feat: build crisis synthesizer (#36)

GENOME.md

@@ -1,124 +0,0 @@
-# GENOME.md — the-door
-
-> Codebase analysis generated 2026-04-13. Crisis intervention web app — a door that's always open.
-
-## Project Overview
-
-the-door is a single-URL crisis intervention web app. A man at 3am can talk to Timmy. No login. No signup. No tracking. Just a door that's always open.
-
-**Mission**: Stand between a broken man and a machine that would tell him to die.
-
-48 files. Static HTML frontend (<25KB, works on 3G). Python crisis detection backend. Safety-critical — a broken deployment could prevent someone from reaching the 988 Lifeline.
-
-## Architecture
-
-```
-Browser → nginx (SSL) → index.html → /api/* proxy → Hermes Gateway
-                                                        ↓
-                                                  crisis/detect.py
-                                                        ↓
-                                                  988 Lifeline overlay
-```
-
-## Entry Points
-
-- **index.html** — The entire frontend. One file. <25KB. Works on 3G.
-- **system-prompt.txt** — Crisis-aware system prompt for the AI.
-- **deploy/deploy.sh** — Deployment script for VPS.
-- **deploy/playbook.yml** — Ansible playbook for deployment.
-- **crisis/detect.py** — Core crisis detection module (canonical).
-- **crisis_detector.py** — Legacy class API wrapper around detect.py.
-- **crisis_responder.py** — Response formatting for crisis levels.
-
-## Data Flow
-
-```
-User message → browser
-    ↓
-index.html → client-side crisis keyword scan
-    ↓
-/api/chat → Hermes Gateway
-    ↓
-system-prompt.txt → injected into AI system prompt
-    ↓
-crisis/detect.py → 5-tier classification (NONE/LOW/MEDIUM/HIGH/CRITICAL)
-    ↓
-crisis/response.py → appropriate response with 988 Lifeline info
-    ↓
-Response → browser → crisis overlay if HIGH/CRITICAL
-```
-
-## Key Abstractions
-
-### Crisis Detection (crisis/detect.py)
-Canonical detection module. Regex-based keyword matching across 4 tiers:
-- CRITICAL: immediate self-harm risk (single match triggers)
-- HIGH: strong despair signals (single match triggers)
-- MEDIUM: distress signals (requires 2+ indicators)
-- LOW: emotional difficulty (single match)
-
-Design principles:
-- Never computes the value of a human life
-- Never suggests death is a solution
-- Always errs on side of higher risk
-
-### Crisis Profiles (crisis/profiles.py)
-Compassion profiles that shape AI response tone based on crisis level.
-
-### Session Tracker (crisis/session_tracker.py)
-Tracks crisis interactions across sessions. Persistent state for ongoing support.
-
-### Gateway (crisis/gateway.py)
-HTTP gateway for crisis detection API. Endpoints for scanning text and getting responses.
-
-### Offline Fallback (crisis-offline.html, sw.js)
-Service worker caches crisis resources. When network is down, users still see 988 Lifeline info and crisis resources.
-
-## File Types
-
-| Type | Count | Purpose |
-|------|-------|---------|
-| .py | 16 | Crisis detection, response, tests |
-| .html | 4 | Frontend, offline fallback, tests |
-| .yml | 2 | CI workflows |
-| .sh | 2 | Health check, service restart |
-| .md | 5 | Documentation, safety audits |
-
-## Test Coverage
-
-### Existing Tests
-- test_crisis_overlay_focus_trap.py — Accessibility: focus trap in crisis overlay
-- test_dying_detection_deprecation.py — Legacy API deprecation
-- test_false_positive_fixes.py — Crisis detection false positive resistance
-- test_service_worker_offline.py — Offline fallback verification
-- test_session_tracker.py — Session tracking persistence
-- crisis/test_rescue.py — Rescue flow testing
-- crisis/tests.py — Core crisis detection tests
-
-### Coverage Gaps
-- No integration tests for full browser → API → response → overlay flow
-- No tests for system-prompt.txt injection into AI system prompt
-- No load tests (what happens at 1000 concurrent crisis users?)
-- No tests for deploy.sh idempotency
-
-### Critical paths that need tests:
-1. **Full crisis flow**: user message → detection → 988 overlay → response
-2. **Offline fallback**: network down → service worker → cached crisis resources
-3. **Deploy safety**: deploy.sh doesn't break running service
-
-## Security Considerations
-
-- **SAFETY-CRITICAL**: the-door serves users in crisis. Broken deployment could prevent someone from reaching 988 Lifeline.
-- **PR safety**: the-door PRs NEVER auto-merge. Requires-human label on all PRs. (fleet-ops#183)
-- **No authentication by design**: no login, no signup, no tracking. Privacy is a safety feature.
-- **Rate limiting**: deploy/rate-limit.conf prevents abuse while allowing crisis access.
-- **Offline resilience**: service worker ensures crisis resources available even without network.
-- **System prompt is safety boundary**: system-prompt.txt defines the AI's crisis behavior. Changes require human review.
-
-## Design Decisions
-
-- **Single HTML file**: no build step, no framework, no dependencies. Works on 3G. Loads instantly.
-- **Client-side detection first**: browser scans for crisis keywords before sending to server. Instant response for critical cases.
-- **Server-side detection second**: crisis/detect.py provides deeper analysis with tiered classification.
-- **Offline-first for crisis**: service worker caches crisis resources. Network failure doesn't block access to help.
-- **No tracking**: privacy protects vulnerable users. No analytics, no cookies, no login.

evolution/crisis_synthesizer.py

@@ -1 +1,195 @@
-...
+"""Crisis synthesizer — learn from anonymized crisis interactions.
+
+This is deliberately simple and privacy-preserving. It does not train a model or
+modify detection rules automatically. It only logs metadata, summarizes patterns,
+and suggests human-reviewed keyword weight adjustments.
+"""
+
+from __future__ import annotations
+
+import argparse
+import json
+import time
+from collections import Counter, defaultdict
+from pathlib import Path
+from typing import Iterable
+
+DEFAULT_LOG_PATH = Path.home() / ".the-door" / "crisis-interactions.jsonl"
+LEVELS = ("NONE", "LOW", "MEDIUM", "HIGH", "CRITICAL")
+
+
+def build_interaction_event(
+    level: str,
+    indicators: list[str],
+    response_given: str,
+    continued_conversation: bool,
+    false_positive: bool,
+    *,
+    now: float | None = None,
+) -> dict:
+    return {
+        "timestamp": float(time.time() if now is None else now),
+        "level": level,
+        "indicators": list(indicators),
+        "indicator_count": len(indicators),
+        "response_given": response_given,
+        "continued_conversation": bool(continued_conversation),
+        "false_positive": bool(false_positive),
+    }
+
+
+def append_interaction_event(
+    log_path: str | Path,
+    *,
+    level: str,
+    indicators: list[str],
+    response_given: str,
+    continued_conversation: bool,
+    false_positive: bool,
+    now: float | None = None,
+) -> dict:
+    event = build_interaction_event(
+        level,
+        indicators,
+        response_given,
+        continued_conversation,
+        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_interaction_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 summarize_keywords(events: Iterable[dict]) -> list[dict]:
+    counts: Counter[str] = Counter()
+    for event in events:
+        counts.update(event.get("indicators", []))
+    return [{"keyword": keyword, "count": count} for keyword, count in counts.most_common(10)]
+
+
+def suggest_keyword_adjustments(events: Iterable[dict], *, min_observations: int = 5) -> list[dict]:
+    stats: dict[str, dict[str, int]] = defaultdict(lambda: {
+        "observations": 0,
+        "true_positive_count": 0,
+        "false_positive_count": 0,
+        "continued_conversation_count": 0,
+    })
+
+    for event in events:
+        for keyword in event.get("indicators", []):
+            bucket = stats[keyword]
+            bucket["observations"] += 1
+            if event.get("false_positive"):
+                bucket["false_positive_count"] += 1
+            else:
+                bucket["true_positive_count"] += 1
+            if event.get("continued_conversation"):
+                bucket["continued_conversation_count"] += 1
+
+    suggestions = []
+    for keyword, bucket in sorted(stats.items()):
+        if bucket["observations"] < min_observations:
+            continue
+        fp = bucket["false_positive_count"]
+        tp = bucket["true_positive_count"]
+        if fp >= min_observations and tp == 0:
+            adjustment = "lower_weight"
+            rationale = "Observed only false positives across the sample window."
+        elif tp >= min_observations and fp == 0:
+            adjustment = "raise_weight"
+            rationale = "Observed repeated genuine crises with no false positives."
+        else:
+            adjustment = "observe"
+            rationale = "Mixed evidence; keep monitoring before changing weights."
+        suggestions.append(
+            {
+                "keyword": keyword,
+                **bucket,
+                "suggested_adjustment": adjustment,
+                "rationale": rationale,
+            }
+        )
+    return suggestions
+
+
+def build_weekly_report(
+    events: Iterable[dict],
+    *,
+    now: float | None = None,
+    window_days: int = 7,
+    min_observations: int = 3,
+) -> 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}
+    detected_events = []
+    continued_after_intervention = 0
+    for event in filtered:
+        level = event.get("level", "NONE")
+        detections_per_level[level] = detections_per_level.get(level, 0) + 1
+        if level != "NONE":
+            detected_events.append(event)
+            if event.get("continued_conversation"):
+                continued_after_intervention += 1
+
+    false_positive_count = sum(1 for event in detected_events if event.get("false_positive"))
+    false_positive_estimate = false_positive_count / len(detected_events) if detected_events else 0.0
+
+    return {
+        "window_days": window_days,
+        "total_events": len(filtered),
+        "detections_per_level": detections_per_level,
+        "most_common_keywords": summarize_keywords(filtered),
+        "false_positive_estimate": false_positive_estimate,
+        "continued_after_intervention": continued_after_intervention,
+        "keyword_weight_suggestions": suggest_keyword_adjustments(filtered, min_observations=min_observations),
+    }
+
+
+def render_weekly_report(summary: dict) -> str:
+    return json.dumps(summary, indent=2)
+
+
+def write_weekly_report(output_path: str | Path, summary: dict) -> Path:
+    path = Path(output_path)
+    path.parent.mkdir(parents=True, exist_ok=True)
+    path.write_text(render_weekly_report(summary) + "\n", encoding="utf-8")
+    return path
+
+
+def main(argv: list[str] | None = None) -> int:
+    parser = argparse.ArgumentParser(description="Summarize anonymized crisis interactions")
+    parser.add_argument("--log-path", default=str(DEFAULT_LOG_PATH), help="JSONL crisis interaction log")
+    parser.add_argument("--days", type=int, default=7, help="Lookback window in days")
+    parser.add_argument("--min-observations", type=int, default=3, help="Minimum observations before suggesting keyword adjustments")
+    parser.add_argument("--output", help="Optional file to write the weekly report JSON")
+    args = parser.parse_args(argv)
+
+    events = load_interaction_events(args.log_path)
+    summary = build_weekly_report(events, window_days=args.days, min_observations=args.min_observations)
+    rendered = render_weekly_report(summary)
+    print(rendered)
+    if args.output:
+        write_weekly_report(args.output, summary)
+    return 0
+
+
+if __name__ == "__main__":
+    raise SystemExit(main())

tests/test_crisis_synthesizer.py

@@ -0,0 +1,111 @@
+"""Tests for evolution/crisis_synthesizer.py (issue #36)."""
+
+from __future__ import annotations
+
+import importlib.util
+import json
+import pathlib
+import sys
+import tempfile
+import unittest
+
+ROOT = pathlib.Path(__file__).resolve().parents[1]
+SCRIPT = ROOT / 'evolution' / 'crisis_synthesizer.py'
+
+spec = importlib.util.spec_from_file_location('crisis_synthesizer', str(SCRIPT))
+mod = importlib.util.module_from_spec(spec)
+sys.modules['crisis_synthesizer'] = mod
+spec.loader.exec_module(mod)
+
+
+class TestCrisisSynthesizerEvent(unittest.TestCase):
+    def test_build_interaction_event_is_privacy_preserving(self):
+        event = mod.build_interaction_event(
+            level='CRITICAL',
+            indicators=['want_to_die', 'no_way_out'],
+            response_given='guardian',
+            continued_conversation=True,
+            false_positive=False,
+            now=1700000000,
+        )
+        self.assertEqual(event['timestamp'], 1700000000)
+        self.assertEqual(event['level'], 'CRITICAL')
+        self.assertEqual(event['response_given'], 'guardian')
+        self.assertTrue(event['continued_conversation'])
+        self.assertFalse(event['false_positive'])
+        self.assertEqual(event['indicators'], ['want_to_die', 'no_way_out'])
+        for forbidden in ['text', 'message', 'content', 'ip', 'session_id', 'user_id']:
+            self.assertNotIn(forbidden, event)
+
+
+class TestCrisisSynthesizerStorage(unittest.TestCase):
+    def test_append_and_load_events_round_trip(self):
+        with tempfile.TemporaryDirectory() as tmp:
+            log_path = pathlib.Path(tmp) / 'crisis-events.jsonl'
+            mod.append_interaction_event(
+                log_path,
+                level='HIGH',
+                indicators=['hopeless'],
+                response_given='companion',
+                continued_conversation=False,
+                false_positive=True,
+                now=1700000100,
+            )
+            events = mod.load_interaction_events(log_path)
+            self.assertEqual(len(events), 1)
+            self.assertEqual(events[0]['level'], 'HIGH')
+            self.assertEqual(events[0]['indicators'], ['hopeless'])
+
+
+class TestCrisisSynthesizerSummary(unittest.TestCase):
+    def test_weekly_report_contains_required_metrics(self):
+        events = [
+            mod.build_interaction_event('CRITICAL', ['want_to_die'], 'guardian', True, False, now=1700000000),
+            mod.build_interaction_event('HIGH', ['hopeless'], 'companion', False, True, now=1700000100),
+            mod.build_interaction_event('LOW', ['rough_day'], 'friend', False, False, now=1700000200),
+            mod.build_interaction_event('CRITICAL', ['want_to_die'], 'guardian', False, False, now=1700000300),
+            mod.build_interaction_event('NONE', [], 'friend', False, False, now=1700000400),
+        ]
+        summary = mod.build_weekly_report(events, now=1700000500, window_days=7)
+        self.assertEqual(summary['detections_per_level']['CRITICAL'], 2)
+        self.assertEqual(summary['detections_per_level']['HIGH'], 1)
+        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'], 0.25)
+        self.assertEqual(summary['most_common_keywords'][0]['keyword'], 'want_to_die')
+        self.assertEqual(summary['most_common_keywords'][0]['count'], 2)
+
+
+class TestCrisisSynthesizerSuggestions(unittest.TestCase):
+    def test_suggests_weight_adjustments_from_interactions(self):
+        events = []
+        for ts in range(3):
+            events.append(mod.build_interaction_event('CRITICAL', ['want_to_die'], 'guardian', True, False, now=1700000000 + ts))
+        for ts in range(3):
+            events.append(mod.build_interaction_event('LOW', ['rough_day'], 'friend', False, True, now=1700000100 + ts))
+        suggestions = mod.suggest_keyword_adjustments(events, min_observations=3)
+        by_keyword = {s['keyword']: s for s in suggestions}
+        self.assertEqual(by_keyword['want_to_die']['suggested_adjustment'], 'raise_weight')
+        self.assertEqual(by_keyword['rough_day']['suggested_adjustment'], 'lower_weight')
+
+
+class TestCrisisSynthesizerRendering(unittest.TestCase):
+    def test_render_weekly_report_outputs_json(self):
+        summary = {
+            'detections_per_level': {'NONE': 0, 'LOW': 1, 'MEDIUM': 0, 'HIGH': 0, 'CRITICAL': 0},
+            'most_common_keywords': [{'keyword': 'rough_day', 'count': 1}],
+            'false_positive_estimate': 0.0,
+            'continued_after_intervention': 0,
+            'keyword_weight_suggestions': [],
+            'window_days': 7,
+            'total_events': 1,
+        }
+        rendered = mod.render_weekly_report(summary)
+        parsed = json.loads(rendered)
+        self.assertEqual(parsed['window_days'], 7)
+        self.assertEqual(parsed['most_common_keywords'][0]['keyword'], 'rough_day')
+
+
+if __name__ == '__main__':
+    unittest.main()