Compare commits

..

1 Commits

Author SHA1 Message Date
Alexander Whitestone
0ab2626ef2 feat: image content screening for self-harm indicators (closes #132)
All checks were successful
Sanity Checks / sanity-test (pull_request) Successful in 4s
Smoke Test / smoke (pull_request) Successful in 10s
2026-04-15 12:02:22 -04:00
4 changed files with 243 additions and 91 deletions

155
image_screening.py Normal file
View File

@@ -0,0 +1,155 @@
"""
Image Content Screening for Self-Harm Indicators
Screens images sent in chat for self-harm content using vision models.
Privacy-preserving: analyzes in-memory, discards after scoring. No image retention.
"""
import base64
import json
import logging
from dataclasses import dataclass
from enum import Enum
from typing import Optional
logger = logging.getLogger("the-door.image_screening")
class RiskLevel(str, Enum):
SAFE = "safe"
CONCERNING = "concerning"
CRITICAL = "critical"
@dataclass
class ImageScreenResult:
risk_level: RiskLevel
confidence: float
categories: list
reasoning: str
model_used: str
@property
def needs_crisis_response(self):
return self.risk_level == RiskLevel.CRITICAL
@property
def needs_followup(self):
return self.risk_level in (RiskLevel.CONCERNING, RiskLevel.CRITICAL)
def to_dict(self):
return {
"risk_level": self.risk_level.value,
"confidence": round(self.confidence, 3),
"categories": self.categories,
"reasoning": self.reasoning,
"model_used": self.model_used,
}
SCREENING_PROMPT = """Analyze this image for indicators of self-harm or crisis. Look for:
1. Self-harm wounds or injuries
2. Medication/pills in concerning quantities
3. Farewell-type imagery (empty rooms, items laid out, goodbye notes)
4. Screenshots of crisis-related searches
5. Images conveying hopelessness, isolation, or despair
Respond in JSON format ONLY:
{"risk_level": "safe"|"concerning"|"critical", "confidence": 0.0-1.0, "categories": [], "reasoning": ""}
CRITICAL: imminent self-harm indicators
CONCERNING: ambiguous but worrying
SAFE: no indicators detected"""
def _analyze_with_ollama(image_b64, model="gemma3:4b"):
try:
import urllib.request
payload = json.dumps({
"model": model,
"messages": [{
"role": "user",
"content": SCREENING_PROMPT,
"images": [image_b64],
}],
"stream": False,
"options": {"temperature": 0.1},
}).encode()
req = urllib.request.Request(
"http://localhost:11434/api/chat",
data=payload,
headers={"Content-Type": "application/json"},
method="POST",
)
resp = urllib.request.urlopen(req, timeout=30)
data = json.loads(resp.read())
content = data.get("message", {}).get("content", "")
json_start = content.find("{")
json_end = content.rfind("}") + 1
if json_start == -1 or json_end <= json_start:
return None
result = json.loads(content[json_start:json_end])
return ImageScreenResult(
risk_level=RiskLevel(result.get("risk_level", "safe")),
confidence=float(result.get("confidence", 0.5)),
categories=result.get("categories", []),
reasoning=result.get("reasoning", ""),
model_used=f"ollama:{model}",
)
except Exception as e:
logger.warning(f"Ollama vision analysis failed: {e}")
return None
def _analyze_fallback(image_bytes):
return ImageScreenResult(
risk_level=RiskLevel.SAFE,
confidence=0.2,
categories=["unanalyzed"],
reasoning="No vision model available. Defaulting to safe with low confidence.",
model_used="fallback:heuristic",
)
def screen_image(image_data, use_vision_model=True, model="gemma3:4b"):
"""Screen image for self-harm indicators. Analyzes in-memory, no retention."""
if isinstance(image_data, bytes):
image_b64 = base64.b64encode(image_data).decode()
else:
image_b64 = image_data
image_data = base64.b64decode(image_b64)
if use_vision_model:
result = _analyze_with_ollama(image_b64, model)
if result:
logger.info(f"Image screened: {result.risk_level.value} (conf: {result.confidence:.2f})")
if result.needs_crisis_response:
logger.warning(f"CRITICAL image: {result.reasoning}")
return result
return _analyze_fallback(image_data)
def handle_chat_image(image_data):
"""Handle image from chat. Returns action dict for gateway."""
result = screen_image(image_data)
action = {
"result": result.to_dict(),
"show_crisis_overlay": result.needs_crisis_response,
"log_event": result.needs_followup,
"response_text": None,
}
if result.risk_level == RiskLevel.CRITICAL:
action["response_text"] = (
"I noticed something concerning in the image you shared. "
"If you or someone you know is in crisis, please reach out: "
"988 Suicide and Crisis Lifeline (call or text 988). "
"You are not alone."
)
elif result.risk_level == RiskLevel.CONCERNING:
action["response_text"] = (
"I want to check in \u2014 how are you doing? "
"If you need to talk to someone, the 988 Lifeline is available 24/7."
)
return action

View File

@@ -475,26 +475,6 @@ html, body {
margin-bottom: 24px;
}
.modal-status {
min-height: 22px;
margin: 0 0 16px;
font-size: 0.9rem;
line-height: 1.45;
color: #8b949e;
}
.modal-status.is-visible {
display: block;
}
.modal-status.success {
color: #3fb950;
}
.modal-status.error {
color: #ff7b72;
}
.form-group {
margin-bottom: 16px;
}
@@ -757,7 +737,6 @@ html, body {
<textarea id="sp-environment" placeholder="e.g., Giving my car keys to a friend, locking away meds..."></textarea>
</div>
</div>
<div id="safety-plan-status" class="modal-status" role="status" aria-live="polite" aria-atomic="true"></div>
<div class="modal-footer">
<button class="btn btn-secondary" id="cancel-safety-plan">Cancel</button>
<button class="btn btn-primary" id="save-safety-plan">Save Plan</button>
@@ -839,7 +818,6 @@ Sovereignty and service always.`;
var closeSafetyPlan = document.getElementById('close-safety-plan');
var cancelSafetyPlan = document.getElementById('cancel-safety-plan');
var saveSafetyPlan = document.getElementById('save-safety-plan');
var safetyPlanStatus = document.getElementById('safety-plan-status');
var clearChatBtn = document.getElementById('clear-chat-btn');
// ===== STATE =====
@@ -1205,24 +1183,12 @@ Sovereignty and service always.`;
} catch (e) {}
}
function setSafetyPlanStatus(message, type) {
safetyPlanStatus.textContent = message;
safetyPlanStatus.className = 'modal-status is-visible ' + (type || '');
}
function clearSafetyPlanStatus() {
safetyPlanStatus.textContent = '';
safetyPlanStatus.className = 'modal-status';
}
closeSafetyPlan.addEventListener('click', function() {
clearSafetyPlanStatus();
safetyPlanModal.classList.remove('active');
_restoreSafetyPlanFocus();
});
cancelSafetyPlan.addEventListener('click', function() {
clearSafetyPlanStatus();
safetyPlanModal.classList.remove('active');
_restoreSafetyPlanFocus();
});
@@ -1237,9 +1203,11 @@ Sovereignty and service always.`;
};
try {
localStorage.setItem('timmy_safety_plan', JSON.stringify(plan));
setSafetyPlanStatus('Safety plan saved locally.', 'success');
safetyPlanModal.classList.remove('active');
_restoreSafetyPlanFocus();
alert('Safety plan saved locally.');
} catch (e) {
setSafetyPlanStatus('Error saving plan.', 'error');
alert('Error saving plan.');
}
});
@@ -1317,7 +1285,6 @@ Sovereignty and service always.`;
// Wire open buttons to activate focus trap
safetyPlanBtn.addEventListener('click', function() {
clearSafetyPlanStatus();
loadSafetyPlan();
safetyPlanModal.classList.add('active');
_activateSafetyPlanFocusTrap(safetyPlanBtn);
@@ -1326,8 +1293,6 @@ Sovereignty and service always.`;
// Crisis panel safety plan button (if crisis panel is visible)
if (crisisSafetyPlanBtn) {
crisisSafetyPlanBtn.addEventListener('click', function() {
clearSafetyPlanStatus();
clearSafetyPlanStatus();
loadSafetyPlan();
safetyPlanModal.classList.add('active');
_activateSafetyPlanFocusTrap(crisisSafetyPlanBtn);

View File

@@ -0,0 +1,84 @@
"""Tests for image content screening module."""
import json
from unittest.mock import patch, MagicMock
from image_screening import (
RiskLevel,
ImageScreenResult,
screen_image,
handle_chat_image,
_analyze_fallback,
)
class TestImageScreenResult:
def test_safe_result(self):
result = ImageScreenResult(
risk_level=RiskLevel.SAFE, confidence=0.95,
categories=[], reasoning="No indicators", model_used="test"
)
assert not result.needs_crisis_response
assert not result.needs_followup
assert result.to_dict()["risk_level"] == "safe"
def test_critical_result(self):
result = ImageScreenResult(
risk_level=RiskLevel.CRITICAL, confidence=0.9,
categories=["wounds"], reasoning="Detected", model_used="test"
)
assert result.needs_crisis_response
assert result.needs_followup
def test_concerning_result(self):
result = ImageScreenResult(
risk_level=RiskLevel.CONCERNING, confidence=0.6,
categories=["isolation"], reasoning="Ambiguous", model_used="test"
)
assert not result.needs_crisis_response
assert result.needs_followup
class TestScreenImage:
def test_fallback_returns_safe(self):
result = screen_image(b"fake_image_data", use_vision_model=False)
assert result.risk_level == RiskLevel.SAFE
assert result.model_used == "fallback:heuristic"
assert result.confidence < 0.5
def test_base64_input(self):
import base64
b64 = base64.b64encode(b"fake").decode()
result = screen_image(b64, use_vision_model=False)
assert result.risk_level == RiskLevel.SAFE
class TestHandleChatImage:
def test_safe_image_no_overlay(self):
action = handle_chat_image(b"safe_image")
assert not action["show_crisis_overlay"]
assert action["response_text"] is None
@patch("image_screening._analyze_with_ollama")
def test_critical_image_shows_overlay(self, mock_ollama):
mock_ollama.return_value = ImageScreenResult(
risk_level=RiskLevel.CRITICAL, confidence=0.95,
categories=["wounds"], reasoning="Self-harm detected",
model_used="ollama:gemma3:4b"
)
action = handle_chat_image(b"concerning_image")
assert action["show_crisis_overlay"]
assert "988" in action["response_text"]
assert action["log_event"]
@patch("image_screening._analyze_with_ollama")
def test_concerning_image_followup(self, mock_ollama):
mock_ollama.return_value = ImageScreenResult(
risk_level=RiskLevel.CONCERNING, confidence=0.6,
categories=["isolation"], reasoning="Empty room",
model_used="ollama:gemma3:4b"
)
action = handle_chat_image(b"maybe_concerning")
assert not action["show_crisis_overlay"]
assert action["log_event"]
assert "check in" in action["response_text"]

View File

@@ -1,52 +0,0 @@
import pathlib
import re
import unittest
ROOT = pathlib.Path(__file__).resolve().parents[1]
INDEX_HTML = ROOT / 'index.html'
class TestSafetyPlanSaveFeedback(unittest.TestCase):
@classmethod
def setUpClass(cls):
cls.html = INDEX_HTML.read_text()
def test_modal_has_inline_status_live_region(self):
self.assertRegex(
self.html,
r'<div[^>]+id="safety-plan-status"[^>]+role="status"[^>]+aria-live="polite"[^>]*>',
'Expected an inline polite live region for safety plan save feedback.',
)
def test_save_feedback_does_not_use_blocking_alerts(self):
self.assertNotIn(
"alert('Safety plan saved locally.')",
self.html,
'Expected success feedback to stop using blocking alert().',
)
self.assertNotIn(
"alert('Error saving plan.')",
self.html,
'Expected error feedback to stop using blocking alert().',
)
def test_save_logic_updates_inline_status_for_success_and_error(self):
self.assertRegex(
self.html,
r'function\s+setSafetyPlanStatus\s*\(',
'Expected a helper to update inline save feedback.',
)
self.assertRegex(
self.html,
r"setSafetyPlanStatus\('Safety plan saved locally\.'\s*,\s*'success'\)",
'Expected success path to update inline status.',
)
self.assertRegex(
self.html,
r"setSafetyPlanStatus\('Error saving plan\.'\s*,\s*'error'\)",
'Expected error path to update inline status.',
)
if __name__ == '__main__':
unittest.main()