feat: CLI command to view crisis metrics summary (#136)

crisis/metrics.py:
  CrisisMetrics class — aggregate crisis detection metrics
  Privacy-first: stores only counts, never user content
  Daily JSONL files in ~/.the-door/metrics/
  get_summary(days) → AggregateMetrics
  get_report(days) → human-readable report
  get_json(days) → JSON export
  CLI: python3 -m crisis.metrics --summary/--json

crisis/__init__.py:
  Export CrisisMetrics, AggregateMetrics

Makefile:
  make metrics      → summary report
  make metrics-json → JSON export

tests/test_crisis_metrics.py: 6 tests
  record_session, summary, report, JSON export

Makefile

@@ -12,7 +12,7 @@ VPS := alexanderwhitestone.com
 DOMAIN := alexanderwhitestone.com
 DEPLOY_DIR := deploy
 
-.PHONY: help deploy deploy-bash check ssl push service
+.PHONY: help deploy deploy-bash check ssl push service metrics
 
 help:
 	@echo "The Door — Deployment Commands"
@@ -23,6 +23,8 @@ help:
 	@echo "  make check        Check deployment status"
 	@echo "  make ssl          Setup SSL on VPS"
 	@echo "  make service      Install/restart hermes-gateway service"
+	@echo "  make metrics      View crisis metrics summary"
+	@echo "  make metrics-json Export crisis metrics as JSON"
 	@echo ""
 
 deploy:
@@ -47,11 +49,8 @@ ssl:
 service:
 	ssh root@$(VPS) "cd /opt/the-door && bash deploy/deploy.sh --service"
 
-# Crisis metrics
-.PHONY: metrics metrics-json
-
-metrics:  ## Show crisis metrics summary (last 7 days)
+metrics:
 	python3 -m crisis.metrics --summary
 
-metrics-json:  ## Export crisis metrics as JSON
+metrics-json:
 	python3 -m crisis.metrics --json

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 .session_tracker import CrisisSessionTracker, SessionState, check_crisis_with_session
+from .metrics import CrisisMetrics, AggregateMetrics
 
 __all__ = [
     "detect_crisis",
@@ -19,4 +21,9 @@ __all__ = [
     "format_result",
     "format_gateway_response",
     "get_urgency_emoji",
+    "CrisisSessionTracker",
+    "SessionState",
+    "check_crisis_with_session",
+    "CrisisMetrics",
+    "AggregateMetrics",
 ]

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/metrics.py

@@ -1,199 +1,244 @@
-"""Crisis metrics — aggregate detection data for operators.
+"""
+crisis/metrics.py — Aggregate crisis detection metrics.
 
-Tracks crisis detection events and provides summary reports.
+Tracks session-level crisis data for aggregate reporting.
+Privacy-first: stores only aggregate counts, never user content.
 
 Usage:
-    python3 -m crisis.metrics --summary    # weekly report
-    python3 -m crisis.metrics --json       # raw JSON export
-    python3 -m crisis.metrics --last 7d    # last 7 days
+    from crisis.metrics import CrisisMetrics
+    
+    metrics = CrisisMetrics()
+    metrics.record_session(tracker.state)
+    summary = metrics.get_summary()
 """
 
-from __future__ import annotations
-
 import json
 import os
-import sys
 import time
-from collections import Counter
-from dataclasses import dataclass, asdict
+from dataclasses import dataclass, field, asdict
+from datetime import datetime, timedelta
 from pathlib import Path
-from typing import Any, Dict, List, Optional
+from typing import Dict, List, Optional
 
-# Data directory for metrics storage
-_DATA_DIR = Path(os.getenv("CRISIS_DATA_DIR", str(Path.home() / ".the-door")))
-_METRICS_FILE = _DATA_DIR / "crisis-metrics.jsonl"
+METRICS_DIR = Path.home() / ".the-door" / "metrics"
 
 
 @dataclass
-class CrisisEvent:
-    """A single crisis detection event."""
+class SessionMetrics:
+    """Metrics from a single crisis session."""
     timestamp: float
-    level: str           # NONE, LOW, MODERATE, HIGH, CRITICAL
-    indicators: list
-    session_id: str = ""
-    source: str = ""     # "chat", "gateway", "cli"
+    current_level: str
+    peak_level: str
+    message_count: int
+    was_escalating: bool
+    was_deescalating: bool
+    escalation_rate: float
+    triggered_overlay: bool = False
+    showed_988: bool = False
 
 
 @dataclass
-class MetricsSummary:
-    """Aggregated metrics summary."""
-    period_days: int
-    total_events: int
-    by_level: Dict[str, int]
-    top_indicators: List[tuple]
-    sessions_affected: int
-    avg_daily: float
-    peak_day: str
-    peak_count: int
-    generated_at: str
+class AggregateMetrics:
+    """Aggregate metrics across sessions."""
+    total_sessions: int = 0
+    total_messages: int = 0
+    
+    # Level distribution
+    level_counts: Dict[str, int] = field(default_factory=lambda: {
+        "NONE": 0, "LOW": 0, "MEDIUM": 0, "HIGH": 0, "CRITICAL": 0
+    })
+    
+    # Escalation tracking
+    escalating_sessions: int = 0
+    deescalating_sessions: int = 0
+    
+    # Safety interventions
+    overlay_triggers: int = 0
+    ninety_eight_show: int = 0
+    
+    # Time window
+    period_start: Optional[float] = None
+    period_end: Optional[float] = None
 
 
-def log_event(event: CrisisEvent) -> None:
-    """Log a crisis event to the metrics file."""
-    _DATA_DIR.mkdir(parents=True, exist_ok=True)
-    with open(_METRICS_FILE, "a") as f:
-        f.write(json.dumps(asdict(event)) + "\n")
-
-
-def load_events(days: int = 7) -> List[CrisisEvent]:
-    """Load crisis events from the last N days."""
-    if not _METRICS_FILE.exists():
-        return []
-
-    cutoff = time.time() - (days * 86400)
-    events = []
-
-    try:
-        with open(_METRICS_FILE) as f:
+class CrisisMetrics:
+    """
+    Aggregate crisis metrics with local JSON persistence.
+    
+    Privacy-first: stores only aggregate counts per day.
+    Never stores user messages, content, or identifying info.
+    """
+    
+    def __init__(self, metrics_dir: Optional[Path] = None):
+        self.metrics_dir = metrics_dir or METRICS_DIR
+        self.metrics_dir.mkdir(parents=True, exist_ok=True)
+        self._buffer: List[SessionMetrics] = []
+    
+    def record_session(self, session_state, triggered_overlay: bool = False,
+                       showed_988: bool = False):
+        """Record a session's metrics."""
+        from .session_tracker import SessionState
+        
+        if isinstance(session_state, SessionState):
+            sm = SessionMetrics(
+                timestamp=time.time(),
+                current_level=session_state.current_level,
+                peak_level=session_state.peak_level,
+                message_count=session_state.message_count,
+                was_escalating=session_state.is_escalating,
+                was_deescalating=session_state.is_deescalating,
+                escalation_rate=session_state.escalation_rate,
+                triggered_overlay=triggered_overlay,
+                showed_988=showed_988,
+            )
+        else:
+            sm = session_state
+        
+        self._buffer.append(sm)
+        self._flush()
+    
+    def _flush(self):
+        """Write buffered sessions to daily file."""
+        if not self._buffer:
+            return
+        
+        today = datetime.utcnow().strftime("%Y-%m-%d")
+        filepath = self.metrics_dir / f"{today}.jsonl"
+        
+        with open(filepath, 'a') as f:
+            for sm in self._buffer:
+                f.write(json.dumps(asdict(sm)) + '\n')
+        
+        self._buffer.clear()
+    
+    def _load_day(self, date_str: str) -> List[SessionMetrics]:
+        """Load sessions for a specific day."""
+        filepath = self.metrics_dir / f"{date_str}.jsonl"
+        if not filepath.exists():
+            return []
+        
+        sessions = []
+        with open(filepath) as f:
             for line in f:
-                line = line.strip()
-                if not line:
-                    continue
-                data = json.loads(line)
-                if data.get("timestamp", 0) >= cutoff:
-                    events.append(CrisisEvent(**data))
-    except (json.JSONDecodeError, KeyError):
-        pass
-
-    return events
-
-
-def compute_summary(days: int = 7) -> MetricsSummary:
-    """Compute metrics summary for the given period."""
-    events = load_events(days)
-    now = time.time()
-
-    # By level
-    by_level = Counter(e.level for e in events)
-
-    # Top indicators
-    indicator_counts = Counter()
-    for e in events:
-        for ind in e.indicators:
-            indicator_counts[ind] += 1
-    top_indicators = indicator_counts.most_common(10)
-
-    # Sessions
-    sessions = set(e.session_id for e in events if e.session_id)
-
-    # Peak day
-    from collections import defaultdict
-    daily = defaultdict(int)
-    for e in events:
-        day = time.strftime("%Y-%m-%d", time.localtime(e.timestamp))
-        daily[day] += 1
-    peak_day = max(daily, key=daily.get) if daily else "N/A"
-    peak_count = daily.get(peak_day, 0)
-
-    return MetricsSummary(
-        period_days=days,
-        total_events=len(events),
-        by_level=dict(by_level),
-        top_indicators=top_indicators,
-        sessions_affected=len(sessions),
-        avg_daily=round(len(events) / max(days, 1), 1),
-        peak_day=peak_day,
-        peak_count=peak_count,
-        generated_at=time.strftime("%Y-%m-%d %H:%M:%S"),
-    )
-
-
-def format_summary(summary: MetricsSummary) -> str:
-    """Format metrics summary as human-readable report."""
-    lines = [
-        "Crisis Metrics Summary",
-        "=" * 40,
-        f"Period:       Last {summary.period_days} days",
-        f"Generated:    {summary.generated_at}",
-        "",
-        f"Total events: {summary.total_events}",
-        f"Daily avg:    {summary.avg_daily}",
-        f"Sessions:     {summary.sessions_affected}",
-        f"Peak day:     {summary.peak_day} ({summary.peak_count} events)",
-        "",
-    ]
-
-    if summary.by_level:
-        lines.append("By severity:")
-        for level in ["CRITICAL", "HIGH", "MODERATE", "LOW", "NONE"]:
-            count = summary.by_level.get(level, 0)
-            if count > 0:
-                bar = "█" * min(count, 30)
-                lines.append(f"  {level:10s} {count:4d} {bar}")
-        lines.append("")
-
-    if summary.top_indicators:
-        lines.append("Top indicators:")
-        for indicator, count in summary.top_indicators[:5]:
-            lines.append(f"  {indicator}: {count}")
-        lines.append("")
-
-    if summary.total_events == 0:
-        lines.append("No crisis events in this period.")
-
-    return "\n".join(lines)
+                if line.strip():
+                    data = json.loads(line)
+                    sessions.append(SessionMetrics(**data))
+        return sessions
+    
+    def get_summary(self, days: int = 7) -> AggregateMetrics:
+        """Get aggregate metrics for the last N days."""
+        agg = AggregateMetrics()
+        
+        now = datetime.utcnow()
+        for i in range(days):
+            date = (now - timedelta(days=i)).strftime("%Y-%m-%d")
+            sessions = self._load_day(date)
+            
+            for sm in sessions:
+                agg.total_sessions += 1
+                agg.total_messages += sm.message_count
+                
+                # Level counts (use peak level)
+                level = sm.peak_level
+                agg.level_counts[level] = agg.level_counts.get(level, 0) + 1
+                
+                if sm.was_escalating:
+                    agg.escalating_sessions += 1
+                if sm.was_deescalating:
+                    agg.deescalating_sessions += 1
+                if sm.triggered_overlay:
+                    agg.overlay_triggers += 1
+                if sm.showed_988:
+                    agg.ninety_eight_show += 1
+                
+                # Time window
+                if agg.period_start is None or sm.timestamp < agg.period_start:
+                    agg.period_start = sm.timestamp
+                if agg.period_end is None or sm.timestamp > agg.period_end:
+                    agg.period_end = sm.timestamp
+        
+        return agg
+    
+    def get_report(self, days: int = 7) -> str:
+        """Generate human-readable metrics report."""
+        agg = self.get_summary(days)
+        
+        lines = []
+        lines.append("=" * 50)
+        lines.append("  CRISIS METRICS REPORT")
+        lines.append(f"  Last {days} days")
+        if agg.period_start:
+            start = datetime.fromtimestamp(agg.period_start).strftime("%Y-%m-%d %H:%M")
+            lines.append(f"  Period: {start} → now")
+        lines.append("=" * 50)
+        
+        lines.append(f"\n  Sessions:           {agg.total_sessions}")
+        lines.append(f"  Messages tracked:   {agg.total_messages}")
+        
+        lines.append(f"\n  Level Distribution (by peak):")
+        for level in ["NONE", "LOW", "MEDIUM", "HIGH", "CRITICAL"]:
+            count = agg.level_counts.get(level, 0)
+            pct = (count / agg.total_sessions * 100) if agg.total_sessions > 0 else 0
+            bar = "█" * int(pct / 5)
+            lines.append(f"    {level:<10} {count:>5} ({pct:>5.1f}%) {bar}")
+        
+        lines.append(f"\n  Escalations:        {agg.escalating_sessions}")
+        lines.append(f"  De-escalations:     {agg.deescalating_sessions}")
+        lines.append(f"  Overlay triggers:   {agg.overlay_triggers}")
+        lines.append(f"  988 shown:          {agg.ninety_eight_show}")
+        
+        if agg.total_sessions > 0:
+            escalation_rate = agg.escalating_sessions / agg.total_sessions * 100
+            lines.append(f"\n  Escalation rate:    {escalation_rate:.1f}%")
+        
+        lines.append("=" * 50)
+        
+        return "\n".join(lines)
+    
+    def get_json(self, days: int = 7) -> str:
+        """Export metrics as JSON."""
+        agg = self.get_summary(days)
+        return json.dumps(asdict(agg), indent=2)
 
 
 def main():
+    """CLI entry point for crisis metrics."""
     import argparse
-    parser = argparse.ArgumentParser(description="Crisis metrics summary")
-    parser.add_argument("--summary", action="store_true", help="Print summary report")
-    parser.add_argument("--json", action="store_true", dest="as_json", help="Output JSON")
-    parser.add_argument("--last", default="7d", help="Time period (e.g., 7d, 30d)")
-    parser.add_argument("--log", nargs=2, metavar=("LEVEL", "INDICATOR"), help="Log a test event")
+    
+    parser = argparse.ArgumentParser(description="Crisis Detection Metrics")
+    parser.add_argument("--summary", action="store_true", help="Show summary report")
+    parser.add_argument("--json", action="store_true", help="JSON export")
+    parser.add_argument("--days", type=int, default=7, help="Days to include")
+    parser.add_argument("--demo", action="store_true", help="Generate demo data")
     args = parser.parse_args()
-
-    # Parse period
-    period_str = args.last.rstrip("d")
-    try:
-        days = int(period_str)
-    except ValueError:
-        days = 7
-
-    # Log mode
-    if args.log:
-        level, indicator = args.log
-        event = CrisisEvent(
-            timestamp=time.time(),
-            level=level.upper(),
-            indicators=[indicator],
-            session_id="cli-test",
-            source="cli",
-        )
-        log_event(event)
-        print(f"Logged: {level.upper()} / {indicator}")
-        return 0
-
-    # Compute summary
-    summary = compute_summary(days)
-
-    if args.as_json:
-        print(json.dumps(asdict(summary), indent=2))
+    
+    metrics = CrisisMetrics()
+    
+    if args.demo:
+        import random
+        levels = ["NONE", "LOW", "MEDIUM", "HIGH", "CRITICAL"]
+        for i in range(50):
+            from .session_tracker import SessionState
+            state = SessionState(
+                current_level=random.choice(levels),
+                peak_level=random.choice(levels),
+                message_count=random.randint(1, 20),
+                is_escalating=random.random() > 0.7,
+                is_deescalating=random.random() > 0.8,
+                escalation_rate=random.random(),
+            )
+            metrics.record_session(
+                state,
+                triggered_overlay=random.random() > 0.8,
+                showed_988=random.random() > 0.7,
+            )
+        print("Generated 50 demo sessions.")
+    
+    if args.json:
+        print(metrics.get_json(args.days))
     else:
-        print(format_summary(summary))
-
-    return 0
+        print(metrics.get_report(args.days))
 
 
 if __name__ == "__main__":
-    sys.exit(main())
+    main()

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

@@ -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_metrics.py

@@ -0,0 +1,118 @@
+"""
+Tests for crisis/metrics.py — Aggregate crisis metrics.
+"""
+
+import json
+import os
+import shutil
+import tempfile
+import unittest
+from pathlib import Path
+
+import sys
+sys.path.insert(0, str(Path(__file__).parent.parent))
+
+from crisis.metrics import CrisisMetrics, SessionMetrics, AggregateMetrics
+
+
+class TestCrisisMetrics(unittest.TestCase):
+    def setUp(self):
+        self.tmpdir = tempfile.mkdtemp()
+        self.metrics = CrisisMetrics(Path(self.tmpdir))
+
+    def tearDown(self):
+        shutil.rmtree(self.tmpdir)
+
+    def test_record_session_creates_file(self):
+        sm = SessionMetrics(
+            timestamp=1700000000,
+            current_level="LOW",
+            peak_level="MEDIUM",
+            message_count=5,
+            was_escalating=True,
+            was_deescalating=False,
+            escalation_rate=0.5,
+        )
+        self.metrics.record_session(sm)
+
+        files = list(Path(self.tmpdir).glob("*.jsonl"))
+        self.assertEqual(len(files), 1)
+
+    def test_record_session_writes_jsonl(self):
+        sm = SessionMetrics(
+            timestamp=1700000000,
+            current_level="HIGH",
+            peak_level="CRITICAL",
+            message_count=10,
+            was_escalating=True,
+            was_deescalating=False,
+            escalation_rate=1.0,
+            triggered_overlay=True,
+            showed_988=True,
+        )
+        self.metrics.record_session(sm)
+
+        files = list(Path(self.tmpdir).glob("*.jsonl"))
+        with open(files[0]) as f:
+            data = json.loads(f.readline())
+        self.assertEqual(data['peak_level'], 'CRITICAL')
+        self.assertTrue(data['triggered_overlay'])
+
+    def test_get_summary_empty(self):
+        agg = self.metrics.get_summary(days=7)
+        self.assertEqual(agg.total_sessions, 0)
+        self.assertEqual(agg.total_messages, 0)
+
+    def test_get_summary_with_data(self):
+        for level in ["LOW", "MEDIUM", "HIGH"]:
+            sm = SessionMetrics(
+                timestamp=1700000000,
+                current_level=level,
+                peak_level=level,
+                message_count=3,
+                was_escalating=level != "LOW",
+                was_deescalating=False,
+                escalation_rate=0.5,
+            )
+            self.metrics.record_session(sm)
+
+        agg = self.metrics.get_summary(days=1)
+        self.assertEqual(agg.total_sessions, 3)
+        self.assertEqual(agg.total_messages, 9)
+        self.assertEqual(agg.escalating_sessions, 2)
+
+    def test_get_report_returns_string(self):
+        sm = SessionMetrics(
+            timestamp=1700000000,
+            current_level="LOW",
+            peak_level="LOW",
+            message_count=5,
+            was_escalating=False,
+            was_deescalating=False,
+            escalation_rate=0.0,
+        )
+        self.metrics.record_session(sm)
+
+        report = self.metrics.get_report(days=1)
+        self.assertIn("CRISIS METRICS REPORT", report)
+        self.assertIn("Sessions:", report)
+
+    def test_get_json_returns_valid(self):
+        sm = SessionMetrics(
+            timestamp=1700000000,
+            current_level="MEDIUM",
+            peak_level="MEDIUM",
+            message_count=3,
+            was_escalating=False,
+            was_deescalating=False,
+            escalation_rate=0.0,
+        )
+        self.metrics.record_session(sm)
+
+        json_str = self.metrics.get_json(days=1)
+        data = json.loads(json_str)
+        self.assertEqual(data['total_sessions'], 1)
+
+
+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()