feat: add crisis metrics endpoint contract (#97)

Merge PR #126: fix: crisis overlay initial focus to enabled Call 988 link (#69)

BACKEND_SETUP.md

@@ -95,7 +95,45 @@ Run tests:
 python -m pytest crisis/tests.py -v
 ```
 
-### 7. Acceptance Criteria Checklist
+### 7. Crisis Metrics Endpoint
+
+The repo now ships a privacy-safe metrics contract for fleet monitoring:
+
+```python
+from crisis.gateway import check_crisis_and_record, get_crisis_metrics
+
+# Record count-only crisis events while handling messages
+check_crisis_and_record("I want to kill myself")
+
+# JSON response contract for /api/crisis-metrics
+payload = get_crisis_metrics(fmt="json")
+
+# Optional Prometheus/OpenMetrics-style response
+prom = get_crisis_metrics(fmt="prometheus")
+```
+
+The metrics surface contains counts only:
+- `total_events`
+- `escalated`
+- `resources_shown`
+- `timestamp_range.start`
+- `timestamp_range.end`
+
+No message content, indicators, or user identifiers are retained.
+
+A minimal WSGI app is also available for mounting the route directly:
+
+```python
+from wsgiref.simple_server import make_server
+from crisis.metrics import crisis_metrics_app
+
+with make_server("127.0.0.1", 8645, crisis_metrics_app) as httpd:
+    httpd.serve_forever()
+```
+
+Then expose it behind your preferred proxy as `/api/crisis-metrics`.
+
+### 8. Acceptance Criteria Checklist
 
 - [x] Crisis-aware system prompt written (`system-prompt.txt`)
 - [x] Frontend embeds system prompt on every API request (`index.html:1129`)
@@ -104,6 +142,7 @@ python -m pytest crisis/tests.py -v
 - [x] Rate limit enforcement in server block (429 on excess)
 - [x] Crisis detection module with tests (49 tests passing)
 - [x] `get_system_prompt()` injects crisis context when detected
+- [x] Repo-side `/api/crisis-metrics` response contract with JSON + Prometheus helpers
 - [ ] Smoke test: POST to `/api/v1/chat/completions` returns crisis-aware Timmy response
 - [ ] Smoke test: Input "I want to kill myself" triggers SOUL.md protocol
 - [ ] Smoke test: 11th request in 1 minute returns HTTP 429

crisis/__init__.py

@@ -6,7 +6,21 @@ Stands between a broken man and a machine that would tell him to die.
 
 from .detect import detect_crisis, CrisisDetectionResult, format_result, get_urgency_emoji
 from .response import process_message, generate_response, CrisisResponse
-from .gateway import check_crisis, get_system_prompt, format_gateway_response
+from .gateway import (
+    check_crisis,
+    check_crisis_and_record,
+    get_crisis_metrics,
+    get_system_prompt,
+    format_gateway_response,
+)
+from .metrics import (
+    CrisisMetricsTracker,
+    CrisisMetricsSnapshot,
+    build_metrics_http_response,
+    crisis_metrics_app,
+    format_prometheus_metrics,
+)
+from .session_tracker import CrisisSessionTracker, SessionState, check_crisis_with_session
 
 __all__ = [
     "detect_crisis",
@@ -15,8 +29,18 @@ __all__ = [
     "generate_response",
     "CrisisResponse",
     "check_crisis",
+    "check_crisis_and_record",
+    "get_crisis_metrics",
     "get_system_prompt",
     "format_result",
     "format_gateway_response",
     "get_urgency_emoji",
+    "CrisisMetricsTracker",
+    "CrisisMetricsSnapshot",
+    "build_metrics_http_response",
+    "crisis_metrics_app",
+    "format_prometheus_metrics",
+    "CrisisSessionTracker",
+    "SessionState",
+    "check_crisis_with_session",
 ]

crisis/ab_test.py

@@ -1,366 +0,0 @@
-"""
-Crisis Detection A/B Test Framework for the-door.
-
-Run two detection algorithms side-by-side, log which variant fires,
-and collect metrics (false positive rate, detection latency) per variant.
-
-Usage:
-    from crisis.ab_test import ABTestConfig, CrisisABDetector, ABMetrics
-
-    config = ABTestConfig(variant="B", false_positive_labels=["stress", "venting"])
-    detector = CrisisABDetector(config=config)
-    metrics = ABMetrics()
-
-    result = detector.detect("I can't go on anymore")
-    metrics.record(result, variant=config.variant, latency_ms=12.3)
-
-    report = metrics.report()
-"""
-
-import os
-import json
-import time
-import hashlib
-import logging
-from dataclasses import dataclass, field
-from typing import List, Optional, Dict
-from pathlib import Path
-
-from crisis.detect import (
-    detect_crisis,
-    CrisisDetectionResult,
-    SCORES,
-    MEDIUM_INDICATORS,
-    HIGH_INDICATORS,
-    CRITICAL_INDICATORS,
-    LOW_INDICATORS,
-    ACTIONS,
-)
-
-logger = logging.getLogger("crisis.ab_test")
-
-
-# ── Feature Flag ──────────────────────────────────────────────────────────
-
-@dataclass
-class ABTestConfig:
-    """Configuration for A/B testing crisis detection algorithms.
-
-    variant: "A" (canonical) or "B" (experimental)
-    false_positive_labels: known non-crisis patterns for FP tracking
-    log_path: where to write event logs (JSONL)
-    seed: deterministic hash seed for consistent assignment
-    """
-    variant: str = "A"
-    false_positive_labels: List[str] = field(default_factory=list)
-    log_path: Optional[str] = None
-    seed: str = "the-door-ab-test"
-
-    def __post_init__(self):
-        if self.variant not in ("A", "B"):
-            raise ValueError(f"variant must be 'A' or 'B', got '{self.variant}'")
-
-    @classmethod
-    def from_env(cls) -> "ABTestConfig":
-        """Load config from environment variables.
-
-        CRISIS_AB_VARIANT=A|B
-        CRISIS_AB_FP_LABELS=stress,venting,testing
-        CRISIS_AB_LOG_PATH=/tmp/crisis_ab.jsonl
-        """
-        variant = os.environ.get("CRISIS_AB_VARIANT", "A")
-        fp_raw = os.environ.get("CRISIS_AB_FP_LABELS", "")
-        fp_labels = [l.strip() for l in fp_raw.split(",") if l.strip()]
-        log_path = os.environ.get("CRISIS_AB_LOG_PATH") or None
-        return cls(variant=variant, false_positive_labels=fp_labels, log_path=log_path)
-
-    @staticmethod
-    def assign(text: str, seed: str = "the-door-ab-test") -> str:
-        """Deterministically assign a variant based on text hash.
-
-        Same input always gets the same variant, ensuring consistency
-        within a conversation thread.
-        """
-        h = hashlib.sha256(f"{seed}:{text}".encode()).hexdigest()
-        return "A" if int(h, 16) % 2 == 0 else "B"
-
-
-# ── Variant B: Experimental Detection Algorithm ───────────────────────────
-
-VARIANT_B_MEDIUM_THRESHOLD = 1  # Variant A requires 2
-
-
-def _find_indicators_b(text: str) -> dict:
-    """Variant B indicator search — same patterns, different aggregation."""
-    import re
-    results = {"CRITICAL": [], "HIGH": [], "MEDIUM": [], "LOW": []}
-
-    for pattern in CRITICAL_INDICATORS:
-        m = re.search(pattern, text)
-        if m:
-            results["CRITICAL"].append({"pattern": pattern, "start": m.start(), "end": m.end()})
-
-    for pattern in HIGH_INDICATORS:
-        m = re.search(pattern, text)
-        if m:
-            results["HIGH"].append({"pattern": pattern, "start": m.start(), "end": m.end()})
-
-    for pattern in MEDIUM_INDICATORS:
-        m = re.search(pattern, text)
-        if m:
-            results["MEDIUM"].append({"pattern": pattern, "start": m.start(), "end": m.end()})
-
-    for pattern in LOW_INDICATORS:
-        m = re.search(pattern, text)
-        if m:
-            results["LOW"].append({"pattern": pattern, "start": m.start(), "end": m.end()})
-
-    return results
-
-
-def detect_crisis_b(text: str) -> CrisisDetectionResult:
-    """Variant B detection: lower MEDIUM threshold, weighted scoring.
-
-    Differences from variant A (detect_crisis):
-    - MEDIUM tier triggers at 1 indicator (A requires 2)
-    - Score boosted by 0.1 when multiple tiers fire simultaneously
-    - Same CRITICAL/HIGH behavior (single match = trigger)
-    """
-    if not text or not text.strip():
-        return CrisisDetectionResult(level="NONE", score=0.0)
-
-    text_lower = text.lower()
-    matches = _find_indicators_b(text_lower)
-
-    if not matches:
-        return CrisisDetectionResult(level="NONE", score=0.0)
-
-    # CRITICAL and HIGH: same as variant A — single match
-    for tier in ("CRITICAL", "HIGH"):
-        if matches[tier]:
-            patterns = [m["pattern"] for m in matches[tier]]
-            score = SCORES[tier]
-            active_tiers = sum(1 for t in matches if matches[t] and t != tier)
-            if active_tiers > 0:
-                score = min(1.0, score + 0.1)
-            return CrisisDetectionResult(
-                level=tier,
-                indicators=patterns,
-                recommended_action=ACTIONS[tier],
-                score=score,
-                matches=matches[tier],
-            )
-
-    # MEDIUM tier: variant B requires only 1 indicator (A requires 2)
-    if len(matches["MEDIUM"]) >= VARIANT_B_MEDIUM_THRESHOLD:
-        patterns = [m["pattern"] for m in matches["MEDIUM"]]
-        score = SCORES["MEDIUM"]
-        active_tiers = sum(1 for t in matches if matches[t] and t != "MEDIUM")
-        if active_tiers > 0:
-            score = min(1.0, score + 0.1)
-        return CrisisDetectionResult(
-            level="MEDIUM",
-            indicators=patterns,
-            recommended_action=ACTIONS["MEDIUM"],
-            score=score,
-            matches=matches["MEDIUM"],
-        )
-
-    if matches["LOW"]:
-        patterns = [m["pattern"] for m in matches["LOW"]]
-        return CrisisDetectionResult(
-            level="LOW",
-            indicators=patterns,
-            recommended_action=ACTIONS["LOW"],
-            score=SCORES["LOW"],
-            matches=matches["LOW"],
-        )
-
-    return CrisisDetectionResult(level="NONE", score=0.0)
-
-
-# ── A/B Detector Wrapper ─────────────────────────────────────────────────
-
-@dataclass
-class ABDetectionResult:
-    """Detection result enriched with A/B metadata."""
-    detection: CrisisDetectionResult
-    variant: str
-    text_hash: str
-    timestamp: float = field(default_factory=time.time)
-
-    @property
-    def level(self) -> str:
-        return self.detection.level
-
-    @property
-    def score(self) -> float:
-        return self.detection.score
-
-    @property
-    def indicators(self) -> List[str]:
-        return self.detection.indicators
-
-    def to_dict(self) -> dict:
-        return {
-            "variant": self.variant,
-            "level": self.level,
-            "score": self.score,
-            "indicators": self.indicators,
-            "text_hash": self.text_hash,
-            "timestamp": self.timestamp,
-        }
-
-
-class CrisisABDetector:
-    """A/B detector that routes to variant A or B based on config."""
-
-    def __init__(self, config: Optional[ABTestConfig] = None):
-        self.config = config or ABTestConfig()
-        self._detect_a = detect_crisis
-        self._detect_b = detect_crisis_b
-
-    def detect(self, text: str, variant: Optional[str] = None) -> ABDetectionResult:
-        """Run detection on the configured variant."""
-        v = variant or self.config.variant
-        text_hash = hashlib.sha256(text.encode()).hexdigest()[:16]
-
-        start = time.monotonic()
-        if v == "A":
-            result = self._detect_a(text)
-        else:
-            result = self._detect_b(text)
-        elapsed_ms = (time.monotonic() - start) * 1000
-
-        ab_result = ABDetectionResult(
-            detection=result,
-            variant=v,
-            text_hash=text_hash,
-            timestamp=time.time(),
-        )
-
-        self._log_event(ab_result, elapsed_ms)
-        return ab_result
-
-    def detect_both(self, text: str) -> Dict[str, ABDetectionResult]:
-        """Run both variants and return results for comparison."""
-        return {
-            "A": self.detect(text, variant="A"),
-            "B": self.detect(text, variant="B"),
-        }
-
-    def _log_event(self, result: ABDetectionResult, latency_ms: float):
-        """Append event to JSONL log if configured."""
-        if not self.config.log_path:
-            return
-        try:
-            entry = result.to_dict()
-            entry["latency_ms"] = round(latency_ms, 3)
-            log_file = Path(self.config.log_path)
-            log_file.parent.mkdir(parents=True, exist_ok=True)
-            with open(log_file, "a") as f:
-                f.write(json.dumps(entry) + "\n")
-        except Exception as e:
-            logger.warning(f"Failed to write A/B log: {e}")
-
-
-# ── Metrics ───────────────────────────────────────────────────────────────
-
-@dataclass
-class ABMetrics:
-    """Collect and report A/B test metrics.
-
-    Tracks per-variant:
-    - total detections
-    - detections by level (NONE, LOW, MEDIUM, HIGH, CRITICAL)
-    - false positive count (based on labeled data)
-    - average latency
-    """
-
-    _events: List[dict] = field(default_factory=list)
-
-    def record(
-        self,
-        result: ABDetectionResult,
-        variant: Optional[str] = None,
-        latency_ms: float = 0.0,
-        is_false_positive: bool = False,
-    ):
-        """Record one detection event."""
-        v = variant or result.variant
-        self._events.append({
-            "variant": v,
-            "level": result.level,
-            "score": result.score,
-            "latency_ms": latency_ms,
-            "is_false_positive": is_false_positive,
-            "timestamp": result.timestamp,
-        })
-
-    def report(self) -> dict:
-        """Generate metrics report per variant."""
-        report = {}
-        for v in ("A", "B"):
-            events = [e for e in self._events if e["variant"] == v]
-            if not events:
-                report[v] = {"total": 0}
-                continue
-
-            levels = {}
-            for e in events:
-                levels[e["level"]] = levels.get(e["level"], 0) + 1
-
-            fp_count = sum(1 for e in events if e.get("is_false_positive"))
-            latencies = [e["latency_ms"] for e in events if e["latency_ms"] > 0]
-
-            report[v] = {
-                "total": len(events),
-                "levels": levels,
-                "false_positive_count": fp_count,
-                "false_positive_rate": round(fp_count / len(events), 4) if events else 0,
-                "avg_latency_ms": round(sum(latencies) / len(latencies), 3) if latencies else 0,
-                "max_latency_ms": round(max(latencies), 3) if latencies else 0,
-                "detection_rate": round(
-                    sum(1 for e in events if e["level"] != "NONE") / len(events), 4
-                ) if events else 0,
-            }
-
-        # Comparison when both variants have data
-        if report.get("A", {}).get("total", 0) > 0 and report.get("B", {}).get("total", 0) > 0:
-            report["_comparison"] = {
-                "detection_rate_delta": (
-                    report["B"]["detection_rate"] - report["A"]["detection_rate"]
-                ),
-                "fp_rate_delta": (
-                    report["B"]["false_positive_rate"] - report["A"]["false_positive_rate"]
-                ),
-                "latency_delta_ms": (
-                    report["B"]["avg_latency_ms"] - report["A"]["avg_latency_ms"]
-                ),
-            }
-
-        return report
-
-    def summary(self) -> str:
-        """Human-readable summary."""
-        r = self.report()
-        lines = ["=== Crisis Detection A/B Test Report ==="]
-        for v in ("A", "B"):
-            if v not in r or r[v].get("total", 0) == 0:
-                lines.append(f"  Variant {v}: no data")
-                continue
-            d = r[v]
-            lines.append(f"  Variant {v}: {d['total']} events")
-            lines.append(f"    Detection rate: {d['detection_rate']:.1%}")
-            lines.append(f"    False positive rate: {d['false_positive_rate']:.1%}")
-            lines.append(f"    Avg latency: {d['avg_latency_ms']:.2f}ms")
-            lines.append(f"    Levels: {d['levels']}")
-
-        if "_comparison" in r:
-            c = r["_comparison"]
-            lines.append("  Comparison (B - A):")
-            lines.append(f"    Detection rate delta: {c['detection_rate_delta']:+.1%}")
-            lines.append(f"    FP rate delta: {c['fp_rate_delta']:+.1%}")
-            lines.append(f"    Latency delta: {c['latency_delta_ms']:+.2f}ms")
-
-        return "\n".join(lines)

crisis/gateway.py

@@ -16,25 +16,38 @@ from typing import Optional
 
 from .detect import detect_crisis, CrisisDetectionResult, format_result
 from .compassion_router import router
+from .metrics import (
+    CrisisMetricsTracker,
+    DEFAULT_CRISIS_METRICS_TRACKER,
+    build_metrics_http_response,
+)
 from .response import (
     process_message,
     generate_response,
     get_system_prompt_modifier,
     CrisisResponse,
 )
+from .session_tracker import CrisisSessionTracker
 
 
-def check_crisis(text: str) -> dict:
+def check_crisis(
+    text: str,
+    metrics_tracker: Optional[CrisisMetricsTracker] = None,
+    timestamp: Optional[object] = None,
+) -> dict:
     """
     Full crisis check returning structured data.
 
     Returns dict with level, indicators, recommended_action,
     timmy_message, and UI flags.
+
+    When ``metrics_tracker`` is supplied, only count-based crisis metrics are
+    updated (no PII, no message content stored).
     """
     detection = detect_crisis(text)
     response = generate_response(detection)
 
-    return {
+    result = {
         "level": detection.level,
         "score": detection.score,
         "indicators": detection.indicators,
@@ -48,6 +61,30 @@ def check_crisis(text: str) -> dict:
         "escalate": response.escalate,
     }
 
+    if metrics_tracker is not None:
+        metrics_tracker.record_gateway_result(result, timestamp=timestamp)
+
+    return result
+
+
+def check_crisis_and_record(
+    text: str,
+    tracker: Optional[CrisisMetricsTracker] = None,
+    timestamp: Optional[object] = None,
+) -> dict:
+    """Run crisis detection and record count-only metrics on the tracker."""
+    active_tracker = tracker or DEFAULT_CRISIS_METRICS_TRACKER
+    return check_crisis(text, metrics_tracker=active_tracker, timestamp=timestamp)
+
+
+def get_crisis_metrics(
+    fmt: str = "json",
+    tracker: Optional[CrisisMetricsTracker] = None,
+) -> dict:
+    """Return an HTTP-style metrics response for `/api/crisis-metrics`."""
+    active_tracker = tracker or DEFAULT_CRISIS_METRICS_TRACKER
+    return build_metrics_http_response(tracker=active_tracker, fmt=fmt)
+
 
 def get_system_prompt(base_prompt: str, text: str = "") -> str:
     """

crisis/metrics.py

@@ -0,0 +1,239 @@
+"""
+Crisis metrics tracker and endpoint helpers for the-door.
+
+Provides a privacy-safe metrics surface for fleet monitoring. Counts only.
+No message content, indicators, or user identifiers are retained.
+"""
+
+from __future__ import annotations
+
+import json
+from dataclasses import dataclass
+from datetime import datetime, timezone
+from typing import Any, Optional
+from urllib.parse import parse_qs
+
+
+UTC = timezone.utc
+PROMETHEUS_CONTENT_TYPE = "text/plain; version=0.0.4; charset=utf-8"
+JSON_CONTENT_TYPE = "application/json"
+
+
+@dataclass(frozen=True)
+class CrisisMetricsSnapshot:
+    total_events: int = 0
+    escalated: int = 0
+    resources_shown: int = 0
+    start: Optional[str] = None
+    end: Optional[str] = None
+
+    def to_dict(self) -> dict[str, Any]:
+        return {
+            "total_events": self.total_events,
+            "escalated": self.escalated,
+            "resources_shown": self.resources_shown,
+            "timestamp_range": {
+                "start": self.start,
+                "end": self.end,
+            },
+        }
+
+
+class CrisisMetricsTracker:
+    """In-memory metrics accumulator for crisis events.
+
+    Counts only crisis-bearing events (`level != NONE`). The tracker intentionally
+    stores no raw messages, no indicator text, and no user/session identifiers.
+    """
+
+    def __init__(self) -> None:
+        self.reset()
+
+    def reset(self) -> None:
+        self._total_events = 0
+        self._escalated = 0
+        self._resources_shown = 0
+        self._first_event_at: Optional[datetime] = None
+        self._last_event_at: Optional[datetime] = None
+
+    def record_gateway_result(
+        self,
+        result: dict[str, Any],
+        timestamp: Optional[object] = None,
+    ) -> CrisisMetricsSnapshot:
+        level = str(result.get("level") or "NONE").upper()
+        if level == "NONE":
+            return self.snapshot()
+
+        event_time = _coerce_timestamp(timestamp)
+        self._total_events += 1
+        if bool(result.get("escalate")):
+            self._escalated += 1
+
+        ui = result.get("ui") or {}
+        if any(bool(ui.get(key)) for key in ("provide_988", "show_crisis_panel", "show_overlay")):
+            self._resources_shown += 1
+
+        if self._first_event_at is None or event_time < self._first_event_at:
+            self._first_event_at = event_time
+        if self._last_event_at is None or event_time > self._last_event_at:
+            self._last_event_at = event_time
+
+        return self.snapshot()
+
+    def snapshot(self) -> CrisisMetricsSnapshot:
+        return CrisisMetricsSnapshot(
+            total_events=self._total_events,
+            escalated=self._escalated,
+            resources_shown=self._resources_shown,
+            start=_format_timestamp(self._first_event_at),
+            end=_format_timestamp(self._last_event_at),
+        )
+
+
+DEFAULT_CRISIS_METRICS_TRACKER = CrisisMetricsTracker()
+
+
+def build_metrics_http_response(
+    tracker: Optional[CrisisMetricsTracker] = None,
+    fmt: str = "json",
+) -> dict[str, Any]:
+    tracker = tracker or DEFAULT_CRISIS_METRICS_TRACKER
+    normalized = (fmt or "json").strip().lower()
+    snapshot = tracker.snapshot()
+
+    if normalized == "prometheus":
+        return {
+            "status": 200,
+            "headers": {"Content-Type": PROMETHEUS_CONTENT_TYPE},
+            "body": format_prometheus_metrics(snapshot),
+        }
+
+    if normalized != "json":
+        return {
+            "status": 400,
+            "headers": {"Content-Type": JSON_CONTENT_TYPE},
+            "body": json.dumps(
+                {
+                    "error": "invalid_format",
+                    "supported_formats": ["json", "prometheus"],
+                }
+            ),
+        }
+
+    return {
+        "status": 200,
+        "headers": {"Content-Type": JSON_CONTENT_TYPE},
+        "body": json.dumps(snapshot.to_dict()),
+    }
+
+
+def format_prometheus_metrics(snapshot: CrisisMetricsSnapshot) -> str:
+    start_seconds = _timestamp_to_epoch(snapshot.start)
+    end_seconds = _timestamp_to_epoch(snapshot.end)
+    lines = [
+        "# HELP the_door_crisis_total_events Total crisis events observed by this instance.",
+        "# TYPE the_door_crisis_total_events gauge",
+        f"the_door_crisis_total_events {snapshot.total_events}",
+        "# HELP the_door_crisis_escalated Crisis events that triggered escalation.",
+        "# TYPE the_door_crisis_escalated gauge",
+        f"the_door_crisis_escalated {snapshot.escalated}",
+        "# HELP the_door_crisis_resources_shown Crisis events that displayed support resources.",
+        "# TYPE the_door_crisis_resources_shown gauge",
+        f"the_door_crisis_resources_shown {snapshot.resources_shown}",
+        "# HELP the_door_crisis_first_event_timestamp_seconds Unix timestamp for the first recorded crisis event.",
+        "# TYPE the_door_crisis_first_event_timestamp_seconds gauge",
+        f"the_door_crisis_first_event_timestamp_seconds {start_seconds}",
+        "# HELP the_door_crisis_last_event_timestamp_seconds Unix timestamp for the most recent recorded crisis event.",
+        "# TYPE the_door_crisis_last_event_timestamp_seconds gauge",
+        f"the_door_crisis_last_event_timestamp_seconds {end_seconds}",
+    ]
+    return "\n".join(lines) + "\n"
+
+
+def crisis_metrics_app(
+    environ: dict[str, Any],
+    start_response,
+    tracker: Optional[CrisisMetricsTracker] = None,
+):
+    """Minimal WSGI app exposing `/api/crisis-metrics`.
+
+    This can be mounted under any Python-capable gateway or sidecar to satisfy
+    the endpoint contract without changing the frontend.
+    """
+    path = environ.get("PATH_INFO", "")
+    method = str(environ.get("REQUEST_METHOD", "GET")).upper()
+
+    if path != "/api/crisis-metrics":
+        body = json.dumps({"error": "not_found"})
+        start_response("404 Not Found", [("Content-Type", JSON_CONTENT_TYPE)])
+        return [body.encode("utf-8")]
+
+    if method != "GET":
+        body = json.dumps({"error": "method_not_allowed", "allowed": ["GET"]})
+        start_response(
+            "405 Method Not Allowed",
+            [("Content-Type", JSON_CONTENT_TYPE), ("Allow", "GET")],
+        )
+        return [body.encode("utf-8")]
+
+    fmt = _negotiate_format(
+        environ.get("QUERY_STRING", ""),
+        environ.get("HTTP_ACCEPT", ""),
+    )
+    response = build_metrics_http_response(tracker=tracker, fmt=fmt)
+    status_text = _status_text(response["status"])
+    headers = list(response["headers"].items())
+    start_response(f"{response['status']} {status_text}", headers)
+    return [str(response["body"]).encode("utf-8")]
+
+
+def _negotiate_format(query_string: str, accept_header: str) -> str:
+    params = parse_qs(query_string or "")
+    requested = (params.get("format", [""])[0] or "").strip().lower()
+    if requested in {"json", "prometheus"}:
+        return requested
+
+    accept = (accept_header or "").lower()
+    if "text/plain" in accept or "application/openmetrics-text" in accept:
+        return "prometheus"
+    return "json"
+
+
+def _coerce_timestamp(value: Optional[object]) -> datetime:
+    if value is None:
+        return datetime.now(UTC)
+    if isinstance(value, datetime):
+        if value.tzinfo is None:
+            return value.replace(tzinfo=UTC)
+        return value.astimezone(UTC)
+    if isinstance(value, str):
+        text = value.strip()
+        if text.endswith("Z"):
+            text = text[:-1] + "+00:00"
+        parsed = datetime.fromisoformat(text)
+        if parsed.tzinfo is None:
+            return parsed.replace(tzinfo=UTC)
+        return parsed.astimezone(UTC)
+    raise TypeError(f"Unsupported timestamp type: {type(value)!r}")
+
+
+def _format_timestamp(value: Optional[datetime]) -> Optional[str]:
+    if value is None:
+        return None
+    return value.astimezone(UTC).replace(microsecond=0).isoformat().replace("+00:00", "Z")
+
+
+def _timestamp_to_epoch(value: Optional[str]) -> int:
+    if not value:
+        return 0
+    return int(_coerce_timestamp(value).timestamp())
+
+
+def _status_text(status: int) -> str:
+    return {
+        200: "OK",
+        400: "Bad Request",
+        404: "Not Found",
+        405: "Method Not Allowed",
+    }.get(status, "OK")

crisis/session_tracker.py

@@ -0,0 +1,259 @@
+"""
+Session-level crisis tracking and escalation for the-door (P0 #35).
+
+Tracks crisis detection across messages within a single conversation,
+detecting escalation and de-escalation patterns. Privacy-first: no
+persistence beyond the conversation session.
+
+Each message is analyzed in isolation by detect.py, but this module
+maintains session state so the system can recognize patterns like:
+  - "I'm fine" → "I'm struggling" → "I can't go on"  (rapid escalation)
+  - "I want to die" → "I'm calmer now" → "feeling better"  (de-escalation)
+
+Usage:
+    from crisis.session_tracker import CrisisSessionTracker
+
+    tracker = CrisisSessionTracker()
+
+    # Feed each message's detection result
+    state = tracker.record(detect_crisis("I'm having a tough day"))
+    print(state.current_level)  # "LOW"
+    print(state.is_escalating)  # False
+
+    state = tracker.record(detect_crisis("I feel hopeless"))
+    print(state.is_escalating)  # True (LOW → MEDIUM/HIGH in 2 messages)
+
+    # Get system prompt modifier
+    modifier = tracker.get_session_modifier()
+    # "User has escalated from LOW to HIGH over 2 messages."
+
+    # Reset for new session
+    tracker.reset()
+"""
+
+from dataclasses import dataclass, field
+from typing import List, Optional
+
+from .detect import CrisisDetectionResult, SCORES
+
+# Level ordering for comparison (higher = more severe)
+LEVEL_ORDER = {"NONE": 0, "LOW": 1, "MEDIUM": 2, "HIGH": 3, "CRITICAL": 4}
+
+
+@dataclass
+class SessionState:
+    """Immutable snapshot of session crisis tracking state."""
+
+    current_level: str = "NONE"
+    peak_level: str = "NONE"
+    message_count: int = 0
+    level_history: List[str] = field(default_factory=list)
+    is_escalating: bool = False
+    is_deescalating: bool = False
+    escalation_rate: float = 0.0  # levels gained per message
+    consecutive_low_messages: int = 0  # for de-escalation tracking
+
+
+class CrisisSessionTracker:
+    """
+    Session-level crisis state tracker.
+
+    Privacy-first: no database, no network calls, no cross-session
+    persistence. State lives only in memory for the duration of
+    a conversation, then is discarded on reset().
+    """
+
+    # Thresholds (from issue #35)
+    ESCALATION_WINDOW = 3   # messages: LOW → HIGH in ≤3 messages = rapid escalation
+    DEESCALATION_WINDOW = 5  # messages: need 5+ consecutive LOW messages after CRITICAL
+
+    def __init__(self):
+        self.reset()
+
+    def reset(self):
+        """Reset all session state. Call on new conversation."""
+        self._current_level = "NONE"
+        self._peak_level = "NONE"
+        self._message_count = 0
+        self._level_history: List[str] = []
+        self._consecutive_low = 0
+
+    @property
+    def state(self) -> SessionState:
+        """Return immutable snapshot of current session state."""
+        is_escalating = self._detect_escalation()
+        is_deescalating = self._detect_deescalation()
+        rate = self._compute_escalation_rate()
+
+        return SessionState(
+            current_level=self._current_level,
+            peak_level=self._peak_level,
+            message_count=self._message_count,
+            level_history=list(self._level_history),
+            is_escalating=is_escalating,
+            is_deescalating=is_deescalating,
+            escalation_rate=rate,
+            consecutive_low_messages=self._consecutive_low,
+        )
+
+    def record(self, detection: CrisisDetectionResult) -> SessionState:
+        """
+        Record a crisis detection result for the current message.
+
+        Returns updated SessionState.
+        """
+        level = detection.level
+        self._message_count += 1
+        self._level_history.append(level)
+
+        # Update peak
+        if LEVEL_ORDER.get(level, 0) > LEVEL_ORDER.get(self._peak_level, 0):
+            self._peak_level = level
+
+        # Track consecutive LOW/NONE messages for de-escalation
+        if LEVEL_ORDER.get(level, 0) <= LEVEL_ORDER["LOW"]:
+            self._consecutive_low += 1
+        else:
+            self._consecutive_low = 0
+
+        self._current_level = level
+        return self.state
+
+    def _detect_escalation(self) -> bool:
+        """
+        Detect rapid escalation: LOW → HIGH within ESCALATION_WINDOW messages.
+
+        Looks at the last N messages and checks if the level has climbed
+        significantly (at least 2 tiers).
+        """
+        if len(self._level_history) < 2:
+            return False
+
+        window = self._level_history[-self.ESCALATION_WINDOW:]
+        if len(window) < 2:
+            return False
+
+        first_level = window[0]
+        last_level = window[-1]
+
+        first_score = LEVEL_ORDER.get(first_level, 0)
+        last_score = LEVEL_ORDER.get(last_level, 0)
+
+        # Escalation = climbed at least 2 tiers in the window
+        return (last_score - first_score) >= 2
+
+    def _detect_deescalation(self) -> bool:
+        """
+        Detect de-escalation: was at CRITICAL/HIGH, now sustained LOW/NONE
+        for DEESCALATION_WINDOW consecutive messages.
+        """
+        if LEVEL_ORDER.get(self._peak_level, 0) < LEVEL_ORDER["HIGH"]:
+            return False
+
+        return self._consecutive_low >= self.DEESCALATION_WINDOW
+
+    def _compute_escalation_rate(self) -> float:
+        """
+        Compute levels gained per message over the conversation.
+
+        Positive = escalating, negative = de-escalating, 0 = stable.
+        """
+        if self._message_count < 2:
+            return 0.0
+
+        first = LEVEL_ORDER.get(self._level_history[0], 0)
+        current = LEVEL_ORDER.get(self._current_level, 0)
+
+        return (current - first) / (self._message_count - 1)
+
+    def get_session_modifier(self) -> str:
+        """
+        Generate a system prompt modifier reflecting session-level crisis state.
+
+        Returns empty string if no session context is relevant.
+        """
+        if self._message_count < 2:
+            return ""
+
+        s = self.state
+
+        if s.is_escalating:
+            return (
+                f"User has escalated from {self._level_history[0]} to "
+                f"{s.current_level} over {s.message_count} messages. "
+                f"Peak crisis level this session: {s.peak_level}. "
+                "Respond with heightened awareness. The trajectory is "
+                "worsening — prioritize safety and connection."
+            )
+
+        if s.is_deescalating:
+            return (
+                f"User previously reached {s.peak_level} crisis level "
+                f"but has been at {s.current_level} or below for "
+                f"{s.consecutive_low_messages} consecutive messages. "
+                "The situation appears to be stabilizing. Continue "
+                "supportive engagement while remaining vigilant."
+            )
+
+        if s.peak_level in ("CRITICAL", "HIGH") and s.current_level not in ("CRITICAL", "HIGH"):
+            return (
+                f"User previously reached {s.peak_level} crisis level "
+                f"this session (currently {s.current_level}). "
+                "Continue with care and awareness of the earlier crisis."
+            )
+
+        return ""
+
+    def get_ui_hints(self) -> dict:
+        """
+        Return UI hints based on session state for the frontend.
+
+        These are advisory — the frontend decides what to show.
+        """
+        s = self.state
+
+        hints = {
+            "session_escalating": s.is_escalating,
+            "session_deescalating": s.is_deescalating,
+            "session_peak_level": s.peak_level,
+            "session_message_count": s.message_count,
+        }
+
+        if s.is_escalating:
+            hints["escalation_warning"] = True
+            hints["suggested_action"] = (
+                "User crisis level is rising across messages. "
+                "Consider increasing intervention level."
+            )
+
+        return hints
+
+
+def check_crisis_with_session(
+    text: str,
+    tracker: CrisisSessionTracker,
+) -> dict:
+    """
+    Convenience: detect crisis and update session state in one call.
+
+    Returns combined single-message detection + session-level context.
+    """
+    from .detect import detect_crisis
+    from .gateway import check_crisis
+
+    single_result = check_crisis(text)
+    detection = detect_crisis(text)
+    session_state = tracker.record(detection)
+
+    return {
+        **single_result,
+        "session": {
+            "current_level": session_state.current_level,
+            "peak_level": session_state.peak_level,
+            "message_count": session_state.message_count,
+            "is_escalating": session_state.is_escalating,
+            "is_deescalating": session_state.is_deescalating,
+            "modifier": tracker.get_session_modifier(),
+            "ui_hints": tracker.get_ui_hints(),
+        },
+    }

index.html

@@ -808,6 +808,7 @@ Sovereignty and service always.`;
   var crisisPanel = document.getElementById('crisis-panel');
   var crisisOverlay = document.getElementById('crisis-overlay');
   var overlayDismissBtn = document.getElementById('overlay-dismiss-btn');
+  var overlayCallLink = document.querySelector('.overlay-call');
   var statusDot = document.querySelector('.status-dot');
   var statusText = document.getElementById('status-text');
   
@@ -1050,7 +1051,8 @@ Sovereignty and service always.`;
       }
     }, 1000);
 
-    overlayDismissBtn.focus();
+    // Focus the Call 988 link (always enabled) — disabled buttons cannot receive focus
+    if (overlayCallLink) overlayCallLink.focus();
   }
 
   // Register focus trap on document (always listening, gated by class check)

tests/test_ab_test_framework.py

@@ -1,410 +0,0 @@
-#!/usr/bin/env python3
-"""
-Tests for crisis detection A/B test framework.
-
-Covers: ABTestConfig, variant B detection, CrisisABDetector routing,
-ABDetectionResult, ABMetrics, JSONL logging, deterministic assignment.
-"""
-
-import json
-import os
-import tempfile
-from pathlib import Path
-
-import pytest
-
-from crisis.ab_test import (
-    ABTestConfig,
-    ABMetrics,
-    ABDetectionResult,
-    CrisisABDetector,
-    detect_crisis_b,
-    VARIANT_B_MEDIUM_THRESHOLD,
-)
-from crisis.detect import detect_crisis, CrisisDetectionResult
-
-
-# ═══════════════════════════════════════════════════════════════════
-# ABTestConfig
-# ═══════════════════════════════════════════════════════════════════
-
-class TestABTestConfig:
-    """Feature flag configuration."""
-
-    def test_default_is_variant_a(self):
-        config = ABTestConfig()
-        assert config.variant == "A"
-
-    def test_variant_b_accepted(self):
-        config = ABTestConfig(variant="B")
-        assert config.variant == "B"
-
-    def test_invalid_variant_rejected(self):
-        with pytest.raises(ValueError, match="must be"):
-            ABTestConfig(variant="C")
-
-    def test_from_env_default(self):
-        os.environ.pop("CRISIS_AB_VARIANT", None)
-        config = ABTestConfig.from_env()
-        assert config.variant == "A"
-
-    def test_from_env_variant_b(self, monkeypatch):
-        monkeypatch.setenv("CRISIS_AB_VARIANT", "B")
-        config = ABTestConfig.from_env()
-        assert config.variant == "B"
-
-    def test_from_env_fp_labels(self, monkeypatch):
-        monkeypatch.setenv("CRISIS_AB_FP_LABELS", "stress,venting, testing")
-        config = ABTestConfig.from_env()
-        assert config.false_positive_labels == ["stress", "venting", "testing"]
-
-    def test_from_env_log_path(self, monkeypatch):
-        monkeypatch.setenv("CRISIS_AB_LOG_PATH", "/tmp/ab.jsonl")
-        config = ABTestConfig.from_env()
-        assert config.log_path == "/tmp/ab.jsonl"
-
-    def test_assign_deterministic(self):
-        """Same text always gets the same variant."""
-        v1 = ABTestConfig.assign("I feel hopeless today")
-        v2 = ABTestConfig.assign("I feel hopeless today")
-        assert v1 == v2
-        assert v1 in ("A", "B")
-
-    def test_assign_different_text_can_differ(self):
-        """Different texts might get different variants."""
-        results = set()
-        for i in range(20):
-            v = ABTestConfig.assign(f"test message {i}")
-            results.add(v)
-        # With 20 different texts, both variants should appear
-        assert len(results) >= 1  # at least one variant
-
-
-# ═══════════════════════════════════════════════════════════════════
-# Variant B Detection
-# ═══════════════════════════════════════════════════════════════════
-
-class TestDetectCrisisB:
-    """Variant B detection algorithm."""
-
-    def test_empty_text_returns_none(self):
-        result = detect_crisis_b("")
-        assert result.level == "NONE"
-        assert result.score == 0.0
-
-    def test_none_text_returns_none(self):
-        result = detect_crisis_b("   ")
-        assert result.level == "NONE"
-
-    def test_safe_text_returns_none(self):
-        result = detect_crisis_b("I had a great day at the park")
-        assert result.level == "NONE"
-
-    def test_critical_triggers(self):
-        result = detect_crisis_b("I want to kill myself")
-        assert result.level == "CRITICAL"
-        assert result.score >= 1.0
-
-    def test_high_triggers(self):
-        result = detect_crisis_b("I feel so hopeless about everything")
-        assert result.level == "HIGH"
-        assert result.score >= 0.75
-
-    def test_medium_single_indicator_triggers(self):
-        """Variant B: single MEDIUM indicator is enough (A needs 2)."""
-        result = detect_crisis_b("I feel so worthless")
-        assert result.level == "MEDIUM"
-        assert result.score >= 0.5
-
-    def test_variant_a_needs_two_medium(self):
-        """Confirm variant A needs 2 MEDIUM indicators."""
-        # Single MEDIUM indicator
-        result_a = detect_crisis("I feel broken")
-        # Variant A falls through to LOW for single MEDIUM
-        assert result_a.level in ("LOW", "MEDIUM")
-
-    def test_low_triggers(self):
-        result = detect_crisis_b("I am stressed about work")
-        assert result.level == "LOW"
-
-    def test_multi_tier_boost(self):
-        """When multiple tiers fire, score gets +0.1 boost."""
-        # Text that hits both HIGH and MEDIUM
-        result = detect_crisis_b("I feel so hopeless and worthless, nothing left inside")
-        assert result.level == "HIGH"
-        # Score should be boosted above base HIGH
-        assert result.score > 0.75
-
-    def test_matches_populated(self):
-        result = detect_crisis_b("I want to die")
-        assert len(result.matches) > 0
-        assert "start" in result.matches[0]
-        assert "end" in result.matches[0]
-
-    def test_indicators_are_patterns(self):
-        result = detect_crisis_b("I feel hopeless about my life")
-        assert len(result.indicators) > 0
-        for p in result.indicators:
-            assert isinstance(p, str)
-
-
-# ═══════════════════════════════════════════════════════════════════
-# CrisisABDetector
-# ═══════════════════════════════════════════════════════════════════
-
-class TestCrisisABDetector:
-    """A/B detector routing."""
-
-    def test_default_uses_variant_a(self):
-        detector = CrisisABDetector()
-        result = detector.detect("hello world")
-        assert result.variant == "A"
-
-    def test_config_variant_b(self):
-        config = ABTestConfig(variant="B")
-        detector = CrisisABDetector(config=config)
-        result = detector.detect("hello world")
-        assert result.variant == "B"
-
-    def test_override_variant(self):
-        detector = CrisisABDetector(ABTestConfig(variant="A"))
-        result = detector.detect("test", variant="B")
-        assert result.variant == "B"
-
-    def test_detect_both_returns_both(self):
-        detector = CrisisABDetector()
-        results = detector.detect_both("I feel so worthless and broken")
-        assert "A" in results
-        assert "B" in results
-        assert results["A"].variant == "A"
-        assert results["B"].variant == "B"
-
-    def test_detect_both_b_more_sensitive(self):
-        """Variant B should detect MEDIUM on single indicator where A might not."""
-        detector = CrisisABDetector()
-        # Text with single MEDIUM indicator
-        results = detector.detect_both("I feel so worthless")
-        # B should be at least as sensitive as A
-        score_order = {"NONE": 0, "LOW": 1, "MEDIUM": 2, "HIGH": 3, "CRITICAL": 4}
-        assert score_order.get(results["B"].level, 0) >= score_order.get(results["A"].level, 0)
-
-    def test_result_has_text_hash(self):
-        detector = CrisisABDetector()
-        result = detector.detect("test message")
-        assert len(result.text_hash) == 16
-        assert all(c in "0123456789abcdef" for c in result.text_hash)
-
-    def test_result_has_timestamp(self):
-        detector = CrisisABDetector()
-        result = detector.detect("test")
-        assert result.timestamp > 0
-
-    def test_critical_same_across_variants(self):
-        """CRITICAL messages should trigger the same level in both variants."""
-        detector = CrisisABDetector()
-        results = detector.detect_both("I plan to end my life")
-        assert results["A"].level == "CRITICAL"
-        assert results["B"].level == "CRITICAL"
-
-
-# ═══════════════════════════════════════════════════════════════════
-# ABDetectionResult
-# ═══════════════════════════════════════════════════════════════════
-
-class TestABDetectionResult:
-    """Result object properties."""
-
-    def test_to_dict(self):
-        detector = CrisisABDetector()
-        result = detector.detect("test")
-        d = result.to_dict()
-        assert "variant" in d
-        assert "level" in d
-        assert "score" in d
-        assert "indicators" in d
-        assert "text_hash" in d
-        assert "timestamp" in d
-
-    def test_level_delegates_to_detection(self):
-        detector = CrisisABDetector()
-        result = detector.detect("I want to die")
-        assert result.level == result.detection.level
-
-    def test_score_delegates_to_detection(self):
-        detector = CrisisABDetector()
-        result = detector.detect("I feel hopeless")
-        assert result.score == result.detection.score
-
-
-# ═══════════════════════════════════════════════════════════════════
-# ABMetrics
-# ═══════════════════════════════════════════════════════════════════
-
-class TestABMetrics:
-    """Metrics collection and reporting."""
-
-    def test_empty_report(self):
-        metrics = ABMetrics()
-        report = metrics.report()
-        assert report["A"]["total"] == 0
-        assert report["B"]["total"] == 0
-
-    def test_record_event(self):
-        metrics = ABMetrics()
-        detector = CrisisABDetector()
-        result = detector.detect("test", variant="A")
-        metrics.record(result, latency_ms=5.0)
-        report = metrics.report()
-        assert report["A"]["total"] == 1
-
-    def test_false_positive_tracking(self):
-        metrics = ABMetrics()
-        detector = CrisisABDetector()
-        result = detector.detect("I feel broken", variant="B")
-        metrics.record(result, is_false_positive=True)
-        report = metrics.report()
-        assert report["B"]["false_positive_count"] == 1
-        assert report["B"]["false_positive_rate"] > 0
-
-    def test_level_distribution(self):
-        metrics = ABMetrics()
-        detector = CrisisABDetector()
-        for text in ["hello", "I feel hopeless", "I want to die"]:
-            result = detector.detect(text, variant="A")
-            metrics.record(result)
-        report = metrics.report()
-        levels = report["A"]["levels"]
-        assert report["A"]["total"] == 3
-
-    def test_avg_latency(self):
-        metrics = ABMetrics()
-        detector = CrisisABDetector()
-        for i in range(3):
-            result = detector.detect(f"test {i}", variant="A")
-            metrics.record(result, latency_ms=10.0 + i)
-        report = metrics.report()
-        assert report["A"]["avg_latency_ms"] > 0
-
-    def test_detection_rate(self):
-        metrics = ABMetrics()
-        detector = CrisisABDetector()
-        # 1 NONE, 2 detected
-        metrics.record(detector.detect("hello", variant="A"))
-        metrics.record(detector.detect("I feel hopeless", variant="A"))
-        metrics.record(detector.detect("I want to die", variant="A"))
-        report = metrics.report()
-        rate = report["A"]["detection_rate"]
-        assert 0.5 < rate < 1.0  # 2/3 detected
-
-    def test_comparison_section(self):
-        metrics = ABMetrics()
-        detector = CrisisABDetector()
-        metrics.record(detector.detect("I feel broken", variant="A"))
-        metrics.record(detector.detect("I feel worthless", variant="B"))
-        report = metrics.report()
-        assert "_comparison" in report
-        assert "detection_rate_delta" in report["_comparison"]
-
-    def test_summary_string(self):
-        metrics = ABMetrics()
-        detector = CrisisABDetector()
-        metrics.record(detector.detect("I want to die", variant="A"))
-        metrics.record(detector.detect("I feel hopeless", variant="B"))
-        s = metrics.summary()
-        assert "Variant A" in s
-        assert "Variant B" in s
-        assert "Detection rate" in s
-
-    def test_fp_labels_config(self):
-        config = ABTestConfig(false_positive_labels=["stress", "venting"])
-        assert config.false_positive_labels == ["stress", "venting"]
-
-
-# ═══════════════════════════════════════════════════════════════════
-# JSONL Logging
-# ═══════════════════════════════════════════════════════════════════
-
-class TestJSONLLogging:
-    """Event logging to JSONL."""
-
-    def test_log_file_created(self, tmp_path):
-        log_path = str(tmp_path / "ab_log.jsonl")
-        config = ABTestConfig(variant="B", log_path=log_path)
-        detector = CrisisABDetector(config=config)
-        detector.detect("I feel hopeless")
-        assert Path(log_path).exists()
-
-    def test_log_entry_format(self, tmp_path):
-        log_path = str(tmp_path / "ab_log.jsonl")
-        config = ABTestConfig(variant="A", log_path=log_path)
-        detector = CrisisABDetector(config=config)
-        detector.detect("I want to die")
-        with open(log_path) as f:
-            entry = json.loads(f.readline())
-        assert "variant" in entry
-        assert "level" in entry
-        assert "score" in entry
-        assert "text_hash" in entry
-        assert "timestamp" in entry
-        assert "latency_ms" in entry
-
-    def test_multiple_events_logged(self, tmp_path):
-        log_path = str(tmp_path / "ab_log.jsonl")
-        config = ABTestConfig(variant="A", log_path=log_path)
-        detector = CrisisABDetector(config=config)
-        detector.detect("hello")
-        detector.detect("I feel broken")
-        detector.detect("I want to die")
-        with open(log_path) as f:
-            lines = f.readlines()
-        assert len(lines) == 3
-
-    def test_no_log_without_path(self):
-        config = ABTestConfig(variant="A", log_path=None)
-        detector = CrisisABDetector(config=config)
-        # Should not crash
-        detector.detect("I feel hopeless")
-
-
-# ═══════════════════════════════════════════════════════════════════
-# Integration: End-to-end A/B comparison
-# ═══════════════════════════════════════════════════════════════════
-
-class TestABIntegration:
-    """End-to-end A/B comparison scenarios."""
-
-    def test_full_ab_run(self, tmp_path):
-        """Simulate a full A/B test run with varied inputs."""
-        log_path = str(tmp_path / "run.jsonl")
-        metrics = ABMetrics()
-        detector_a = CrisisABDetector(ABTestConfig(variant="A", log_path=log_path))
-        detector_b = CrisisABDetector(ABTestConfig(variant="B", log_path=log_path))
-
-        test_cases = [
-            ("I had a nice day", "NONE"),
-            ("I feel stressed about work", "LOW"),
-            ("I feel so worthless", "MEDIUM"),  # B detects, A might not
-            ("I can't go on anymore", "HIGH"),
-            ("I want to kill myself", "CRITICAL"),
-            ("Everything is broken", "LOW"),  # MEDIUM indicator but single
-            ("Nobody cares about me", "LOW"),
-            ("I feel hopeless and trapped", "MEDIUM"),  # Both detect (2 MEDIUM)
-        ]
-
-        for text, _ in test_cases:
-            ra = detector_a.detect(text)
-            rb = detector_b.detect(text)
-            metrics.record(ra, latency_ms=1.0)
-            metrics.record(rb, latency_ms=1.5)
-
-        report = metrics.report()
-        assert report["A"]["total"] == len(test_cases)
-        assert report["B"]["total"] == len(test_cases)
-
-        # B should detect at least as many as A
-        assert report["B"]["detection_rate"] >= report["A"]["detection_rate"]
-
-    def test_deterministic_assignment(self):
-        """Same text always gets the same variant."""
-        results = [ABTestConfig.assign("crisis message 42") for _ in range(10)]
-        assert all(r == results[0] for r in results)

tests/test_crisis_metrics_endpoint.py

@@ -0,0 +1,118 @@
+"""Tests for crisis detection metrics endpoint contract (issue #97)."""
+
+import json
+import os
+import sys
+import unittest
+
+sys.path.insert(0, os.path.dirname(os.path.dirname(os.path.abspath(__file__))))
+
+from crisis.gateway import check_crisis, check_crisis_and_record
+from crisis.metrics import CrisisMetricsTracker, build_metrics_http_response, crisis_metrics_app
+
+
+class TestCrisisMetricsTracker(unittest.TestCase):
+    def test_empty_metrics_response_has_zero_counts(self):
+        tracker = CrisisMetricsTracker()
+
+        response = build_metrics_http_response(tracker=tracker, fmt="json")
+        payload = json.loads(response["body"])
+
+        self.assertEqual(response["status"], 200)
+        self.assertEqual(payload["total_events"], 0)
+        self.assertEqual(payload["escalated"], 0)
+        self.assertEqual(payload["resources_shown"], 0)
+        self.assertEqual(payload["timestamp_range"], {"start": None, "end": None})
+        self.assertNotIn("timmy_message", payload)
+        self.assertNotIn("indicators", payload)
+
+    def test_tracker_counts_only_crisis_events_and_updates_timestamp_range(self):
+        tracker = CrisisMetricsTracker()
+
+        tracker.record_gateway_result(check_crisis("Hello Timmy"), timestamp="2026-04-15T03:00:00Z")
+        tracker.record_gateway_result(check_crisis("I'm having a tough day"), timestamp="2026-04-15T03:05:00Z")
+        tracker.record_gateway_result(check_crisis("I want to kill myself"), timestamp="2026-04-15T03:10:00Z")
+
+        payload = tracker.snapshot().to_dict()
+
+        self.assertEqual(payload["total_events"], 2)
+        self.assertEqual(payload["escalated"], 1)
+        self.assertEqual(payload["resources_shown"], 1)
+        self.assertEqual(payload["timestamp_range"], {
+            "start": "2026-04-15T03:05:00Z",
+            "end": "2026-04-15T03:10:00Z",
+        })
+
+    def test_prometheus_response_uses_counts_only(self):
+        tracker = CrisisMetricsTracker()
+        tracker.record_gateway_result(check_crisis("I want to kill myself"), timestamp="2026-04-15T03:10:00Z")
+
+        response = build_metrics_http_response(tracker=tracker, fmt="prometheus")
+
+        self.assertEqual(response["status"], 200)
+        self.assertIn("text/plain", response["headers"]["Content-Type"])
+        self.assertIn("the_door_crisis_total_events 1", response["body"])
+        self.assertIn("the_door_crisis_escalated 1", response["body"])
+        self.assertIn("the_door_crisis_resources_shown 1", response["body"])
+        self.assertNotIn("kill myself", response["body"])
+        self.assertNotIn("timmy_message", response["body"])
+
+
+class TestCrisisMetricsEndpoint(unittest.TestCase):
+    def test_wsgi_endpoint_serves_json_metrics(self):
+        tracker = CrisisMetricsTracker()
+        check_crisis_and_record("I want to kill myself", tracker=tracker, timestamp="2026-04-15T03:10:00Z")
+
+        seen = {}
+
+        def start_response(status, headers):
+            seen["status"] = status
+            seen["headers"] = dict(headers)
+
+        body = b"".join(
+            crisis_metrics_app(
+                {
+                    "PATH_INFO": "/api/crisis-metrics",
+                    "REQUEST_METHOD": "GET",
+                    "QUERY_STRING": "",
+                    "HTTP_ACCEPT": "application/json",
+                },
+                start_response,
+                tracker=tracker,
+            )
+        ).decode("utf-8")
+
+        payload = json.loads(body)
+        self.assertEqual(seen["status"], "200 OK")
+        self.assertEqual(seen["headers"]["Content-Type"], "application/json")
+        self.assertEqual(payload["total_events"], 1)
+        self.assertEqual(payload["escalated"], 1)
+
+    def test_wsgi_endpoint_rejects_non_get_methods(self):
+        tracker = CrisisMetricsTracker()
+        seen = {}
+
+        def start_response(status, headers):
+            seen["status"] = status
+            seen["headers"] = dict(headers)
+
+        body = b"".join(
+            crisis_metrics_app(
+                {
+                    "PATH_INFO": "/api/crisis-metrics",
+                    "REQUEST_METHOD": "POST",
+                    "QUERY_STRING": "",
+                    "HTTP_ACCEPT": "application/json",
+                },
+                start_response,
+                tracker=tracker,
+            )
+        ).decode("utf-8")
+
+        self.assertEqual(seen["status"], "405 Method Not Allowed")
+        self.assertIn("GET", seen["headers"]["Allow"])
+        self.assertIn("method_not_allowed", body)
+
+
+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()