessionTracker] = None,
+    voice_score: Optional[float] = None,
+    image_score: Optional[float] = None,
+    behavioral_score: Optional[float] = None,
+    audit_log_path: Optional[Path] = None,
+    weights: Optional[dict] = None,
+) -> dict:
+    """Combine text, voice, image, and behavioral signals into one crisis assessment."""
+    detection = detect_crisis(text)
+    session_state = tracker.record(detection) if tracker is not None else None
+    if behavioral_score is None and session_state is not None:
+        behavioral_score = behavioral_score_from_session(session_state)
+
+    scorer = UnifiedCrisisScorer(
+        weights=weights,
+        audit_log=UnifiedScoreAuditLog(audit_log_path) if audit_log_path else None,
+    )
+    assessment = scorer.score(
+        text_score=detection.score,
+        voice_score=voice_score,
+        image_score=image_score,
+        behavioral_score=behavioral_score,
+        source_text=text,
+    )
+
+    unified_detection = CrisisDetectionResult(
+        level=assessment.level.value,
+        indicators=detection.indicators,
+        recommended_action=detection.recommended_action,
+        score=assessment.combined_score,
+        matches=detection.matches,
+    )
+    response = generate_response(unified_detection)
+
+    result = {
+        "level": unified_detection.level,
+        "score": unified_detection.score,
+        "indicators": detection.indicators,
+        "recommended_action": unified_detection.recommended_action,
+        "timmy_message": response.timmy_message,
+        "ui": {
+            "show_crisis_panel": response.show_crisis_panel,
+            "show_overlay": response.show_overlay,
+            "provide_988": response.provide_988,
+        },
+        "escalate": response.escalate,
+        "unified": {
+            "level": assessment.level.value,
+            "combined_score": assessment.combined_score,
+            "weights": assessment.weights,
+            "modalities": assessment.modalities,
+            "present_modalities": assessment.present_modalities,
+        },
+    }
+    if session_state is not None:
+        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,
+        }
+    return result
+
+
 def get_system_prompt(base_prompt: str, text: str = "") -> str:
     """
     Sovereign Heart System Prompt Override.

crisis/session_tracker.py

@@ -0,0 +1,259 @@
+"""
+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 .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
+
+
+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
+
+    @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()
+
+        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,
+        )
+
+    def record(self, detection: CrisisDetectionResult) -> SessionState:
+        """
+        Record a crisis detection result for the current message.
+
+        Returns updated SessionState.
+        """
+        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
+
+        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."
+            )
+
+        if s.peak_level in ("CRITICAL", "HIGH") and s.current_level not in ("CRITICAL", "HIGH"):
+            return (
+                f"User previously reached {s.peak_level} crisis level "
+                f"this session (currently {s.current_level}). "
+                "Continue with care and awareness of the earlier crisis."
+            )
+
+        return ""
+
+    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,
+        }
+
+        if s.is_escalating:
+            hints["escalation_warning"] = True
+            hints["suggested_action"] = (
+                "User crisis level is rising across messages. "
+                "Consider increasing intervention level."
+            )
+
+        return hints
+
+
+def check_crisis_with_session(
+    text: str,
+    tracker: CrisisSessionTracker,
+) -> 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)
+
+    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(),
+        },
+    }

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_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()

tests/test_unified_gateway.py

@@ -0,0 +1,19 @@
+from crisis.gateway import check_crisis_multimodal
+from crisis.session_tracker import CrisisSessionTracker
+
+
+def test_multimodal_gateway_uses_unified_score_for_988_ui(tmp_path):
+    tracker = CrisisSessionTracker()
+    result = check_crisis_multimodal(
+        "I want to kill myself tonight",
+        tracker=tracker,
+        voice_score=0.92,
+        image_score=0.6,
+        audit_log_path=tmp_path / "audit.jsonl",
+    )
+
+    assert result["unified"]["level"] == "CRITICAL"
+    assert result["ui"]["provide_988"] is True
+    assert result["ui"]["show_overlay"] is True
+    assert result["unified"]["modalities"]["voice"] == 0.92
+    assert result["unified"]["modalities"]["behavioral"] >= 0.0

tests/test_unified_scorer.py

@@ -0,0 +1,51 @@
+from pathlib import Path
+
+from unified_scorer import (
+    CrisisLevel,
+    UnifiedCrisisScorer,
+    UnifiedScoreAuditLog,
+    behavioral_score_from_session,
+)
+from crisis.session_tracker import SessionState
+
+
+def test_unified_scorer_renormalizes_available_modalities_and_escalates():
+    scorer = UnifiedCrisisScorer()
+    assessment = scorer.score(
+        text_score=1.0,
+        voice_score=0.8,
+        image_score=None,
+        behavioral_score=0.7,
+    )
+
+    assert assessment.level is CrisisLevel.CRITICAL
+    assert assessment.combined_score > 0.8
+    assert assessment.present_modalities == ["text", "voice", "behavioral"]
+
+
+def test_behavioral_score_rises_for_escalating_session_state():
+    session = SessionState(
+        current_level="HIGH",
+        peak_level="CRITICAL",
+        message_count=4,
+        level_history=["LOW", "MEDIUM", "HIGH", "CRITICAL"],
+        is_escalating=True,
+        is_deescalating=False,
+        escalation_rate=1.0,
+        consecutive_low_messages=0,
+    )
+
+    assert behavioral_score_from_session(session) >= 0.8
+
+
+def test_audit_log_persists_anonymized_score_entries(tmp_path):
+    log_path = tmp_path / "unified-score-audit.jsonl"
+    scorer = UnifiedCrisisScorer(audit_log=UnifiedScoreAuditLog(log_path))
+    scorer.score(text_score=0.75, voice_score=0.2, image_score=0.1, behavioral_score=0.6, source_text="I feel trapped and hopeless")
+
+    lines = log_path.read_text().strip().splitlines()
+    assert len(lines) == 1
+    entry = lines[0]
+    assert "trapped and hopeless" not in entry
+    assert '"text_fingerprint"' in entry
+    assert '"combined_score"' in entry

unified_scorer.py

@@ -0,0 +1,126 @@
+"""Unified multimodal crisis scoring for the-door."""
+
+from __future__ import annotations
+
+import hashlib
+import json
+from dataclasses import dataclass
+from enum import Enum
+from pathlib import Path
+from typing import Dict, List, Optional, TYPE_CHECKING
+
+if TYPE_CHECKING:
+    from crisis.session_tracker import SessionState
+
+
+SCORE_BY_LEVEL = {"NONE": 0.0, "LOW": 0.25, "MEDIUM": 0.5, "HIGH": 0.75, "CRITICAL": 1.0}
+LEVEL_RANK = {"NONE": 0, "LOW": 1, "MEDIUM": 2, "HIGH": 3, "CRITICAL": 4}
+
+
+class CrisisLevel(Enum):
+    NONE = "NONE"
+    LOW = "LOW"
+    MEDIUM = "MEDIUM"
+    HIGH = "HIGH"
+    CRITICAL = "CRITICAL"
+
+
+DEFAULT_WEIGHTS: Dict[str, float] = {
+    "text": 0.4,
+    "voice": 0.25,
+    "behavioral": 0.2,
+    "image": 0.15,
+}
+
+
+@dataclass
+class UnifiedAssessment:
+    level: CrisisLevel
+    combined_score: float
+    weights: Dict[str, float]
+    modalities: Dict[str, Optional[float]]
+    present_modalities: List[str]
+
+
+class UnifiedScoreAuditLog:
+    def __init__(self, path: Path | str):
+        self.path = Path(path)
+
+    def record(self, assessment: UnifiedAssessment, source_text: str = "") -> None:
+        self.path.parent.mkdir(parents=True, exist_ok=True)
+        fingerprint = hashlib.sha256(source_text.encode("utf-8")).hexdigest()[:12] if source_text else None
+        payload = {
+            "level": assessment.level.value,
+            "combined_score": round(assessment.combined_score, 4),
+            "weights": assessment.weights,
+            "modalities": assessment.modalities,
+            "present_modalities": assessment.present_modalities,
+            "text_fingerprint": fingerprint,
+        }
+        with self.path.open("a", encoding="utf-8") as fh:
+            fh.write(json.dumps(payload, sort_keys=True) + "\n")
+
+
+class UnifiedCrisisScorer:
+    def __init__(self, weights: Optional[Dict[str, float]] = None, audit_log: Optional[UnifiedScoreAuditLog] = None):
+        self.weights = dict(DEFAULT_WEIGHTS)
+        if weights:
+            self.weights.update(weights)
+        self.audit_log = audit_log
+
+    def _normalize(self, modalities: Dict[str, Optional[float]]) -> Dict[str, float]:
+        present = [name for name, score in modalities.items() if score is not None]
+        if not present:
+            return {}
+        total = sum(self.weights[name] for name in present)
+        return {name: self.weights[name] / total for name in present}
+
+    def _level_for_score(self, score: float) -> CrisisLevel:
+        if score > 0.8:
+            return CrisisLevel.CRITICAL
+        if score > 0.6:
+            return CrisisLevel.HIGH
+        if score > 0.4:
+            return CrisisLevel.MEDIUM
+        if score > 0.0:
+            return CrisisLevel.LOW
+        return CrisisLevel.NONE
+
+    def score(
+        self,
+        *,
+        text_score: Optional[float],
+        voice_score: Optional[float] = None,
+        image_score: Optional[float] = None,
+        behavioral_score: Optional[float] = None,
+        source_text: str = "",
+    ) -> UnifiedAssessment:
+        modalities = {
+            "text": text_score,
+            "voice": voice_score,
+            "behavioral": behavioral_score,
+            "image": image_score,
+        }
+        normalized = self._normalize(modalities)
+        combined = 0.0
+        for name, weight in normalized.items():
+            combined += float(modalities[name]) * weight
+        assessment = UnifiedAssessment(
+            level=self._level_for_score(combined),
+            combined_score=combined,
+            weights=normalized,
+            modalities=modalities,
+            present_modalities=[name for name, score in modalities.items() if score is not None],
+        )
+        if self.audit_log:
+            self.audit_log.record(assessment, source_text=source_text)
+        return assessment
+
+
+def behavioral_score_from_session(session: 'SessionState') -> float:
+    current = SCORE_BY_LEVEL.get(session.current_level, 0.0)
+    peak_bonus = 0.1 if LEVEL_RANK.get(session.peak_level, 0) >= LEVEL_RANK["HIGH"] else 0.0
+    escalation_bonus = 0.15 if session.is_escalating else 0.0
+    rate_bonus = min(max(session.escalation_rate, 0.0), 1.0) * 0.1
+    deescalation_penalty = 0.15 if session.is_deescalating else 0.0
+    return max(0.0, min(1.0, current + peak_bonus + escalation_bonus + rate_bonus - deescalation_penalty))