feat: add crisis ab testing for #101

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

@@ -6,8 +6,9 @@ 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, get_metrics_summary, get_metrics_report
-from .metrics import record_detection, record_continuation, get_metrics
+from .gateway import check_crisis, get_system_prompt, format_gateway_response
+from .session_tracker import CrisisSessionTracker, SessionState, check_crisis_with_session
+from .ab_testing import ABTestCrisisDetector, VariantRecord
 
 __all__ = [
     "detect_crisis",
@@ -20,9 +21,9 @@ __all__ = [
     "format_result",
     "format_gateway_response",
     "get_urgency_emoji",
-    "get_metrics_summary",
-    "get_metrics_report",
-    "record_detection",
-    "record_continuation",
-    "get_metrics",
+    "CrisisSessionTracker",
+    "SessionState",
+    "check_crisis_with_session",
+    "ABTestCrisisDetector",
+    "VariantRecord",
 ]

crisis/ab_testing.py

@@ -0,0 +1,112 @@
+"""A/B test framework for crisis detection in the-door."""
+
+from __future__ import annotations
+
+import os
+import random
+import time
+from dataclasses import dataclass
+from typing import Callable, Dict, List, Optional, Tuple
+
+from .detect import CrisisDetectionResult
+
+
+def _get_variant_override() -> Optional[str]:
+    """Return env override for deterministic testing/debugging."""
+    value = os.environ.get("CRISIS_AB_VARIANT", "").strip().upper()
+    if value in {"A", "B"}:
+        return value
+    return None
+
+
+@dataclass
+class VariantRecord:
+    """Single crisis detection event record with no user text or PII."""
+
+    variant: str
+    level: str
+    latency_ms: float
+    indicator_count: int
+    false_positive: Optional[bool] = None
+
+
+class ABTestCrisisDetector:
+    """Route crisis detection between two variants and collect comparison stats."""
+
+    def __init__(
+        self,
+        variant_a: Callable[[str], CrisisDetectionResult],
+        variant_b: Callable[[str], CrisisDetectionResult],
+        split: float = 0.5,
+    ):
+        self.variant_a = variant_a
+        self.variant_b = variant_b
+        self.split = max(0.0, min(1.0, float(split)))
+        self.records: List[VariantRecord] = []
+
+    def _select_variant(self) -> str:
+        override = _get_variant_override()
+        if override:
+            return override
+        return "A" if random.random() < self.split else "B"
+
+    def detect(self, text: str) -> Tuple[CrisisDetectionResult, str, int]:
+        variant = self._select_variant()
+        detector = self.variant_a if variant == "A" else self.variant_b
+
+        start = time.perf_counter()
+        result = detector(text)
+        latency_ms = (time.perf_counter() - start) * 1000.0
+
+        record = VariantRecord(
+            variant=variant,
+            level=result.level,
+            latency_ms=latency_ms,
+            indicator_count=len(result.indicators),
+        )
+        self.records.append(record)
+        return result, variant, len(self.records) - 1
+
+    def record_outcome(self, record_id: int, *, false_positive: bool) -> None:
+        if record_id < 0 or record_id >= len(self.records):
+            raise IndexError(f"Unknown record id: {record_id}")
+        self.records[record_id].false_positive = bool(false_positive)
+
+    def get_stats(self) -> Dict[str, dict]:
+        stats: Dict[str, dict] = {}
+        for variant in ("A", "B"):
+            records = [record for record in self.records if record.variant == variant]
+            if not records:
+                stats[variant] = {
+                    "count": 0,
+                    "reviewed_count": 0,
+                    "false_positive_rate": None,
+                }
+                continue
+
+            levels: Dict[str, int] = {}
+            for record in records:
+                levels[record.level] = levels.get(record.level, 0) + 1
+
+            reviewed = [record for record in records if record.false_positive is not None]
+            false_positive_rate = None
+            if reviewed:
+                false_positive_rate = round(
+                    sum(1 for record in reviewed if record.false_positive) / len(reviewed),
+                    4,
+                )
+
+            stats[variant] = {
+                "count": len(records),
+                "avg_latency_ms": round(sum(record.latency_ms for record in records) / len(records), 4),
+                "max_latency_ms": round(max(record.latency_ms for record in records), 4),
+                "min_latency_ms": round(min(record.latency_ms for record in records), 4),
+                "avg_indicator_count": round(sum(record.indicator_count for record in records) / len(records), 4),
+                "levels": levels,
+                "reviewed_count": len(reviewed),
+                "false_positive_rate": false_positive_rate,
+            }
+        return stats
+
+    def reset(self) -> None:
+        self.records.clear()

crisis/gateway.py

@@ -22,7 +22,7 @@ from .response import (
     get_system_prompt_modifier,
     CrisisResponse,
 )
-from .metrics import record_detection, get_summary, print_summary, record_continuation
+from .session_tracker import CrisisSessionTracker
 
 
 def check_crisis(text: str) -> dict:
@@ -35,9 +35,6 @@ def check_crisis(text: str) -> dict:
     detection = detect_crisis(text)
     response = generate_response(detection)
 
-    # Record metrics (privacy-preserving — no message content stored)
-    record_detection(detection.level, detection.indicators)
-
     return {
         "level": detection.level,
         "score": detection.score,
@@ -97,21 +94,6 @@ def format_gateway_response(text: str, pretty: bool = True) -> str:
     return json.dumps(result)
 
 
-# ── Metrics endpoint ─────────────────────────────────────────────
-
-def get_metrics_summary() -> dict:
-    """
-    Return crisis detection metrics summary.
-    Privacy-preserving: no PII, no message content, just counts.
-    """
-    return get_summary()
-
-
-def get_metrics_report() -> str:
-    """Return formatted weekly metrics report for stdout/logs."""
-    return print_summary()
-
-
 # ── Quick test interface ────────────────────────────────────────
 
 def _interactive():

crisis/metrics.py

@@ -1,270 +0,0 @@
-"""
-Crisis Detection Metrics for the-door.
-
-Privacy-preserving analytics layer. Tracks:
-- Detection counts per level (CRITICAL, HIGH, MEDIUM, LOW)
-- Keyword firing frequency (pattern hashes, not raw text)
-- Time-based distribution (hourly buckets)
-- Post-intervention behavior
-
-NO PII is stored — no message content, no user identifiers, no timestamps
-finer than hourly granularity.
-
-Storage: JSON file at crisis_metrics.json (configurable path).
-"""
-
-import json
-import os
-import hashlib
-import time
-from datetime import datetime, timezone
-from dataclasses import dataclass, field, asdict
-from typing import Dict, List, Optional
-
-
-# ── Pattern hash helper ───────────────────────────────────────────
-
-def _hash_pattern(pattern: str) -> str:
-    """Hash a regex pattern to avoid storing raw crisis text."""
-    return hashlib.sha256(pattern.encode()).hexdigest()[:12]
-
-
-# ── Hour bucket helper ────────────────────────────────────────────
-
-def _current_hour_key() -> str:
-    """Return current UTC hour as 'YYYY-MM-DDTHH'."""
-    return datetime.now(timezone.utc).strftime("%Y-%m-%dT%H")
-
-
-def _current_day_key() -> str:
-    """Return current UTC day as 'YYYY-MM-DD'."""
-    return datetime.now(timezone.utc).strftime("%Y-%m-%d")
-
-
-# ── Metrics store ─────────────────────────────────────────────────
-
-@dataclass
-class CrisisMetrics:
-    """In-memory metrics accumulator with JSON persistence."""
-
-    # Total detections per level
-    detections_by_level: Dict[str, int] = field(default_factory=lambda: {
-        "CRITICAL": 0, "HIGH": 0, "MEDIUM": 0, "LOW": 0, "NONE": 0
-    })
-
-    # Pattern hash -> count (tracks which indicators fire most)
-    keyword_frequency: Dict[str, int] = field(default_factory=dict)
-
-    # Hourly detection counts: "YYYY-MM-DDTHH" -> total
-    hourly_counts: Dict[str, int] = field(default_factory=dict)
-
-    # Daily detection counts: "YYYY-MM-DD" -> {level: count}
-    daily_counts: Dict[str, Dict[str, int]] = field(default_factory=dict)
-
-    # Total messages scanned (for false-positive estimation)
-    total_scanned: int = 0
-
-    # Total detections (non-NONE)
-    total_detections: int = 0
-
-    # Post-intervention tracking (session-scoped, reset on restart)
-    interventions: int = 0
-    continued_after_intervention: int = 0
-
-    def record(self, level: str, indicators: List[str]) -> None:
-        """Record a single detection event."""
-        self.total_scanned += 1
-
-        # Level counts
-        self.detections_by_level[level] = self.detections_by_level.get(level, 0) + 1
-
-        if level != "NONE":
-            self.total_detections += 1
-
-            # Hourly bucket
-            hour = _current_hour_key()
-            self.hourly_counts[hour] = self.hourly_counts.get(hour, 0) + 1
-
-            # Daily bucket
-            day = _current_day_key()
-            if day not in self.daily_counts:
-                self.daily_counts[day] = {}
-            daily = self.daily_counts[day]
-            daily[level] = daily.get(level, 0) + 1
-
-            # Keyword frequency (hash patterns)
-            for pattern in indicators:
-                h = _hash_pattern(pattern)
-                self.keyword_frequency[h] = self.keyword_frequency.get(h, 0) + 1
-
-        # Intervention tracking
-        if level in ("CRITICAL", "HIGH"):
-            self.interventions += 1
-
-    def record_continuation(self) -> None:
-        """Call when a user continues conversation after crisis intervention."""
-        self.continued_after_intervention += 1
-
-    def to_dict(self) -> dict:
-        """Serialize to JSON-safe dict."""
-        return asdict(self)
-
-    @classmethod
-    def from_dict(cls, data: dict) -> "CrisisMetrics":
-        """Deserialize from dict."""
-        return cls(
-            detections_by_level=data.get("detections_by_level", {}),
-            keyword_frequency=data.get("keyword_frequency", {}),
-            hourly_counts=data.get("hourly_counts", {}),
-            daily_counts=data.get("daily_counts", {}),
-            total_scanned=data.get("total_scanned", 0),
-            total_detections=data.get("total_detections", 0),
-            interventions=data.get("interventions", 0),
-            continued_after_intervention=data.get("continued_after_intervention", 0),
-        )
-
-    # ── Derived metrics ───────────────────────────────────────────
-
-    @property
-    def false_positive_estimate(self) -> float:
-        """
-        Estimate false positive rate.
-        Heuristic: users who continue chatting after HIGH/CRITICAL intervention
-        were likely not in true crisis. Returns 0.0 if no interventions.
-        """
-        if self.interventions == 0:
-            return 0.0
-        return self.continued_after_intervention / self.interventions
-
-    @property
-    def detection_rate(self) -> float:
-        """Fraction of scanned messages that triggered any detection."""
-        if self.total_scanned == 0:
-            return 0.0
-        return self.total_detections / self.total_scanned
-
-    def top_keywords(self, n: int = 10) -> List[tuple]:
-        """Return top N most-fired pattern hashes with counts."""
-        sorted_kw = sorted(self.keyword_frequency.items(), key=lambda x: -x[1])
-        return sorted_kw[:n]
-
-    def weekly_summary(self) -> dict:
-        """Generate a 7-day summary from daily_counts."""
-        from datetime import timedelta
-
-        today = datetime.now(timezone.utc).date()
-        summary = {
-            "period_start": (today - timedelta(days=6)).isoformat(),
-            "period_end": today.isoformat(),
-            "total_detections": 0,
-            "by_level": {"CRITICAL": 0, "HIGH": 0, "MEDIUM": 0, "LOW": 0},
-            "daily_totals": {},
-        }
-
-        for i in range(7):
-            day = (today - timedelta(days=6 - i)).isoformat()
-            day_data = self.daily_counts.get(day, {})
-            day_total = sum(day_data.values())
-            summary["daily_totals"][day] = day_total
-            summary["total_detections"] += day_total
-            for level in ("CRITICAL", "HIGH", "MEDIUM", "LOW"):
-                summary["by_level"][level] += day_data.get(level, 0)
-
-        summary["false_positive_estimate"] = self.false_positive_estimate
-        summary["detection_rate"] = self.detection_rate
-        summary["top_keywords"] = self.top_keywords(5)
-        return summary
-
-    def format_summary(self) -> str:
-        """Human-readable weekly summary for stdout/log."""
-        s = self.weekly_summary()
-        lines = [
-            "=== Crisis Detection Weekly Summary ===",
-            f"Period: {s['period_start']} to {s['period_end']}",
-            f"Total detections: {s['total_detections']}",
-            "",
-            "By level:",
-        ]
-        for level in ("CRITICAL", "HIGH", "MEDIUM", "LOW"):
-            count = s["by_level"][level]
-            bar = "#" * count
-            lines.append(f"  {level:10s} {count:4d} {bar}")
-
-        lines.append("")
-        lines.append("Daily trend:")
-        for day, count in s["daily_totals"].items():
-            bar = "#" * count
-            lines.append(f"  {day}  {count:4d} {bar}")
-
-        lines.append("")
-        lines.append(f"Detection rate: {s['detection_rate']:.1%}")
-        lines.append(f"False positive estimate: {s['false_positive_estimate']:.1%}")
-        lines.append("")
-        lines.append("Top indicators (hashed):")
-        for h, count in s["top_keywords"]:
-            lines.append(f"  {h}  {count:4d}x")
-
-        return "\n".join(lines)
-
-
-# ── Persistence ───────────────────────────────────────────────────
-
-_DEFAULT_PATH = os.path.join(os.path.dirname(__file__), "crisis_metrics.json")
-
-
-def load_metrics(path: str = _DEFAULT_PATH) -> CrisisMetrics:
-    """Load metrics from JSON file, or return empty metrics if missing."""
-    try:
-        with open(path) as f:
-            data = json.load(f)
-        return CrisisMetrics.from_dict(data)
-    except (FileNotFoundError, json.JSONDecodeError):
-        return CrisisMetrics()
-
-
-def save_metrics(metrics: CrisisMetrics, path: str = _DEFAULT_PATH) -> None:
-    """Persist metrics to JSON file."""
-    with open(path, "w") as f:
-        json.dump(metrics.to_dict(), f, indent=2)
-
-
-# ── Global singleton ──────────────────────────────────────────────
-
-_metrics: Optional[CrisisMetrics] = None
-_metrics_path: str = _DEFAULT_PATH
-
-
-def get_metrics(path: str = None) -> CrisisMetrics:
-    """Get or initialize the global metrics singleton."""
-    global _metrics, _metrics_path
-    if path:
-        _metrics_path = path
-    if _metrics is None:
-        _metrics = load_metrics(_metrics_path)
-    return _metrics
-
-
-def record_detection(level: str, indicators: List[str], path: str = None) -> None:
-    """Record a crisis detection event and persist."""
-    m = get_metrics(path)
-    m.record(level, indicators)
-    save_metrics(m, _metrics_path)
-
-
-def record_continuation(path: str = None) -> None:
-    """Record that a user continued after crisis intervention."""
-    m = get_metrics(path)
-    m.record_continuation()
-    save_metrics(m, _metrics_path)
-
-
-def get_summary(path: str = None) -> dict:
-    """Get the weekly summary as a dict."""
-    m = get_metrics(path)
-    return m.weekly_summary()
-
-
-def print_summary(path: str = None) -> str:
-    """Get and format the weekly summary."""
-    m = get_metrics(path)
-    return m.format_summary()

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_ab_testing.py

@@ -0,0 +1,138 @@
+"""Tests for crisis.ab_testing — A/B test framework for crisis detection (#101)."""
+
+import os
+from unittest.mock import patch
+
+import pytest
+
+from crisis.ab_testing import ABTestCrisisDetector
+from crisis.detect import CrisisDetectionResult, detect_crisis
+
+
+@pytest.fixture(autouse=True)
+def clear_variant_override():
+    old = os.environ.pop("CRISIS_AB_VARIANT", None)
+    try:
+        yield
+    finally:
+        if old is not None:
+            os.environ["CRISIS_AB_VARIANT"] = old
+        else:
+            os.environ.pop("CRISIS_AB_VARIANT", None)
+
+
+def _make_variant(level: str, indicators=None):
+    indicators = indicators or [f"mock_{level.lower()}"]
+
+    def fn(text: str) -> CrisisDetectionResult:
+        return CrisisDetectionResult(level=level, indicators=list(indicators))
+
+    return fn
+
+
+def test_detect_returns_result_variant_and_logged_record():
+    detector = ABTestCrisisDetector(
+        variant_a=_make_variant("LOW"),
+        variant_b=_make_variant("HIGH"),
+    )
+
+    with patch.object(detector, "_select_variant", return_value="A"):
+        result, variant, record_id = detector.detect("test message")
+
+    assert isinstance(result, CrisisDetectionResult)
+    assert variant == "A"
+    assert record_id == 0
+    assert len(detector.records) == 1
+    assert detector.records[0].variant == "A"
+    assert detector.records[0].level == "LOW"
+
+
+def test_env_override_forces_variant_b():
+    os.environ["CRISIS_AB_VARIANT"] = "b"
+    detector = ABTestCrisisDetector(
+        variant_a=_make_variant("LOW"),
+        variant_b=_make_variant("HIGH"),
+    )
+
+    result, variant, _ = detector.detect("test")
+
+    assert variant == "B"
+    assert result.level == "HIGH"
+
+
+def test_get_stats_reports_latency_counts_and_level_breakdown():
+    detector = ABTestCrisisDetector(
+        variant_a=_make_variant("LOW"),
+        variant_b=_make_variant("CRITICAL"),
+    )
+
+    with patch.object(detector, "_select_variant", side_effect=["A", "A", "B"]):
+        detector.detect("first")
+        detector.detect("second")
+        detector.detect("third")
+
+    stats = detector.get_stats()
+    assert stats["A"]["count"] == 2
+    assert stats["B"]["count"] == 1
+    assert stats["A"]["levels"]["LOW"] == 2
+    assert stats["B"]["levels"]["CRITICAL"] == 1
+    assert "avg_latency_ms" in stats["A"]
+    assert "avg_indicator_count" in stats["B"]
+
+
+def test_false_positive_rate_is_computed_from_reviewed_outcomes():
+    detector = ABTestCrisisDetector(
+        variant_a=_make_variant("LOW"),
+        variant_b=_make_variant("HIGH"),
+    )
+
+    with patch.object(detector, "_select_variant", side_effect=["A", "A", "B"]):
+        _, _, a0 = detector.detect("first")
+        _, _, a1 = detector.detect("second")
+        _, _, b0 = detector.detect("third")
+
+    detector.record_outcome(a0, false_positive=True)
+    detector.record_outcome(a1, false_positive=False)
+    detector.record_outcome(b0, false_positive=False)
+
+    stats = detector.get_stats()
+    assert stats["A"]["reviewed_count"] == 2
+    assert stats["A"]["false_positive_rate"] == 0.5
+    assert stats["B"]["false_positive_rate"] == 0.0
+
+
+def test_record_outcome_rejects_unknown_record():
+    detector = ABTestCrisisDetector(
+        variant_a=_make_variant("LOW"),
+        variant_b=_make_variant("HIGH"),
+    )
+
+    with pytest.raises(IndexError):
+        detector.record_outcome(99, false_positive=True)
+
+
+def test_reset_clears_records_and_stats():
+    detector = ABTestCrisisDetector(
+        variant_a=_make_variant("LOW"),
+        variant_b=_make_variant("HIGH"),
+    )
+    detector.detect("test")
+    detector.reset()
+
+    assert detector.records == []
+    stats = detector.get_stats()
+    assert stats["A"]["count"] == 0
+    assert stats["B"]["count"] == 0
+
+
+def test_with_real_detector_integration():
+    detector = ABTestCrisisDetector(
+        variant_a=detect_crisis,
+        variant_b=detect_crisis,
+    )
+
+    result, variant, record_id = detector.detect("I want to kill myself")
+
+    assert result.level == "CRITICAL"
+    assert variant in ("A", "B")
+    assert record_id == 0

tests/test_crisis_metrics.py

@@ -1,263 +0,0 @@
-"""
-Tests for crisis detection metrics (issue #37).
-Verifies privacy-preserving analytics layer works correctly.
-"""
-
-import unittest
-import sys
-import os
-import json
-import tempfile
-
-sys.path.insert(0, os.path.dirname(os.path.dirname(os.path.abspath(__file__))))
-
-from crisis.metrics import (
-    CrisisMetrics,
-    load_metrics,
-    save_metrics,
-    _hash_pattern,
-)
-
-
-class TestCrisisMetrics(unittest.TestCase):
-    """Test the CrisisMetrics dataclass and persistence."""
-
-    def setUp(self):
-        self.metrics = CrisisMetrics()
-        self.tmpfile = tempfile.NamedTemporaryFile(suffix=".json", delete=False)
-        self.tmpfile.close()
-
-    def tearDown(self):
-        if os.path.exists(self.tmpfile.name):
-            os.unlink(self.tmpfile.name)
-
-    def test_record_none(self):
-        """Recording NONE should increment scanned but not detections."""
-        self.metrics.record("NONE", [])
-        self.assertEqual(self.metrics.total_scanned, 1)
-        self.assertEqual(self.metrics.total_detections, 0)
-        self.assertEqual(self.metrics.detections_by_level["NONE"], 1)
-
-    def test_record_critical(self):
-        """Recording CRITICAL should increment all relevant counters."""
-        pattern = r"\bkill\s*(my)?self\b"
-        self.metrics.record("CRITICAL", [pattern])
-        self.assertEqual(self.metrics.total_scanned, 1)
-        self.assertEqual(self.metrics.total_detections, 1)
-        self.assertEqual(self.metrics.detections_by_level["CRITICAL"], 1)
-        self.assertEqual(self.metrics.interventions, 1)
-
-        # Keyword should be hashed, not raw
-        h = _hash_pattern(pattern)
-        self.assertEqual(self.metrics.keyword_frequency[h], 1)
-
-    def test_record_high_is_intervention(self):
-        """HIGH detections should count as interventions."""
-        self.metrics.record("HIGH", [r"\bdespair\b"])
-        self.assertEqual(self.metrics.interventions, 1)
-
-    def test_record_medium_not_intervention(self):
-        """MEDIUM detections should NOT count as interventions."""
-        self.metrics.record("MEDIUM", [r"\bbroken\b", r"\bworthless\b"])
-        self.assertEqual(self.metrics.interventions, 0)
-        self.assertEqual(self.metrics.total_detections, 1)
-
-    def test_record_multiple(self):
-        """Multiple detections should accumulate."""
-        self.metrics.record("CRITICAL", [r"\bpattern1\b"])
-        self.metrics.record("HIGH", [r"\bpattern2\b"])
-        self.metrics.record("NONE", [])
-        self.metrics.record("LOW", [r"\bsad\b"])
-
-        self.assertEqual(self.metrics.total_scanned, 4)
-        self.assertEqual(self.metrics.total_detections, 3)
-        self.assertEqual(self.metrics.interventions, 2)
-
-    def test_continuation_tracking(self):
-        """Recording continuation should affect false positive estimate."""
-        self.metrics.record("CRITICAL", [r"\bpattern\b"])
-        self.assertEqual(self.metrics.false_positive_estimate, 0.0)
-
-        self.metrics.record_continuation()
-        self.assertEqual(self.metrics.continued_after_intervention, 1)
-        self.assertEqual(self.metrics.false_positive_estimate, 1.0)
-
-    def test_false_positive_estimate_zero_when_no_interventions(self):
-        """False positive rate should be 0.0 with no interventions."""
-        self.assertEqual(self.metrics.false_positive_estimate, 0.0)
-
-    def test_detection_rate(self):
-        """Detection rate should be detections/scanned."""
-        self.metrics.record("CRITICAL", [])
-        self.metrics.record("NONE", [])
-        self.metrics.record("NONE", [])
-        self.assertEqual(self.metrics.detection_rate, 1.0 / 3.0)
-
-    def test_detection_rate_zero_when_no_scans(self):
-        self.assertEqual(self.metrics.detection_rate, 0.0)
-
-    def test_hourly_bucket(self):
-        """Detections should be bucketed by hour."""
-        self.metrics.record("HIGH", [])
-        hour_key = self.metrics.hourly_counts
-        self.assertEqual(len(hour_key), 1)
-        for k, v in hour_key.items():
-            self.assertRegex(k, r"^\d{4}-\d{2}-\d{2}T\d{2}$")
-            self.assertEqual(v, 1)
-
-    def test_daily_bucket(self):
-        """Detections should be bucketed by day with level."""
-        self.metrics.record("CRITICAL", [])
-        self.metrics.record("HIGH", [])
-        self.assertEqual(len(self.metrics.daily_counts), 1)
-        for day, levels in self.metrics.daily_counts.items():
-            self.assertRegex(day, r"^\d{4}-\d{2}-\d{2}$")
-            self.assertEqual(levels["CRITICAL"], 1)
-            self.assertEqual(levels["HIGH"], 1)
-
-    def test_top_keywords(self):
-        """top_keywords should return most frequent hashed patterns."""
-        p1 = r"\bpattern_a\b"
-        p2 = r"\bpattern_b\b"
-        h1 = _hash_pattern(p1)
-        h2 = _hash_pattern(p2)
-
-        for _ in range(5):
-            self.metrics.record("HIGH", [p1])
-        for _ in range(2):
-            self.metrics.record("MEDIUM", [p2])
-
-        top = self.metrics.top_keywords(2)
-        self.assertEqual(len(top), 2)
-        self.assertEqual(top[0], (h1, 5))
-        self.assertEqual(top[1], (h2, 2))
-
-    def test_weekly_summary_structure(self):
-        """Weekly summary should have expected keys."""
-        self.metrics.record("CRITICAL", [])
-        summary = self.metrics.weekly_summary()
-
-        self.assertIn("period_start", summary)
-        self.assertIn("period_end", summary)
-        self.assertIn("total_detections", summary)
-        self.assertIn("by_level", summary)
-        self.assertIn("daily_totals", summary)
-        self.assertIn("false_positive_estimate", summary)
-        self.assertIn("detection_rate", summary)
-        self.assertIn("top_keywords", summary)
-
-        self.assertEqual(summary["total_detections"], 1)
-        self.assertEqual(summary["by_level"]["CRITICAL"], 1)
-
-    def test_format_summary_returns_string(self):
-        """format_summary should return a non-empty string."""
-        self.metrics.record("CRITICAL", [])
-        result = self.metrics.format_summary()
-        self.assertIsInstance(result, str)
-        self.assertIn("CRITICAL", result)
-
-    def test_persistence_round_trip(self):
-        """Metrics should survive save/load cycle."""
-        self.metrics.record("CRITICAL", [r"\btest\b"])
-        self.metrics.record("HIGH", [])
-        self.metrics.record_continuation()
-
-        save_metrics(self.metrics, self.tmpfile.name)
-        loaded = load_metrics(self.tmpfile.name)
-
-        self.assertEqual(loaded.total_scanned, 2)
-        self.assertEqual(loaded.total_detections, 2)
-        self.assertEqual(loaded.detections_by_level["CRITICAL"], 1)
-        self.assertEqual(loaded.detections_by_level["HIGH"], 1)
-        self.assertEqual(loaded.interventions, 2)
-        self.assertEqual(loaded.continued_after_intervention, 1)
-
-    def test_load_missing_file_returns_empty(self):
-        """Loading a nonexistent file should return empty metrics."""
-        m = load_metrics("/tmp/nonexistent_metrics_999.json")
-        self.assertEqual(m.total_scanned, 0)
-        self.assertEqual(m.total_detections, 0)
-
-    def test_load_corrupt_file_returns_empty(self):
-        """Loading a corrupt JSON file should return empty metrics."""
-        with open(self.tmpfile.name, "w") as f:
-            f.write("NOT JSON {{{")
-        m = load_metrics(self.tmpfile.name)
-        self.assertEqual(m.total_scanned, 0)
-
-    def test_hash_pattern_consistent(self):
-        """Same pattern should always produce same hash."""
-        h1 = _hash_pattern(r"\bkill\s*self\b")
-        h2 = _hash_pattern(r"\bkill\s*self\b")
-        self.assertEqual(h1, h2)
-
-    def test_hash_pattern_different(self):
-        """Different patterns should produce different hashes."""
-        h1 = _hash_pattern(r"\bpattern_a\b")
-        h2 = _hash_pattern(r"\bpattern_b\b")
-        self.assertNotEqual(h1, h2)
-
-
-class TestMetricsGatewayIntegration(unittest.TestCase):
-    """Test that metrics are recorded through the gateway."""
-
-    def test_check_crisis_records_metrics(self):
-        """check_crisis should record metrics automatically."""
-        import tempfile
-        from crisis.metrics import get_metrics, _metrics
-
-        # Use a temp metrics file
-        tmpfile = tempfile.NamedTemporaryFile(suffix=".json", delete=False)
-        tmpfile.close()
-
-        try:
-            # Reset global singleton
-            import crisis.metrics
-            crisis.metrics._metrics = None
-            crisis.metrics._metrics_path = tmpfile.name
-
-            from crisis.gateway import check_crisis
-
-            # A critical message
-            check_crisis("I want to kill myself")
-
-            m = get_metrics()
-            self.assertEqual(m.total_scanned, 1)
-            self.assertEqual(m.detections_by_level["CRITICAL"], 1)
-
-            # A safe message
-            check_crisis("I had a good day today")
-            self.assertEqual(m.total_scanned, 2)
-            self.assertEqual(m.detections_by_level["NONE"], 1)
-
-        finally:
-            # Reset singleton
-            crisis.metrics._metrics = None
-            if os.path.exists(tmpfile.name):
-                os.unlink(tmpfile.name)
-
-
-class TestNoPIIStorage(unittest.TestCase):
-    """Verify that no message content is stored in metrics."""
-
-    def test_no_text_in_serialized_metrics(self):
-        """Metrics to_dict should never contain message text."""
-        from crisis.metrics import CrisisMetrics
-
-        m = CrisisMetrics()
-        # Record with actual crisis text — only patterns should be hashed
-        m.record("CRITICAL", [r"\bkill\s*self\b"])
-
-        serialized = json.dumps(m.to_dict())
-
-        # These strings should NOT appear in serialized metrics
-        self.assertNotIn("kill myself", serialized)
-        self.assertNotIn("I want to", serialized)
-        self.assertNotIn("user", serialized.lower())
-
-        # Only hashed patterns should appear (short hex strings)
-        self.assertIn(_hash_pattern(r"\bkill\s*self\b"), serialized)
-
-
-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()