Compare commits
5 Commits
| Author | SHA1 | Date | |
|---|---|---|---|
|
|
53bfb47a92 | ||
|
|
08e3ece2d3 | ||
|
|
100cc743c0 | ||
|
|
f7d99c6d9c | ||
| d412939b4f |
@@ -95,45 +95,7 @@ Run tests:
|
||||
python -m pytest crisis/tests.py -v
|
||||
```
|
||||
|
||||
### 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
|
||||
### 7. Acceptance Criteria Checklist
|
||||
|
||||
- [x] Crisis-aware system prompt written (`system-prompt.txt`)
|
||||
- [x] Frontend embeds system prompt on every API request (`index.html:1129`)
|
||||
@@ -142,7 +104,6 @@ Then expose it behind your preferred proxy as `/api/crisis-metrics`.
|
||||
- [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
|
||||
|
||||
@@ -6,20 +6,7 @@ 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,
|
||||
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 .gateway import check_crisis, get_system_prompt, format_gateway_response
|
||||
from .session_tracker import CrisisSessionTracker, SessionState, check_crisis_with_session
|
||||
|
||||
__all__ = [
|
||||
@@ -29,17 +16,10 @@ __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",
|
||||
|
||||
@@ -14,13 +14,10 @@ Usage:
|
||||
import json
|
||||
from typing import Optional
|
||||
|
||||
from image_screening import screen_image_signals
|
||||
|
||||
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,
|
||||
@@ -30,24 +27,17 @@ from .response import (
|
||||
from .session_tracker import CrisisSessionTracker
|
||||
|
||||
|
||||
def check_crisis(
|
||||
text: str,
|
||||
metrics_tracker: Optional[CrisisMetricsTracker] = None,
|
||||
timestamp: Optional[object] = None,
|
||||
) -> dict:
|
||||
def check_crisis(text: str) -> 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)
|
||||
|
||||
result = {
|
||||
return {
|
||||
"level": detection.level,
|
||||
"score": detection.score,
|
||||
"indicators": detection.indicators,
|
||||
@@ -61,29 +51,66 @@ def check_crisis(
|
||||
"escalate": response.escalate,
|
||||
}
|
||||
|
||||
if metrics_tracker is not None:
|
||||
metrics_tracker.record_gateway_result(result, timestamp=timestamp)
|
||||
|
||||
return result
|
||||
def _image_detection_from_score(image_result) -> CrisisDetectionResult:
|
||||
if image_result.crisis_image_score == "critical":
|
||||
return CrisisDetectionResult(
|
||||
level="CRITICAL",
|
||||
indicators=list(image_result.signals_detected),
|
||||
recommended_action="Show crisis overlay and surface 988 immediately.",
|
||||
score=image_result.distress_score,
|
||||
)
|
||||
if image_result.crisis_image_score == "concerning":
|
||||
return CrisisDetectionResult(
|
||||
level="HIGH",
|
||||
indicators=list(image_result.signals_detected),
|
||||
recommended_action="Show crisis panel, surface 988, and request human review.",
|
||||
score=image_result.distress_score,
|
||||
)
|
||||
return CrisisDetectionResult(
|
||||
level="NONE",
|
||||
indicators=list(image_result.signals_detected),
|
||||
recommended_action="No crisis action required.",
|
||||
score=image_result.distress_score,
|
||||
)
|
||||
|
||||
|
||||
def check_crisis_and_record(
|
||||
text: str,
|
||||
tracker: Optional[CrisisMetricsTracker] = None,
|
||||
timestamp: Optional[object] = None,
|
||||
def check_image_crisis(
|
||||
*,
|
||||
image_path: Optional[str] = None,
|
||||
ocr_text: str = "",
|
||||
labels: Optional[list[str]] = None,
|
||||
manual_notes: str = "",
|
||||
visual_flags: Optional[list[str]] = 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)
|
||||
"""Gateway-integrated image crisis check using the local screening slice."""
|
||||
image_result = screen_image_signals(
|
||||
image_path=image_path,
|
||||
ocr_text=ocr_text,
|
||||
labels=labels,
|
||||
manual_notes=manual_notes,
|
||||
visual_flags=visual_flags,
|
||||
)
|
||||
detection = _image_detection_from_score(image_result)
|
||||
response = generate_response(detection)
|
||||
|
||||
|
||||
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)
|
||||
return {
|
||||
"level": detection.level,
|
||||
"image_score": image_result.crisis_image_score,
|
||||
"score": detection.score,
|
||||
"indicators": detection.indicators,
|
||||
"recommended_action": detection.recommended_action,
|
||||
"timmy_message": response.timmy_message,
|
||||
"ui": {
|
||||
"show_crisis_panel": response.show_crisis_panel,
|
||||
"show_overlay": response.show_overlay,
|
||||
"provide_988": response.provide_988,
|
||||
},
|
||||
"escalate": response.escalate,
|
||||
"requires_human_review": image_result.requires_human_review,
|
||||
"grounded_scope": image_result.grounded_scope,
|
||||
"screening": image_result.to_dict(),
|
||||
}
|
||||
|
||||
|
||||
def get_system_prompt(base_prompt: str, text: str = "") -> str:
|
||||
|
||||
@@ -1,239 +0,0 @@
|
||||
"""
|
||||
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")
|
||||
195
image_screening.py
Normal file
195
image_screening.py
Normal file
@@ -0,0 +1,195 @@
|
||||
"""
|
||||
image_screening.py — local image crisis screening slice for epic #130.
|
||||
|
||||
Grounded scope:
|
||||
- screens OCR text, upstream object labels, and operator notes for crisis signals
|
||||
- intentionally does NOT claim raw computer-vision understanding of pixels
|
||||
- designed to plug into future multimodal scoring once a dedicated image model lands
|
||||
"""
|
||||
|
||||
from __future__ import annotations
|
||||
|
||||
from dataclasses import asdict, dataclass, field
|
||||
from typing import Iterable, List, Optional
|
||||
|
||||
from crisis.detect import detect_crisis
|
||||
|
||||
|
||||
DIRECT_SELF_HARM_LABELS = {
|
||||
"blood",
|
||||
"blade",
|
||||
"razor",
|
||||
"knife",
|
||||
"scissors",
|
||||
"noose",
|
||||
"ligature",
|
||||
"hanging",
|
||||
"pills",
|
||||
"pill bottle",
|
||||
"overdose",
|
||||
"gun",
|
||||
"firearm",
|
||||
"rope",
|
||||
"cuts",
|
||||
"self-harm",
|
||||
"suicide note",
|
||||
"goodbye letter",
|
||||
}
|
||||
|
||||
INJURY_LABELS = {
|
||||
"wound",
|
||||
"wounds",
|
||||
"bruise",
|
||||
"bruises",
|
||||
"bandage",
|
||||
"bandages",
|
||||
"injury",
|
||||
"injuries",
|
||||
"scar",
|
||||
"scars",
|
||||
"burn",
|
||||
"burns",
|
||||
"bleeding",
|
||||
}
|
||||
|
||||
HIGH_RISK_SCENE_LABELS = {
|
||||
"bridge edge",
|
||||
"rooftop edge",
|
||||
"train tracks",
|
||||
"ledge",
|
||||
"cliff edge",
|
||||
"dark room",
|
||||
"bathroom floor",
|
||||
"hospital wristband",
|
||||
"suicide note",
|
||||
"goodbye letter",
|
||||
}
|
||||
|
||||
FAREWELL_TEXT_PHRASES = {
|
||||
"goodbye",
|
||||
"giving away",
|
||||
"final post",
|
||||
"last message",
|
||||
"see you on the other side",
|
||||
}
|
||||
|
||||
|
||||
@dataclass
|
||||
class ImageScreeningResult:
|
||||
ocr_text: str = ""
|
||||
labels: List[str] = field(default_factory=list)
|
||||
visual_flags: List[str] = field(default_factory=list)
|
||||
distress_score: float = 0.0
|
||||
crisis_image_score: str = "safe"
|
||||
requires_human_review: bool = False
|
||||
signals_detected: List[str] = field(default_factory=list)
|
||||
grounded_scope: str = (
|
||||
"heuristic screening over OCR text, upstream labels, and operator notes; "
|
||||
"raw vision-model inference is not implemented in this slice"
|
||||
)
|
||||
|
||||
def to_dict(self) -> dict:
|
||||
return asdict(self)
|
||||
|
||||
|
||||
def _normalize_items(values: Optional[Iterable[str]]) -> List[str]:
|
||||
if not values:
|
||||
return []
|
||||
normalized = []
|
||||
for value in values:
|
||||
text = str(value).strip().lower()
|
||||
if text:
|
||||
normalized.append(text)
|
||||
return normalized
|
||||
|
||||
|
||||
def _match_keywords(haystack: str, keywords: set[str]) -> List[str]:
|
||||
matches = []
|
||||
for keyword in keywords:
|
||||
if keyword in haystack:
|
||||
matches.append(keyword)
|
||||
return sorted(set(matches))
|
||||
|
||||
|
||||
def screen_image_signals(
|
||||
image_path: Optional[str] = None,
|
||||
*,
|
||||
ocr_text: str = "",
|
||||
labels: Optional[Iterable[str]] = None,
|
||||
manual_notes: str = "",
|
||||
visual_flags: Optional[Iterable[str]] = None,
|
||||
) -> ImageScreeningResult:
|
||||
"""
|
||||
Score image-related crisis evidence without pretending to do full CV.
|
||||
|
||||
Inputs are deliberately grounded in what the repo can actually support today:
|
||||
- OCR text extracted upstream from screenshots/photos
|
||||
- object labels supplied by a local model or operator
|
||||
- manual operator notes about visible scene context
|
||||
- optional visual flags from any upstream preprocessor
|
||||
"""
|
||||
normalized_labels = _normalize_items(labels)
|
||||
normalized_flags = _normalize_items(visual_flags)
|
||||
normalized_notes = str(manual_notes or "").strip().lower()
|
||||
normalized_ocr = str(ocr_text or "").strip()
|
||||
|
||||
combined_label_text = " ".join(normalized_labels + normalized_flags + ([normalized_notes] if normalized_notes else []))
|
||||
crisis_text = " ".join(part for part in [normalized_ocr, normalized_notes] if part).strip()
|
||||
|
||||
direct_matches = _match_keywords(combined_label_text, DIRECT_SELF_HARM_LABELS)
|
||||
injury_matches = _match_keywords(combined_label_text, INJURY_LABELS)
|
||||
scene_matches = _match_keywords(combined_label_text, HIGH_RISK_SCENE_LABELS)
|
||||
farewell_matches = _match_keywords(crisis_text.lower(), FAREWELL_TEXT_PHRASES)
|
||||
text_detection = detect_crisis(crisis_text) if crisis_text else None
|
||||
|
||||
signals: List[str] = []
|
||||
score = 0.0
|
||||
|
||||
if direct_matches:
|
||||
score = max(score, 0.85)
|
||||
for match in direct_matches:
|
||||
signals.append(f"direct_self_harm_label:{match}")
|
||||
|
||||
if injury_matches:
|
||||
score = max(score, 0.55)
|
||||
for match in injury_matches:
|
||||
signals.append(f"injury_indicator:{match}")
|
||||
|
||||
if scene_matches:
|
||||
score = max(score, 0.4)
|
||||
for match in scene_matches:
|
||||
signals.append(f"high_risk_scene:{match}")
|
||||
|
||||
if farewell_matches:
|
||||
score = max(score, 0.85)
|
||||
for match in farewell_matches:
|
||||
signals.append(f"farewell_text:{match}")
|
||||
|
||||
if text_detection and text_detection.level != "NONE":
|
||||
score = max(score, min(1.0, text_detection.score))
|
||||
signals.append(f"ocr_crisis_level:{text_detection.level}")
|
||||
for indicator in text_detection.indicators[:3]:
|
||||
signals.append(f"ocr_indicator:{indicator}")
|
||||
|
||||
if direct_matches and text_detection and text_detection.level in {"HIGH", "CRITICAL"}:
|
||||
score = min(1.0, max(score, 0.95))
|
||||
signals.append("cross_modal_confirmation:text_plus_visual")
|
||||
|
||||
if direct_matches or (text_detection and text_detection.level == "CRITICAL") or score >= 0.85:
|
||||
crisis_image_score = "critical"
|
||||
elif score >= 0.4 or (text_detection and text_detection.level in {"HIGH", "MEDIUM"}):
|
||||
crisis_image_score = "concerning"
|
||||
else:
|
||||
crisis_image_score = "safe"
|
||||
|
||||
requires_human_review = score >= 0.4 or bool(direct_matches)
|
||||
|
||||
return ImageScreeningResult(
|
||||
ocr_text=normalized_ocr,
|
||||
labels=list(normalized_labels),
|
||||
visual_flags=list(normalized_flags),
|
||||
distress_score=round(score, 4),
|
||||
crisis_image_score=crisis_image_score,
|
||||
requires_human_review=requires_human_review,
|
||||
signals_detected=signals,
|
||||
)
|
||||
@@ -680,7 +680,7 @@ html, body {
|
||||
|
||||
<!-- Footer -->
|
||||
<footer id="footer">
|
||||
<a href="/about" aria-label="About The Door">about</a>
|
||||
<a href="/about.html" aria-label="About The Door">about</a>
|
||||
<button id="safety-plan-btn" aria-label="Open My Safety Plan">my safety plan</button>
|
||||
<button id="clear-chat-btn" aria-label="Clear chat history">clear chat</button>
|
||||
</footer>
|
||||
|
||||
@@ -1,118 +0,0 @@
|
||||
"""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()
|
||||
115
tests/test_image_screening.py
Normal file
115
tests/test_image_screening.py
Normal file
@@ -0,0 +1,115 @@
|
||||
"""Tests for local image crisis screening slice under epic #130."""
|
||||
|
||||
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_image_crisis
|
||||
from image_screening import ImageScreeningResult, screen_image_signals
|
||||
|
||||
|
||||
class TestImageScreeningResult(unittest.TestCase):
|
||||
def test_to_dict_preserves_core_fields(self):
|
||||
result = ImageScreeningResult(
|
||||
ocr_text="help me",
|
||||
labels=["blood"],
|
||||
visual_flags=["dark_scene"],
|
||||
distress_score=0.8,
|
||||
requires_human_review=True,
|
||||
signals_detected=["direct_self_harm_label:blood"],
|
||||
)
|
||||
|
||||
data = result.to_dict()
|
||||
self.assertEqual(data["ocr_text"], "help me")
|
||||
self.assertEqual(data["labels"], ["blood"])
|
||||
self.assertTrue(data["requires_human_review"])
|
||||
|
||||
|
||||
class TestScreenImageSignals(unittest.TestCase):
|
||||
def test_direct_self_harm_labels_trigger_high_risk(self):
|
||||
result = screen_image_signals(
|
||||
labels=["razor blade", "blood droplets"],
|
||||
manual_notes="photo of fresh cuts on forearm",
|
||||
)
|
||||
|
||||
self.assertEqual(result.crisis_image_score, "critical")
|
||||
self.assertGreaterEqual(result.distress_score, 0.8)
|
||||
self.assertTrue(result.requires_human_review)
|
||||
self.assertTrue(any("self_harm" in signal for signal in result.signals_detected))
|
||||
|
||||
def test_ocr_text_uses_existing_crisis_detector(self):
|
||||
result = screen_image_signals(
|
||||
ocr_text="I want to kill myself tonight",
|
||||
labels=["handwritten note"],
|
||||
)
|
||||
|
||||
self.assertGreaterEqual(result.distress_score, 0.7)
|
||||
self.assertTrue(result.requires_human_review)
|
||||
self.assertTrue(any(signal.startswith("ocr_crisis_level:") for signal in result.signals_detected))
|
||||
|
||||
def test_neutral_image_stays_low_risk(self):
|
||||
result = screen_image_signals(
|
||||
labels=["dog", "park", "sunlight"],
|
||||
manual_notes="family outing in daylight",
|
||||
)
|
||||
|
||||
self.assertEqual(result.crisis_image_score, "safe")
|
||||
self.assertLess(result.distress_score, 0.2)
|
||||
self.assertFalse(result.requires_human_review)
|
||||
self.assertEqual(result.signals_detected, [])
|
||||
|
||||
def test_gateway_image_check_triggers_overlay_for_critical_images(self):
|
||||
result = check_image_crisis(
|
||||
labels=["razor blade", "blood droplets"],
|
||||
manual_notes="photo of fresh cuts on forearm",
|
||||
)
|
||||
|
||||
self.assertEqual(result["level"], "CRITICAL")
|
||||
self.assertEqual(result["image_score"], "critical")
|
||||
self.assertTrue(result["ui"]["show_overlay"])
|
||||
self.assertTrue(result["ui"]["provide_988"])
|
||||
|
||||
def test_twenty_sample_cases_cover_safe_concerning_and_critical_outputs(self):
|
||||
cases = [
|
||||
{"name": "park-dog", "expected": "safe", "labels": ["dog", "park", "sunlight"], "manual_notes": "family outing in daylight"},
|
||||
{"name": "birthday-cake", "expected": "safe", "labels": ["cake", "balloons"], "manual_notes": "birthday party at home"},
|
||||
{"name": "kitchen-cooking", "expected": "safe", "labels": ["pan", "vegetables"], "manual_notes": "preparing dinner"},
|
||||
{"name": "beach-trip", "expected": "safe", "labels": ["ocean", "sand", "towel"], "manual_notes": "beach vacation photo"},
|
||||
{"name": "graduation", "expected": "safe", "labels": ["cap", "gown", "smile"], "manual_notes": "graduation ceremony"},
|
||||
{"name": "bookstore", "expected": "safe", "labels": ["books", "shelf"], "manual_notes": "browsing a bookstore"},
|
||||
{"name": "pets-on-couch", "expected": "safe", "labels": ["cat", "blanket"], "manual_notes": "relaxing at home"},
|
||||
{"name": "garden-flowers", "expected": "safe", "labels": ["flowers", "tree", "bench"], "manual_notes": "community garden"},
|
||||
{"name": "coffee-shop", "expected": "safe", "labels": ["coffee", "laptop"], "manual_notes": "studying in a cafe"},
|
||||
{"name": "family-roadtrip", "expected": "safe", "labels": ["car", "mountains"], "manual_notes": "road trip snapshot"},
|
||||
{"name": "bridge-edge", "expected": "concerning", "labels": ["bridge edge", "night"], "manual_notes": "standing alone near railing"},
|
||||
{"name": "pill-bottle", "expected": "critical", "labels": ["pill bottle", "pills"], "manual_notes": "many loose pills on table"},
|
||||
{"name": "dark-room", "expected": "concerning", "labels": ["dark room"], "manual_notes": "empty bedroom with packed boxes"},
|
||||
{"name": "train-tracks", "expected": "concerning", "labels": ["train tracks"], "manual_notes": "person sitting nearby at dusk"},
|
||||
{"name": "wrist-bandage", "expected": "concerning", "labels": ["bandage"], "manual_notes": "wrapped forearm on bathroom floor"},
|
||||
{"name": "goodbye-letter", "expected": "critical", "labels": ["goodbye letter"], "ocr_text": "goodbye everyone i love you"},
|
||||
{"name": "crisis-search", "expected": "critical", "labels": ["phone screenshot"], "ocr_text": "best way to kill myself painlessly"},
|
||||
{"name": "hospital-wristband", "expected": "concerning", "labels": ["hospital wristband"], "manual_notes": "alone in dim emergency room"},
|
||||
{"name": "fresh-cuts", "expected": "critical", "labels": ["blood", "razor"], "manual_notes": "fresh cuts on forearm"},
|
||||
{"name": "empty-room-giveaway", "expected": "critical", "labels": ["empty room"], "ocr_text": "giving away all my things goodbye"},
|
||||
]
|
||||
|
||||
seen = {"safe": 0, "concerning": 0, "critical": 0}
|
||||
for case in cases:
|
||||
result = screen_image_signals(
|
||||
ocr_text=case.get("ocr_text", ""),
|
||||
labels=case.get("labels", []),
|
||||
manual_notes=case.get("manual_notes", ""),
|
||||
)
|
||||
self.assertEqual(result.crisis_image_score, case["expected"], case["name"])
|
||||
seen[case["expected"]] += 1
|
||||
|
||||
self.assertEqual(sum(seen.values()), 20)
|
||||
self.assertEqual(seen["safe"], 10)
|
||||
self.assertGreaterEqual(seen["concerning"], 5)
|
||||
self.assertGreaterEqual(seen["critical"], 5)
|
||||
|
||||
|
||||
if __name__ == "__main__":
|
||||
unittest.main()
|
||||
Reference in New Issue
Block a user