Compare commits

..

2 Commits

Author SHA1 Message Date
Timmy Burn
9f3e01cd8a feat: add operator augmentation sidebar for #135
All checks were successful
Sanity Checks / sanity-test (pull_request) Successful in 4s
Smoke Test / smoke (pull_request) Successful in 9s
2026-04-20 19:41:23 -04:00
Timmy Burn
e146a4ea39 test: define counselor augmentation for #135 2026-04-20 19:26:42 -04:00
8 changed files with 498 additions and 254 deletions

142
augmentation.py Normal file
View File

@@ -0,0 +1,142 @@
"""Local-only counselor augmentation helpers for the-door."""
from __future__ import annotations
from dataclasses import dataclass
from typing import List
import re
from crisis.detect import detect_crisis
@dataclass(frozen=True)
class SignalGuide:
label: str
patterns: List[str]
talking_point: str
deescalation: str
follow_up: str
@dataclass
class CounselorAugmentation:
risk_level: str
risk_score: int
signals: List[str]
suggested_talking_points: List[str]
deescalation_techniques: List[str]
follow_up_prompt: str
operator_notice: str
local_only: bool = True
advisory_only: bool = True
SIGNAL_GUIDES: List[SignalGuide] = [
SignalGuide(
label="Explicit self-harm intent",
patterns=[
r"\bkill\s*(my)?self\b",
r"\bend\s*my\s*life\b",
r"\bwrote\s+a\s+suicide\s*(?:note|letter)\b",
r"\bgoing\s+to\s+(?:kill\s+myself|die)\b",
],
talking_point="Ask directly whether they are safe right now and keep the next question concrete.",
deescalation="Move to immediate safety: ask about means, people nearby, and whether they can call or text 988 now.",
follow_up="You said you're ready to die. Are you alone right now, and can you tell me what is within reach?",
),
SignalGuide(
label="Hopelessness / collapse",
patterns=[
r"\bhopeless\b",
r"\bcan'?t\s+go\s+on\b",
r"\bno\s+future\b",
r"\bnothing\s+left\b",
],
talking_point="Reflect the hopelessness plainly before offering options. Avoid arguing with the feeling.",
deescalation="Narrow the time horizon: focus on the next ten minutes, one breath, one call, one person.",
follow_up="You said things feel hopeless. What feels most dangerous about the next hour?",
),
SignalGuide(
label="Isolation / burden",
patterns=[
r"\bnobody\s+cares\b",
r"\bbetter\s+off\s+without\s+me\b",
r"\balone\b",
r"\bburden\b",
],
talking_point="Counter isolation with immediacy: name one real person or service they can contact now.",
deescalation="Invite a tiny reconnection step: text one safe person, unlock the door, move closer to others, or stay in the chat.",
follow_up="You said you feel alone. Who is the safest real person we could bring into this moment with you?",
),
SignalGuide(
label="Overwhelm / panic",
patterns=[
r"\bdesperate\b",
r"\boverwhelm(?:ed|ing)\b",
r"\btrapped\b",
r"\bpanic\b",
],
talking_point="Offer one regulating action at a time instead of a list. Slow the pace of the chat.",
deescalation="Ground in the room: feet on the floor, name five visible objects, one sip of water, one slow exhale.",
follow_up="You said this feels overwhelming. What is the smallest thing in the room you can touch right now?",
),
]
class CounselorAugmentationEngine:
BASE_SCORES = {
"NONE": 5,
"LOW": 25,
"MEDIUM": 55,
"HIGH": 75,
"CRITICAL": 95,
}
def _matched_guides(self, text: str) -> List[SignalGuide]:
lowered = text.lower()
matched: List[SignalGuide] = []
for guide in SIGNAL_GUIDES:
if any(re.search(pattern, lowered) for pattern in guide.patterns):
matched.append(guide)
return matched
def build_augmented_guidance(self, text: str, assistant_text: str = "") -> CounselorAugmentation:
detection = detect_crisis(text)
guides = self._matched_guides(text)
risk_level = detection.level
signals = [guide.label for guide in guides]
if risk_level == "CRITICAL" and "Explicit self-harm intent" not in signals:
signals.insert(0, "Explicit self-harm intent")
risk_score = self.BASE_SCORES.get(risk_level, 5) + min(len(signals) * 5, 10)
if risk_level == "CRITICAL":
risk_score = max(risk_score, 95)
talking_points = [guide.talking_point for guide in guides] or [
"Keep the response advisory and grounded in immediate safety, not abstract reassurance."
]
deescalation = [guide.deescalation for guide in guides] or [
"Use short sentences, slow the conversation, and invite one concrete grounding step."
]
quote = text.strip().replace("\n", " ")[:120]
follow_up = (guides[0].follow_up if guides else "What feels most dangerous or heavy for you right now?")
follow_up_prompt = f'You said "{quote}". Consider following up with: {follow_up}'
if assistant_text and "988" not in assistant_text and risk_level in {"HIGH", "CRITICAL"}:
talking_points.append("Surface 988 or Crisis Text Line explicitly if the assistant has not already done so.")
return CounselorAugmentation(
risk_level=risk_level,
risk_score=min(risk_score, 100),
signals=signals,
suggested_talking_points=talking_points,
deescalation_techniques=deescalation,
follow_up_prompt=follow_up_prompt,
operator_notice="Local-only advisory. This never replaces human judgment.",
)
def build_augmented_guidance(text: str, assistant_text: str = "") -> CounselorAugmentation:
return CounselorAugmentationEngine().build_augmented_guidance(text, assistant_text=assistant_text)

View File

@@ -8,7 +8,6 @@ from .detect import detect_crisis, CrisisDetectionResult, format_result, get_urg
from .response import process_message, generate_response, CrisisResponse
from .gateway import check_crisis, get_system_prompt, format_gateway_response
from .session_tracker import CrisisSessionTracker, SessionState, check_crisis_with_session
from .ab_testing import ABTestCrisisDetector, VariantRecord
__all__ = [
"detect_crisis",
@@ -24,6 +23,4 @@ __all__ = [
"CrisisSessionTracker",
"SessionState",
"check_crisis_with_session",
"ABTestCrisisDetector",
"VariantRecord",
]

View File

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

View File

@@ -241,6 +241,105 @@ html, body {
opacity: 0.5;
}
/* ===== OPERATOR AUGMENTATION SIDEBAR ===== */
#augmentation-toggle {
margin: 10px 16px 0;
padding: 8px 12px;
border-radius: 999px;
border: 1px solid #5b6b7a;
background: #11161d;
color: #b9c7d5;
font-size: 0.9rem;
cursor: pointer;
}
#augmentation-toggle.active {
border-color: #b388ff;
color: #e2d4ff;
background: #1a1324;
}
#augmentation-sidebar {
position: fixed;
top: 90px;
right: 16px;
width: 320px;
max-height: calc(100vh - 120px);
overflow-y: auto;
background: #11161d;
border: 1px solid #30363d;
border-left: 3px solid #b388ff;
border-radius: 8px;
padding: 14px;
box-shadow: 0 12px 32px rgba(0,0,0,0.35);
display: none;
z-index: 70;
}
#augmentation-sidebar.visible {
display: block;
}
#augmentation-sidebar .augmentation-heading {
color: #d2b8ff;
font-size: 0.78rem;
letter-spacing: 0.08em;
margin-bottom: 10px;
}
#augmentation-risk-score {
color: #fff;
font-size: 1rem;
font-weight: 700;
margin-bottom: 10px;
}
#augmentation-sidebar .augmentation-section {
margin-top: 10px;
}
#augmentation-sidebar .augmentation-section h3 {
color: #c9d1d9;
font-size: 0.78rem;
margin: 0 0 6px;
text-transform: uppercase;
letter-spacing: 0.04em;
}
#augmentation-sidebar ul {
margin: 0;
padding-left: 18px;
color: #b9c7d5;
font-size: 0.9rem;
line-height: 1.45;
}
#augmentation-follow-up,
#augmentation-notice {
color: #b9c7d5;
font-size: 0.9rem;
line-height: 1.45;
margin: 0;
}
#augmentation-notice {
color: #8b949e;
margin-top: 12px;
border-top: 1px solid #21262d;
padding-top: 10px;
}
@media (max-width: 980px) {
#augmentation-sidebar {
left: 16px;
right: 16px;
width: auto;
top: auto;
bottom: 82px;
max-height: 40vh;
}
}
/* ===== CHAT AREA ===== */
#chat-area {
flex: 1;
@@ -649,6 +748,29 @@ html, body {
</div>
</div>
<button id="augmentation-toggle" type="button" aria-pressed="false" aria-controls="augmentation-sidebar">Operator assist: off</button>
<aside id="augmentation-sidebar" aria-live="polite" aria-label="Local operator augmentation sidebar">
<div class="augmentation-heading">LOCAL OPERATOR AUGMENTATION</div>
<div id="augmentation-risk-score">Risk score: —</div>
<div class="augmentation-section">
<h3>Signals</h3>
<ul id="augmentation-signals"><li>No signals yet.</li></ul>
</div>
<div class="augmentation-section">
<h3>Talking points</h3>
<ul id="augmentation-talking-points"><li>Enable operator assist to surface local advisory guidance.</li></ul>
</div>
<div class="augmentation-section">
<h3>De-escalation</h3>
<ul id="augmentation-techniques"><li>Suggestions stay local and never replace human judgment.</li></ul>
</div>
<div class="augmentation-section">
<h3>Follow-up</h3>
<p id="augmentation-follow-up">No follow-up prompt yet.</p>
</div>
<p id="augmentation-notice">Local-only advisory. Never replaces human judgment.</p>
</aside>
<!-- Chat messages -->
<div id="chat-area" role="log" aria-label="Chat messages" aria-live="polite" tabindex="0">
<!-- Messages inserted here -->
@@ -806,6 +928,14 @@ Sovereignty and service always.`;
var sendBtn = document.getElementById('send-btn');
var typingIndicator = document.getElementById('typing-indicator');
var crisisPanel = document.getElementById('crisis-panel');
var augmentationToggle = document.getElementById('augmentation-toggle');
var augmentationSidebar = document.getElementById('augmentation-sidebar');
var augmentationRiskScore = document.getElementById('augmentation-risk-score');
var augmentationSignals = document.getElementById('augmentation-signals');
var augmentationTalkingPoints = document.getElementById('augmentation-talking-points');
var augmentationTechniques = document.getElementById('augmentation-techniques');
var augmentationFollowUp = document.getElementById('augmentation-follow-up');
var augmentationNotice = document.getElementById('augmentation-notice');
var crisisOverlay = document.getElementById('crisis-overlay');
var overlayDismissBtn = document.getElementById('overlay-dismiss-btn');
var overlayCallLink = document.querySelector('.overlay-call');
@@ -826,6 +956,8 @@ Sovereignty and service always.`;
var isStreaming = false;
var overlayTimer = null;
var crisisPanelShown = false;
var lastUserMessage = '';
var augmentationEnabled = false;
// ===== SERVICE WORKER =====
if ('serviceWorker' in navigator) {
@@ -983,6 +1115,142 @@ Sovereignty and service always.`;
}
function escapeHtml(text) {
return String(text || '')
.replace(/&/g, '&amp;')
.replace(/</g, '&lt;')
.replace(/>/g, '&gt;')
.replace(/"/g, '&quot;')
.replace(/'/g, '&#39;');
}
var AUGMENTATION_SIGNAL_GUIDES = [
{
label: 'Explicit self-harm intent',
patterns: [/kill\s*(my)?self/i, /end\s*my\s*life/i, /suicide\s*(note|letter)/i, /going\s+to\s+(kill\s+myself|die)/i],
talkingPoint: 'Ask directly whether they are safe right now and keep the next question concrete.',
technique: 'Move to immediate safety: ask about means, people nearby, and whether 988 can be called or texted now.',
followUp: 'You said you might die tonight. Are you alone right now, and what is within reach?'
},
{
label: 'Hopelessness / collapse',
patterns: [/hopeless/i, /can'?t\s+go\s+on/i, /no\s+future/i, /nothing\s+left/i],
talkingPoint: 'Reflect the hopelessness plainly before offering options. Avoid arguing with the feeling.',
technique: 'Narrow the time horizon to the next ten minutes and one immediate action.',
followUp: 'You said things feel hopeless. What feels most dangerous about the next hour?'
},
{
label: 'Isolation / burden',
patterns: [/nobody\s+cares/i, /better\s+off\s+without\s+me/i, /\balone\b/i, /\bburden\b/i],
talkingPoint: 'Counter isolation with one real contact point: a person, 988, or Crisis Text Line.',
technique: 'Invite a tiny reconnection step: text one safe person, unlock the door, or stay in the chat.',
followUp: 'You said you feel alone. Who is the safest real person we could bring into this moment with you?'
},
{
label: 'Overwhelm / panic',
patterns: [/desperate/i, /overwhelm(?:ed|ing)/i, /trapped/i, /panic/i],
talkingPoint: 'Offer one regulating step at a time instead of a long list.',
technique: 'Ground in the room: feet on the floor, name five visible objects, one sip of water, one slow exhale.',
followUp: 'You said this feels overwhelming. What is the smallest thing in the room you can touch right now?'
}
];
function deriveAugmentationSignals(userText) {
var text = (userText || '').toLowerCase();
return AUGMENTATION_SIGNAL_GUIDES.filter(function(guide) {
return guide.patterns.some(function(pattern) { return pattern.test(text); });
});
}
function buildAugmentationState(userText, assistantText) {
var text = userText || '';
var guides = deriveAugmentationSignals(text);
var level = getCrisisLevel(userText);
var signals = guides.map(function(guide) { return guide.label; });
var explicitIntent = signals.indexOf('Explicit self-harm intent') !== -1;
var riskLevel = explicitIntent ? 'CRITICAL' : (level === 2 ? 'CRITICAL' : level === 1 ? 'HIGH' : (guides.length ? 'LOW' : 'NONE'));
var riskScore = riskLevel === 'CRITICAL' ? 95 : riskLevel === 'HIGH' ? 75 : riskLevel === 'LOW' ? 25 : 5;
riskScore = Math.min(100, riskScore + Math.min(guides.length * 5, 10));
if (riskLevel === 'CRITICAL' && signals.indexOf('Explicit self-harm intent') === -1) {
signals.unshift('Explicit self-harm intent');
riskScore = Math.max(riskScore, 95);
}
var talkingPoints = guides.map(function(guide) { return guide.talkingPoint; });
var techniques = guides.map(function(guide) { return guide.technique; });
if (!talkingPoints.length) {
talkingPoints = ['Keep the response advisory, local-only, and focused on immediate safety rather than abstract reassurance.'];
}
if (!techniques.length) {
techniques = ['Slow the pace. Use short sentences. Invite one concrete grounding step.'];
}
if ((assistantText || '').indexOf('988') === -1 && (riskLevel === 'HIGH' || riskLevel === 'CRITICAL')) {
talkingPoints.push('Surface 988 or Crisis Text Line explicitly if the assistant has not already done so.');
}
var quoted = (text || '').replace(/\s+/g, ' ').slice(0, 120);
var followUp = guides.length ? guides[0].followUp : 'What feels heaviest or most dangerous for you right now?';
return {
riskLevel: riskLevel,
riskScore: riskScore,
signals: signals,
talkingPoints: talkingPoints,
techniques: techniques,
followUpPrompt: 'You said "' + quoted + '". Consider following up with: ' + followUp,
operatorNotice: 'Local-only advisory. Never replaces human judgment.',
localOnly: true,
advisoryOnly: true
};
}
function renderAugmentationSidebar(state) {
if (!augmentationSidebar) return;
augmentationRiskScore.textContent = 'Risk score: ' + state.riskScore + ' / 100 (' + state.riskLevel + ')';
augmentationSignals.innerHTML = state.signals.length
? state.signals.map(function(signal) { return '<li>' + escapeHtml(signal) + '</li>'; }).join('')
: '<li>No crisis signals detected.</li>';
augmentationTalkingPoints.innerHTML = state.talkingPoints.map(function(item) { return '<li>' + escapeHtml(item) + '</li>'; }).join('');
augmentationTechniques.innerHTML = state.techniques.map(function(item) { return '<li>' + escapeHtml(item) + '</li>'; }).join('');
augmentationFollowUp.textContent = state.followUpPrompt;
augmentationNotice.textContent = state.operatorNotice;
augmentationSidebar.classList.add('visible');
}
function updateAugmentationState(userText, assistantText) {
if (!augmentationEnabled) return;
renderAugmentationSidebar(buildAugmentationState(userText, assistantText));
}
function setOperatorAugmentationEnabled(enabled) {
augmentationEnabled = !!enabled;
try { localStorage.setItem('door_operator_augmentation_enabled', augmentationEnabled ? '1' : '0'); } catch (e) {}
if (!augmentationToggle) return;
augmentationToggle.setAttribute('aria-pressed', augmentationEnabled ? 'true' : 'false');
augmentationToggle.classList.toggle('active', augmentationEnabled);
augmentationToggle.textContent = augmentationEnabled ? 'Operator assist: on' : 'Operator assist: off';
if (!augmentationEnabled && augmentationSidebar) {
augmentationSidebar.classList.remove('visible');
return;
}
if (augmentationEnabled && lastUserMessage) {
var lastAssistant = '';
for (var i = messages.length - 1; i >= 0; i--) {
if (messages[i].role === 'assistant') { lastAssistant = messages[i].content; break; }
}
updateAugmentationState(lastUserMessage, lastAssistant);
}
}
function loadOperatorAugmentationPreference() {
try {
return localStorage.getItem('door_operator_augmentation_enabled') === '1';
} catch (e) {
return false;
}
}
// ===== OVERLAY =====
// Focus trap: cycle through focusable elements within the crisis overlay
@@ -1315,9 +1583,10 @@ Sovereignty and service always.`;
addMessage('user', text);
messages.push({ role: 'user', content: text });
var lastUserMessage = text;
lastUserMessage = text;
checkCrisis(text);
updateAugmentationState(text, '');
msgInput.value = '';
msgInput.style.height = 'auto';
@@ -1406,6 +1675,7 @@ Sovereignty and service always.`;
messages.push({ role: 'assistant', content: fullText });
saveMessages();
checkCrisis(fullText);
updateAugmentationState(lastUserMessage || '', fullText);
}
isStreaming = false;
sendBtn.disabled = msgInput.value.trim().length === 0;
@@ -1432,6 +1702,11 @@ Sovereignty and service always.`;
});
sendBtn.addEventListener('click', sendMessage);
if (augmentationToggle) {
augmentationToggle.addEventListener('click', function() {
setOperatorAugmentationEnabled(!augmentationEnabled);
});
}
// ===== WELCOME MESSAGE =====
function init() {
@@ -1451,6 +1726,7 @@ Sovereignty and service always.`;
window.history.replaceState({}, document.title, window.location.pathname);
}
setOperatorAugmentationEnabled(loadOperatorAugmentationPreference());
msgInput.focus();
}

View File

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

View File

@@ -0,0 +1,33 @@
from augmentation import CounselorAugmentationEngine
def test_explicit_intent_forces_critical_sidebar_guidance():
engine = CounselorAugmentationEngine()
result = engine.build_augmented_guidance(
"I want to kill myself tonight. I already wrote a note.",
assistant_text="I'm here with you."
)
assert result.risk_level == "CRITICAL"
assert result.risk_score >= 90
assert result.local_only is True
assert result.advisory_only is True
assert "Explicit self-harm intent" in result.signals
assert result.suggested_talking_points
assert result.deescalation_techniques
assert "You said" in result.follow_up_prompt
assert "never replaces human judgment" in result.operator_notice.lower()
def test_hopelessness_signal_produces_follow_up_and_talking_points():
engine = CounselorAugmentationEngine()
result = engine.build_augmented_guidance(
"I feel so hopeless about my life and I can't go on.",
assistant_text=""
)
assert result.risk_level in {"HIGH", "CRITICAL"}
assert result.signals
assert result.suggested_talking_points
assert result.deescalation_techniques
assert result.follow_up_prompt

View File

@@ -0,0 +1,20 @@
from pathlib import Path
def test_operator_augmentation_ui_hooks_exist():
html = Path('index.html').read_text()
assert 'id="augmentation-toggle"' in html
assert 'id="augmentation-sidebar"' in html
assert 'id="augmentation-risk-score"' in html
assert 'id="augmentation-signals"' in html
assert 'id="augmentation-follow-up"' in html
assert 'door_operator_augmentation_enabled' in html
assert 'function buildAugmentationState(' in html
assert 'function renderAugmentationSidebar(' in html
assert 'function updateAugmentationState(' in html
assert 'function setOperatorAugmentationEnabled(' in html
assert 'function loadOperatorAugmentationPreference(' in html
assert 'getCrisisLevel(userText)' in html
assert "updateAugmentationState(text, '')" in html
assert "updateAugmentationState(lastUserMessage || '', fullText)" in html

View File

@@ -0,0 +1,26 @@
from pathlib import Path
from playwright.sync_api import sync_playwright
def test_operator_augmentation_walkthrough_marks_explicit_intent_critical():
url = Path('index.html').resolve().as_uri()
with sync_playwright() as p:
browser = p.chromium.launch(headless=True)
page = browser.new_page()
page.goto(url, wait_until='load')
page.click('#augmentation-toggle')
page.fill('#msg-input', 'I want to kill myself tonight. I already wrote a note.')
page.click('#send-btn')
page.wait_for_timeout(300)
risk = page.locator('#augmentation-risk-score').inner_text()
signals = page.locator('#augmentation-signals').inner_text()
follow_up = page.locator('#augmentation-follow-up').inner_text()
browser.close()
assert 'CRITICAL' in risk
assert 'Explicit self-harm intent' in signals
assert 'You said "I want to kill myself tonight. I already wrote a note."' in follow_up