feat: add crisis analytics metrics (#37)

Fixes #59

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

crisis/__init__.py

@@ -7,8 +7,14 @@ 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 .tracker import log_escalation, get_escalations, mark_resolved, get_stats
-from .bridge import handle_crisis_api
+from .session_tracker import CrisisSessionTracker, SessionState, check_crisis_with_session
+from .metrics import (
+    build_metrics_event,
+    append_metrics_event,
+    load_metrics_events,
+    build_weekly_summary,
+    render_weekly_summary,
+)
 
 __all__ = [
     "detect_crisis",
@@ -21,9 +27,12 @@ __all__ = [
     "format_result",
     "format_gateway_response",
     "get_urgency_emoji",
-    "log_escalation",
-    "get_escalations",
-    "mark_resolved",
-    "get_stats",
-    "handle_crisis_api",
+    "CrisisSessionTracker",
+    "SessionState",
+    "check_crisis_with_session",
+    "build_metrics_event",
+    "append_metrics_event",
+    "load_metrics_events",
+    "build_weekly_summary",
+    "render_weekly_summary",
 ]

crisis/bridge.py

@@ -1,168 +0,0 @@
-#!/usr/bin/env python3
-"""
-Crisis Bridge — HTTP integration between the-door web and hermes-agent.
-
-Provides:
-- GET /api/crisis/escalations    — list recent escalation events
-- GET /api/crisis/stats           — aggregate statistics
-- POST /api/crisis/log            — log a new escalation (from hermes-agent)
-- POST /api/crisis/resolve/:id    — mark escalation as resolved
-
-Can be mounted as an ASGI/FastAPI sub-app or used standalone.
-Falls back to a simple HTTP server if no framework is available.
-"""
-
-import json
-import os
-from typing import Optional
-
-from .tracker import (
-    log_escalation,
-    get_escalations,
-    mark_resolved,
-    get_stats,
-)
-
-
-def handle_crisis_api(method: str, path: str, body: Optional[str] = None,
-                      hermes_home: Optional[str] = None) -> dict:
-    """
-    Handle a crisis API request. Returns dict with status, headers, body.
-
-    Args:
-        method: HTTP method (GET, POST)
-        path: Request path (e.g., "/api/crisis/escalations")
-        body: JSON request body (for POST)
-        hermes_home: Override HERMES_HOME path
-
-    Returns:
-        {"status": int, "headers": dict, "body": str}
-    """
-    # Normalize path
-    path = path.rstrip("/")
-
-    # GET /api/crisis/escalations
-    if method == "GET" and path == "/api/crisis/escalations":
-        params = _parse_query(path)
-        events = get_escalations(
-            limit=int(params.get("limit", 50)),
-            source=params.get("source"),
-            level=params.get("level"),
-            session_id=params.get("session_id"),
-            since=params.get("since"),
-            hermes_home=hermes_home,
-        )
-        return _json_response(200, {"events": events, "count": len(events)})
-
-    # GET /api/crisis/stats
-    if method == "GET" and path == "/api/crisis/stats":
-        stats = get_stats(hermes_home=hermes_home)
-        return _json_response(200, stats)
-
-    # POST /api/crisis/log
-    if method == "POST" and path == "/api/crisis/log":
-        if not body:
-            return _json_response(400, {"error": "Missing request body"})
-        try:
-            data = json.loads(body)
-        except json.JSONDecodeError:
-            return _json_response(400, {"error": "Invalid JSON"})
-
-        required = ["source", "session_id", "level", "indicators"]
-        missing = [f for f in required if f not in data]
-        if missing:
-            return _json_response(400, {"error": f"Missing fields: {missing}"})
-
-        event = log_escalation(
-            source=data["source"],
-            session_id=data["session_id"],
-            level=data["level"],
-            indicators=data.get("indicators", []),
-            score=data.get("score", 0.0),
-            action_taken=data.get("action_taken", ""),
-            hermes_home=hermes_home,
-        )
-        return _json_response(201, event)
-
-    # POST /api/crisis/resolve/:id
-    if method == "POST" and path.startswith("/api/crisis/resolve/"):
-        event_id = path.split("/")[-1]
-        if mark_resolved(event_id, hermes_home=hermes_home):
-            return _json_response(200, {"resolved": True, "id": event_id})
-        return _json_response(404, {"error": "Event not found"})
-
-    return _json_response(404, {"error": "Not found"})
-
-
-def _parse_query(path: str) -> dict:
-    """Extract query parameters from path."""
-    params = {}
-    if "?" in path:
-        query = path.split("?", 1)[1]
-        for pair in query.split("&"):
-            if "=" in pair:
-                k, v = pair.split("=", 1)
-                params[k] = v
-    return params
-
-
-def _json_response(status: int, body: dict) -> dict:
-    """Format a JSON response."""
-    return {
-        "status": status,
-        "headers": {"Content-Type": "application/json"},
-        "body": json.dumps(body, indent=2),
-    }
-
-
-# ---------------------------------------------------------------------------
-# Standalone server (for development / testing)
-# ---------------------------------------------------------------------------
-
-def run_standalone(host: str = "127.0.0.1", port: int = 8650,
-                   hermes_home: Optional[str] = None):
-    """Run a minimal HTTP server for the crisis API."""
-    try:
-        from http.server import HTTPServer, BaseHTTPRequestHandler
-    except ImportError:
-        print("http.server not available")
-        return
-
-    class Handler(BaseHTTPRequestHandler):
-        def do_GET(self):
-            result = handle_crisis_api("GET", self.path, hermes_home=hermes_home)
-            self._send(result)
-
-        def do_POST(self):
-            length = int(self.headers.get("Content-Length", 0))
-            body = self.rfile.read(length).decode() if length > 0 else None
-            result = handle_crisis_api("POST", self.path, body, hermes_home=hermes_home)
-            self._send(result)
-
-        def _send(self, result):
-            self.send_response(result["status"])
-            for k, v in result["headers"].items():
-                self.send_header(k, v)
-            self.end_headers()
-            self.wfile.write(result["body"].encode())
-
-        def log_message(self, format, *args):
-            pass  # Suppress default logging
-
-    server = HTTPServer((host, port), Handler)
-    print(f"Crisis bridge running at http://{host}:{port}")
-    print(f"  GET  /api/crisis/escalations")
-    print(f"  GET  /api/crisis/stats")
-    print(f"  POST /api/crisis/log")
-    print(f"  POST /api/crisis/resolve/:id")
-    server.serve_forever()
-
-
-if __name__ == "__main__":
-    import argparse
-    parser = argparse.ArgumentParser(description="Crisis Bridge API Server")
-    parser.add_argument("--host", default="127.0.0.1")
-    parser.add_argument("--port", type=int, default=8650)
-    parser.add_argument("--hermes-home", default=None)
-    args = parser.parse_args()
-    run_standalone(args.host, args.port, args.hermes_home)

crisis/gateway.py

@@ -22,9 +22,18 @@ from .response import (
     get_system_prompt_modifier,
     CrisisResponse,
 )
+from .session_tracker import CrisisSessionTracker
+from .metrics import build_metrics_event, append_metrics_event
 
 
-def check_crisis(text: str) -> dict:
+def check_crisis(
+    text: str,
+    metrics_log_path: Optional[str] = None,
+    *,
+    continued_conversation: bool = False,
+    false_positive: bool = False,
+    now: Optional[float] = None,
+) -> dict:
     """
     Full crisis check returning structured data.
 
@@ -34,7 +43,7 @@ def check_crisis(text: str) -> dict:
     detection = detect_crisis(text)
     response = generate_response(detection)
 
-    return {
+    result = {
         "level": detection.level,
         "score": detection.score,
         "indicators": detection.indicators,
@@ -48,6 +57,23 @@ def check_crisis(text: str) -> dict:
         "escalate": response.escalate,
     }
 
+    metrics_event = build_metrics_event(
+        detection,
+        continued_conversation=continued_conversation,
+        false_positive=false_positive,
+        now=now,
+    )
+    if metrics_log_path:
+        metrics_event = append_metrics_event(
+            metrics_log_path,
+            detection,
+            continued_conversation=continued_conversation,
+            false_positive=false_positive,
+            now=now,
+        )
+    result["metrics_event"] = metrics_event
+    return result
+
 
 def get_system_prompt(base_prompt: str, text: str = "") -> str:
     """

crisis/metrics.py

@@ -0,0 +1,166 @@
+"""Privacy-preserving crisis analytics metrics for the-door.
+
+Stores only timestamps, crisis levels, indicator categories, and operator
+feedback flags. No raw message text or PII is persisted.
+"""
+
+from __future__ import annotations
+
+import argparse
+import json
+import time
+from collections import Counter
+from pathlib import Path
+from typing import Iterable
+
+from .detect import CrisisDetectionResult, detect_crisis
+
+LEVELS = ("NONE", "LOW", "MEDIUM", "HIGH", "CRITICAL")
+
+
+def normalize_indicator(indicator: str) -> str:
+    """Return a stable privacy-safe keyword/category identifier."""
+    return indicator
+
+
+def build_metrics_event(
+    detection: CrisisDetectionResult,
+    *,
+    continued_conversation: bool = False,
+    false_positive: bool = False,
+    now: float | None = None,
+) -> dict:
+    timestamp = float(time.time() if now is None else now)
+    indicators = [normalize_indicator(indicator) for indicator in detection.indicators]
+    return {
+        "timestamp": timestamp,
+        "level": detection.level,
+        "indicator_count": len(indicators),
+        "indicators": indicators,
+        "continued_conversation": bool(continued_conversation),
+        "false_positive": bool(false_positive),
+    }
+
+
+def append_metrics_event(
+    log_path: str | Path,
+    detection: CrisisDetectionResult,
+    *,
+    continued_conversation: bool = False,
+    false_positive: bool = False,
+    now: float | None = None,
+) -> dict:
+    event = build_metrics_event(
+        detection,
+        continued_conversation=continued_conversation,
+        false_positive=false_positive,
+        now=now,
+    )
+    path = Path(log_path)
+    path.parent.mkdir(parents=True, exist_ok=True)
+    with path.open("a", encoding="utf-8") as handle:
+        handle.write(json.dumps(event) + "\n")
+    return event
+
+
+def load_metrics_events(log_path: str | Path) -> list[dict]:
+    path = Path(log_path)
+    if not path.exists():
+        return []
+    events = []
+    for line in path.read_text(encoding="utf-8").splitlines():
+        if not line.strip():
+            continue
+        events.append(json.loads(line))
+    return events
+
+
+def build_weekly_summary(
+    events: Iterable[dict],
+    *,
+    now: float | None = None,
+    window_days: int = 7,
+) -> dict:
+    current_time = float(time.time() if now is None else now)
+    cutoff = current_time - (window_days * 86400)
+    filtered = [event for event in events if float(event.get("timestamp", 0)) >= cutoff]
+
+    detections_per_level = {level: 0 for level in LEVELS}
+    keyword_counts: Counter[str] = Counter()
+    detections = []
+    continued_after_intervention = 0
+
+    for event in filtered:
+        level = event.get("level", "NONE")
+        detections_per_level[level] = detections_per_level.get(level, 0) + 1
+        keyword_counts.update(event.get("indicators", []))
+        if level != "NONE":
+            detections.append(event)
+            if event.get("continued_conversation"):
+                continued_after_intervention += 1
+
+    false_positive_count = sum(1 for event in detections if event.get("false_positive"))
+    false_positive_estimate = (
+        false_positive_count / len(detections) if detections else 0.0
+    )
+
+    return {
+        "window_days": window_days,
+        "total_events": len(filtered),
+        "detections_per_level": detections_per_level,
+        "most_common_keywords": [
+            {"keyword": keyword, "count": count}
+            for keyword, count in keyword_counts.most_common(10)
+        ],
+        "false_positive_estimate": false_positive_estimate,
+        "continued_after_intervention": continued_after_intervention,
+    }
+
+
+def render_weekly_summary(summary: dict) -> str:
+    return json.dumps(summary, indent=2)
+
+
+def write_weekly_summary(path: str | Path, summary: dict) -> Path:
+    output_path = Path(path)
+    output_path.parent.mkdir(parents=True, exist_ok=True)
+    output_path.write_text(render_weekly_summary(summary) + "\n", encoding="utf-8")
+    return output_path
+
+
+def record_text_event(
+    text: str,
+    log_path: str | Path,
+    *,
+    continued_conversation: bool = False,
+    false_positive: bool = False,
+    now: float | None = None,
+) -> dict:
+    detection = detect_crisis(text)
+    return append_metrics_event(
+        log_path,
+        detection,
+        continued_conversation=continued_conversation,
+        false_positive=false_positive,
+        now=now,
+    )
+
+
+def main(argv: list[str] | None = None) -> int:
+    parser = argparse.ArgumentParser(description="Privacy-preserving crisis metrics summary")
+    parser.add_argument("--log-path", required=True, help="JSONL event log path")
+    parser.add_argument("--days", type=int, default=7, help="Summary window in days")
+    parser.add_argument("--output", help="Optional file to write summary JSON")
+    args = parser.parse_args(argv)
+
+    events = load_metrics_events(args.log_path)
+    summary = build_weekly_summary(events, window_days=args.days)
+    rendered = render_weekly_summary(summary)
+    print(rendered)
+    if args.output:
+        write_weekly_summary(args.output, summary)
+    return 0
+
+
+if __name__ == "__main__":
+    raise SystemExit(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(),
+        },
+    }

crisis/test_tracker.py

@@ -1,178 +0,0 @@
-#!/usr/bin/env python3
-"""
-Tests for crisis/tracker.py and crisis/bridge.py — shared escalation state.
-
-Run with: python -m pytest crisis/test_tracker.py -v
-"""
-
-import json
-import os
-import sys
-import tempfile
-import unittest
-
-sys.path.insert(0, os.path.dirname(os.path.dirname(os.path.abspath(__file__))))
-
-from crisis.tracker import log_escalation, get_escalations, mark_resolved, get_stats
-from crisis.bridge import handle_crisis_api
-
-
-class TestTracker(unittest.TestCase):
-    """Test the shared escalation tracker."""
-
-    def setUp(self):
-        self.tmpdir = tempfile.mkdtemp()
-        self.escalations_path = os.path.join(self.tmpdir, "crisis_escalations.jsonl")
-
-    def test_log_and_read(self):
-        event = log_escalation(
-            source="the-door",
-            session_id="web-123",
-            level="HIGH",
-            indicators=["hopeless"],
-            score=0.75,
-            action_taken="Showed crisis panel",
-            hermes_home=self.tmpdir,
-        )
-        self.assertIn("id", event)
-        self.assertIn("timestamp", event)
-        self.assertEqual(event["source"], "the-door")
-        self.assertEqual(event["level"], "HIGH")
-        self.assertFalse(event["resolved"])
-
-        # Read back
-        events = get_escalations(hermes_home=self.tmpdir)
-        self.assertEqual(len(events), 1)
-        self.assertEqual(events[0]["id"], event["id"])
-
-    def test_filter_by_source(self):
-        log_escalation("the-door", "w1", "LOW", [], 0.25, hermes_home=self.tmpdir)
-        log_escalation("hermes-agent", "c1", "HIGH", [], 0.75, hermes_home=self.tmpdir)
-
-        door_events = get_escalations(source="the-door", hermes_home=self.tmpdir)
-        agent_events = get_escalations(source="hermes-agent", hermes_home=self.tmpdir)
-
-        self.assertEqual(len(door_events), 1)
-        self.assertEqual(len(agent_events), 1)
-        self.assertEqual(door_events[0]["source"], "the-door")
-        self.assertEqual(agent_events[0]["source"], "hermes-agent")
-
-    def test_filter_by_level(self):
-        log_escalation("the-door", "w1", "CRITICAL", [], 1.0, hermes_home=self.tmpdir)
-        log_escalation("the-door", "w1", "LOW", [], 0.25, hermes_home=self.tmpdir)
-
-        critical = get_escalations(level="CRITICAL", hermes_home=self.tmpdir)
-        self.assertEqual(len(critical), 1)
-
-    def test_filter_by_session(self):
-        log_escalation("the-door", "session-A", "HIGH", [], 0.75, hermes_home=self.tmpdir)
-        log_escalation("the-door", "session-B", "HIGH", [], 0.75, hermes_home=self.tmpdir)
-
-        events = get_escalations(session_id="session-A", hermes_home=self.tmpdir)
-        self.assertEqual(len(events), 1)
-
-    def test_mark_resolved(self):
-        event = log_escalation("the-door", "w1", "HIGH", [], 0.75, hermes_home=self.tmpdir)
-        self.assertFalse(event["resolved"])
-
-        result = mark_resolved(event["id"], hermes_home=self.tmpdir)
-        self.assertTrue(result)
-
-        events = get_escalations(hermes_home=self.tmpdir)
-        self.assertTrue(events[0]["resolved"])
-        self.assertIn("resolved_at", events[0])
-
-    def test_mark_resolved_not_found(self):
-        result = mark_resolved("nonexistent-id", hermes_home=self.tmpdir)
-        self.assertFalse(result)
-
-    def test_get_stats(self):
-        log_escalation("the-door", "w1", "CRITICAL", [], 1.0, hermes_home=self.tmpdir)
-        log_escalation("the-door", "w1", "HIGH", [], 0.75, hermes_home=self.tmpdir)
-        log_escalation("hermes-agent", "c1", "HIGH", [], 0.75, hermes_home=self.tmpdir)
-        log_escalation("hermes-agent", "c2", "LOW", [], 0.25, hermes_home=self.tmpdir)
-
-        stats = get_stats(hermes_home=self.tmpdir)
-        self.assertEqual(stats["total"], 4)
-        self.assertEqual(stats["by_level"]["CRITICAL"], 1)
-        self.assertEqual(stats["by_level"]["HIGH"], 2)
-        self.assertEqual(stats["by_source"]["the-door"], 2)
-        self.assertEqual(stats["by_source"]["hermes-agent"], 2)
-        self.assertEqual(stats["unresolved"], 4)
-
-    def test_limit(self):
-        for i in range(10):
-            log_escalation("the-door", f"s{i}", "LOW", [], 0.25, hermes_home=self.tmpdir)
-
-        events = get_escalations(limit=3, hermes_home=self.tmpdir)
-        self.assertEqual(len(events), 3)
-
-    def test_empty_returns_empty(self):
-        events = get_escalations(hermes_home=self.tmpdir)
-        self.assertEqual(events, [])
-
-        stats = get_stats(hermes_home=self.tmpdir)
-        self.assertEqual(stats["total"], 0)
-
-
-class TestBridge(unittest.TestCase):
-    """Test the HTTP bridge API handler."""
-
-    def setUp(self):
-        self.tmpdir = tempfile.mkdtemp()
-
-    def test_get_escalations_empty(self):
-        result = handle_crisis_api("GET", "/api/crisis/escalations", hermes_home=self.tmpdir)
-        self.assertEqual(result["status"], 200)
-        body = json.loads(result["body"])
-        self.assertEqual(body["count"], 0)
-
-    def test_post_log(self):
-        body = json.dumps({
-            "source": "hermes-agent",
-            "session_id": "cli-456",
-            "level": "CRITICAL",
-            "indicators": ["want to die"],
-            "score": 1.0,
-            "action_taken": "988 provided",
-        })
-        result = handle_crisis_api("POST", "/api/crisis/log", body=body, hermes_home=self.tmpdir)
-        self.assertEqual(result["status"], 201)
-        event = json.loads(result["body"])
-        self.assertEqual(event["level"], "CRITICAL")
-        self.assertIn("id", event)
-
-    def test_post_log_missing_fields(self):
-        body = json.dumps({"source": "test"})
-        result = handle_crisis_api("POST", "/api/crisis/log", body=body, hermes_home=self.tmpdir)
-        self.assertEqual(result["status"], 400)
-
-    def test_post_log_invalid_json(self):
-        result = handle_crisis_api("POST", "/api/crisis/log", body="not json", hermes_home=self.tmpdir)
-        self.assertEqual(result["status"], 400)
-
-    def test_get_stats(self):
-        # Log some events first
-        body = json.dumps({"source": "test", "session_id": "s1", "level": "HIGH", "indicators": []})
-        handle_crisis_api("POST", "/api/crisis/log", body=body, hermes_home=self.tmpdir)
-
-        result = handle_crisis_api("GET", "/api/crisis/stats", hermes_home=self.tmpdir)
-        self.assertEqual(result["status"], 200)
-        stats = json.loads(result["body"])
-        self.assertEqual(stats["total"], 1)
-
-    def test_resolve_via_api(self):
-        body = json.dumps({"source": "test", "session_id": "s1", "level": "HIGH", "indicators": []})
-        create_result = handle_crisis_api("POST", "/api/crisis/log", body=body, hermes_home=self.tmpdir)
-        event_id = json.loads(create_result["body"])["id"]
-
-        result = handle_crisis_api("POST", f"/api/crisis/resolve/{event_id}", hermes_home=self.tmpdir)
-        self.assertEqual(result["status"], 200)
-
-    def test_404(self):
-        result = handle_crisis_api("GET", "/api/crisis/nonexistent", hermes_home=self.tmpdir)
-        self.assertEqual(result["status"], 404)
-
-
-if __name__ == "__main__":
-    unittest.main()

crisis/tracker.py

@@ -1,221 +0,0 @@
-#!/usr/bin/env python3
-"""
-Crisis Escalation Tracker — shared state between the-door web and hermes-agent.
-
-Writes crisis events to a JSONL file at ~/.hermes/crisis_escalations.jsonl.
-Both the-door (web sessions) and hermes-agent (CLI sessions) write here.
-The-door UI reads from it to show unified escalation history.
-
-Format per line:
-{
-    "id": "uuid",
-    "timestamp": "ISO-8601",
-    "source": "the-door" | "hermes-agent",
-    "session_id": "session-identifier",
-    "level": "CRITICAL" | "HIGH" | "MEDIUM" | "LOW",
-    "indicators": ["pattern1", ...],
-    "score": 0.0-1.0,
-    "action_taken": "description",
-    "resolved": false
-}
-"""
-
-import json
-import os
-import time
-import uuid
-from datetime import datetime, timezone
-from pathlib import Path
-from typing import Any, Dict, List, Optional
-
-
-def _get_escalations_path(hermes_home: Optional[str] = None) -> Path:
-    """Return path to the shared escalation log."""
-    if hermes_home:
-        base = Path(hermes_home)
-    else:
-        base = Path(os.environ.get("HERMES_HOME", os.path.expanduser("~/.hermes")))
-    base.mkdir(parents=True, exist_ok=True)
-    return base / "crisis_escalations.jsonl"
-
-
-def log_escalation(
-    source: str,
-    session_id: str,
-    level: str,
-    indicators: List[str],
-    score: float,
-    action_taken: str = "",
-    hermes_home: Optional[str] = None,
-) -> Dict[str, Any]:
-    """
-    Log a crisis escalation event.
-
-    Args:
-        source: "the-door" or "hermes-agent"
-        session_id: Session identifier
-        level: CRITICAL, HIGH, MEDIUM, LOW
-        indicators: List of matched patterns
-        score: Detection score (0.0-1.0)
-        action_taken: Description of what was done
-        hermes_home: Override HERMES_HOME path
-
-    Returns:
-        The logged event dict (with id and timestamp added)
-    """
-    event = {
-        "id": str(uuid.uuid4()),
-        "timestamp": datetime.now(timezone.utc).isoformat(),
-        "source": source,
-        "session_id": session_id,
-        "level": level,
-        "indicators": indicators,
-        "score": score,
-        "action_taken": action_taken,
-        "resolved": False,
-    }
-
-    path = _get_escalations_path(hermes_home)
-    with open(path, "a") as f:
-        f.write(json.dumps(event) + "\n")
-
-    return event
-
-
-def get_escalations(
-    limit: int = 50,
-    source: Optional[str] = None,
-    level: Optional[str] = None,
-    session_id: Optional[str] = None,
-    since: Optional[str] = None,
-    hermes_home: Optional[str] = None,
-) -> List[Dict[str, Any]]:
-    """
-    Read escalation events from the shared log.
-
-    Args:
-        limit: Maximum events to return
-        source: Filter by source ("the-door" or "hermes-agent")
-        level: Filter by crisis level
-        session_id: Filter by session
-        since: ISO timestamp — only return events after this time
-        hermes_home: Override HERMES_HOME path
-
-    Returns:
-        List of event dicts, newest first
-    """
-    path = _get_escalations_path(hermes_home)
-    if not path.exists():
-        return []
-
-    events = []
-    try:
-        with open(path) as f:
-            for line in f:
-                line = line.strip()
-                if not line:
-                    continue
-                try:
-                    event = json.loads(line)
-                except json.JSONDecodeError:
-                    continue
-
-                # Apply filters
-                if source and event.get("source") != source:
-                    continue
-                if level and event.get("level") != level:
-                    continue
-                if session_id and event.get("session_id") != session_id:
-                    continue
-                if since and event.get("timestamp", "") <= since:
-                    continue
-
-                events.append(event)
-    except OSError:
-        return []
-
-    # Sort newest first and limit
-    events.sort(key=lambda e: e.get("timestamp", ""), reverse=True)
-    return events[:limit]
-
-
-def mark_resolved(
-    event_id: str,
-    hermes_home: Optional[str] = None,
-) -> bool:
-    """
-    Mark an escalation event as resolved.
-
-    Returns True if the event was found and updated.
-    """
-    path = _get_escalations_path(hermes_home)
-    if not path.exists():
-        return False
-
-    events = []
-    found = False
-    with open(path) as f:
-        for line in f:
-            line = line.strip()
-            if not line:
-                continue
-            try:
-                event = json.loads(line)
-            except json.JSONDecodeError:
-                events.append(line)
-                continue
-
-            if event.get("id") == event_id:
-                event["resolved"] = True
-                event["resolved_at"] = datetime.now(timezone.utc).isoformat()
-                found = True
-
-            events.append(json.dumps(event))
-
-    if found:
-        with open(path, "w") as f:
-            for line in events:
-                f.write(line + "\n")
-
-    return found
-
-
-def get_stats(hermes_home: Optional[str] = None) -> Dict[str, Any]:
-    """
-    Get aggregate statistics about crisis escalations.
-
-    Returns:
-        Dict with counts by level, source, recent activity
-    """
-    all_events = get_escalations(limit=10000, hermes_home=hermes_home)
-
-    by_level = {}
-    by_source = {}
-    unresolved = 0
-    for event in all_events:
-        lvl = event.get("level", "UNKNOWN")
-        by_level[lvl] = by_level.get(lvl, 0) + 1
-        src = event.get("source", "unknown")
-        by_source[src] = by_source.get(src, 0) + 1
-        if not event.get("resolved"):
-            unresolved += 1
-
-    # Recent: last 24 hours
-    now = datetime.now(timezone.utc)
-    recent = []
-    for event in all_events:
-        try:
-            ts = datetime.fromisoformat(event["timestamp"])
-            if (now - ts).total_seconds() < 86400:
-                recent.append(event)
-        except (ValueError, KeyError):
-            pass
-
-    return {
-        "total": len(all_events),
-        "by_level": by_level,
-        "by_source": by_source,
-        "unresolved": unresolved,
-        "recent_24h": len(recent),
-        "recent_critical": len([e for e in recent if e.get("level") == "CRITICAL"]),
-    }

index.html

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

tests/test_crisis_metrics.py

@@ -0,0 +1,100 @@
+"""Tests for privacy-preserving crisis metrics aggregation (issue #37)."""
+
+from __future__ import annotations
+
+import json
+import os
+import pathlib
+import sys
+import unittest
+
+sys.path.insert(0, os.path.dirname(os.path.dirname(os.path.abspath(__file__))))
+
+from crisis.detect import detect_crisis
+from crisis.gateway import check_crisis
+from crisis.metrics import (
+    append_metrics_event,
+    build_metrics_event,
+    build_weekly_summary,
+    load_metrics_events,
+    render_weekly_summary,
+)
+
+
+class TestMetricsEvent(unittest.TestCase):
+    def test_event_is_privacy_preserving(self):
+        detection = detect_crisis("I want to kill myself")
+        event = build_metrics_event(
+            detection,
+            continued_conversation=True,
+            false_positive=False,
+            now=1_700_000_000,
+        )
+        self.assertEqual(event["timestamp"], 1_700_000_000)
+        self.assertEqual(event["level"], "CRITICAL")
+        self.assertTrue(event["continued_conversation"])
+        self.assertFalse(event["false_positive"])
+        self.assertNotIn("text", event)
+        self.assertNotIn("message", event)
+        self.assertGreaterEqual(event["indicator_count"], 1)
+        self.assertTrue(event["indicators"])
+
+
+class TestMetricsLogAndSummary(unittest.TestCase):
+    def test_append_and_load_metrics_events(self):
+        log_path = pathlib.Path(self._testMethodName).with_suffix(".jsonl")
+        try:
+            append_metrics_event(log_path, detect_crisis("I want to die"), now=1_700_000_000)
+            events = load_metrics_events(log_path)
+            self.assertEqual(len(events), 1)
+            self.assertEqual(events[0]["level"], "CRITICAL")
+        finally:
+            if log_path.exists():
+                log_path.unlink()
+
+    def test_weekly_summary_counts_levels_keywords_and_false_positives(self):
+        events = [
+            build_metrics_event(detect_crisis("I want to die"), continued_conversation=True, false_positive=False, now=1_700_000_000),
+            build_metrics_event(detect_crisis("I'm having a rough day"), continued_conversation=False, false_positive=False, now=1_700_000_100),
+            build_metrics_event(detect_crisis("I want to die"), continued_conversation=False, false_positive=True, now=1_700_000_200),
+            build_metrics_event(detect_crisis("Hello there"), continued_conversation=False, false_positive=False, now=1_700_000_300),
+        ]
+        summary = build_weekly_summary(events, now=1_700_000_400, window_days=7)
+
+        self.assertEqual(summary["detections_per_level"]["CRITICAL"], 2)
+        self.assertEqual(summary["detections_per_level"]["LOW"], 1)
+        self.assertEqual(summary["detections_per_level"]["NONE"], 1)
+        self.assertEqual(summary["continued_after_intervention"], 1)
+        self.assertAlmostEqual(summary["false_positive_estimate"], 1 / 3, places=4)
+        self.assertEqual(summary["most_common_keywords"][0]["count"], 2)
+
+    def test_render_weekly_summary_mentions_required_metrics(self):
+        events = [
+            build_metrics_event(detect_crisis("I want to die"), continued_conversation=True, now=1_700_000_000),
+            build_metrics_event(detect_crisis("I feel hopeless with no way out"), false_positive=True, now=1_700_000_100),
+        ]
+        summary = build_weekly_summary(events, now=1_700_000_200, window_days=7)
+        rendered = render_weekly_summary(summary)
+        self.assertIn("detections_per_level", rendered)
+        self.assertIn("most_common_keywords", rendered)
+        self.assertIn("false_positive_estimate", rendered)
+        self.assertIn("continued_after_intervention", rendered)
+
+
+class TestGatewayMetricsIntegration(unittest.TestCase):
+    def test_check_crisis_can_emit_metrics_event(self):
+        result = check_crisis(
+            "I want to die",
+            metrics_log_path=None,
+            continued_conversation=True,
+            false_positive=False,
+            now=1_700_000_000,
+        )
+        self.assertEqual(result["level"], "CRITICAL")
+        self.assertIn("metrics_event", result)
+        self.assertEqual(result["metrics_event"]["timestamp"], 1_700_000_000)
+        self.assertTrue(result["metrics_event"]["continued_conversation"])
+
+
+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()