Add auto-admin script and route DB/Media to persistent storage

Dockerfile

@@ -1,4 +1,4 @@
-FROM python:3.12-slim
+FROM python:3.12-slim
 
 # Install system dependencies for OpenCV/YOLO
 RUN apt-get update && apt-get install -y \
@@ -46,5 +46,7 @@ CMD cd plant_core && \
     python manage.py collectstatic --no-input && \
     echo "Running migrations..." && \
     python manage.py migrate && \
+    echo "Setting up admin user..." && \
+    python manage.py setup_admin && \
     echo "Starting Gunicorn..." && \
     gunicorn plant_core.wsgi:application --bind 0.0.0.0:7860 --timeout 300 --log-level info

check_classes.py

@@ -0,0 +1,13 @@
+import sys
+import django
+
+sys.path.append('d:/Plant_App/plant_core')
+import os
+os.environ.setdefault('DJANGO_SETTINGS_MODULE', 'plant_core.settings')
+django.setup()
+
+from diseases.ml_logic import CLASS_NAMES
+print("Mango classes:")
+for c in CLASS_NAMES:
+    if "mango" in c.lower():
+        print(c)

plant_core/diseases/management/commands/setup_admin.py

@@ -0,0 +1,20 @@
+from django.core.management.base import BaseCommand
+from django.contrib.auth import get_user_model
+import os
+
+class Command(BaseCommand):
+    help = 'Automatically creates a superuser from ADMIN_USERNAME and ADMIN_PASSWORD environment variables'
+
+    def handle(self, *args, **options):
+        User = get_user_model()
+        
+        username = os.environ.get('ADMIN_USERNAME', 'admin')
+        password = os.environ.get('ADMIN_PASSWORD', 'admin')
+        email = os.environ.get('ADMIN_EMAIL', 'admin@example.com')
+
+        if not User.objects.filter(username=username).exists():
+            self.stdout.write(f"Creating superuser '{username}'...")
+            User.objects.create_superuser(username=username, email=email, password=password)
+            self.stdout.write(self.style.SUCCESS(f"Successfully created superuser '{username}'"))
+        else:
+            self.stdout.write(f"Superuser '{username}' already exists. Skipping creation.")

plant_core/plant_core/settings.py

@@ -32,9 +32,15 @@ BASE_DIR = Path(__file__).resolve().parent.parent
 
 import dj_database_url
 
+# Check if running on Hugging Face Spaces with persistent storage
+if os.path.exists('/data'):
+    DB_PATH = '/data/db.sqlite3'
+else:
+    DB_PATH = BASE_DIR / 'db.sqlite3'
+
 DATABASES = {
     'default': dj_database_url.config(
-        default=f"sqlite:///{BASE_DIR / 'db.sqlite3'}",
+        default=f"sqlite:///{DB_PATH}",
         conn_max_age=600,
         conn_health_checks=True,
     )
@@ -174,7 +180,10 @@ DEFAULT_AUTO_FIELD = 'django.db.models.BigAutoField'
 
 
 MEDIA_URL = '/media/'
-MEDIA_ROOT = os.path.join(BASE_DIR, 'media')
+if os.path.exists('/data'):
+    MEDIA_ROOT = '/data/media'
+else:
+    MEDIA_ROOT = os.path.join(BASE_DIR, 'media')
 
 DATA_UPLOAD_MAX_MEMORY_SIZE = 10_485_760  # 10MB — allow phone camera uploads
 

update_prompt.py

@@ -0,0 +1,52 @@
+import re
+
+with open('plant_core/diseases/ml_logic.py', 'r') as f:
+    content = f.read()
+
+replacement = '''
+        if species_hint:
+            prompt = (
+                "You are a plant pathologist analyzing a single leaf image.\\n\\n"
+                f"CONTEXT: An upstream object-detection model tagged this leaf as species '{species_hint}'. "
+                "That model is NOT trained on all crop types and frequently mislabels out-of-distribution leaves. "
+                "Treat the hint as a weak prior only — never as evidence.\\n\\n"
+                "ANALYSIS STEPS (do this reasoning internally, do not output it):\\n"
+                "1. Confirm this is actually a leaf. If it's not a leaf (or not identifiable as one), set is_plant_leaf to false and every other field to null.\\n"
+                f"2. Identify the species independently from leaf shape, margin, venation pattern, and texture. Only agree with '{species_hint}' if your own visual evidence supports it. If you're not confident in the species, say so via a lower confidence score rather than defaulting to the hint.\\n"
+                "3. Inspect the leaf surface for visible symptoms: discoloration, lesions, spots, mold, wilting, curling, chlorosis, necrosis, pest damage, etc.\\n"
+                "4. Decide health status:\\n"
+                "   - No visible symptoms -> disease_name MUST be \\"healthy\\".\\n"
+                "   - Visible symptoms -> name the disease based on the specific visual pattern you observe (e.g. \\"early blight\\", \\"powdery mildew\\", \\"leaf rust\\"), not a generic guess. If symptoms are visible but you can't confidently name a specific disease, use \\"unidentified leaf disease\\" rather than inventing a specific name.\\n"
+                "5. Only fill organic_cure, chemical_cure, and prevention_tips if disease_name is a real (non-null, non-\\"healthy\\") diagnosis you are reasonably confident in. If disease_name is \\"healthy\\" or \\"unidentified leaf disease\\", set these three fields to null instead of fabricating treatment advice.\\n\\n"
+                "CONFIDENCE CALIBRATION:\\n"
+                "- confidence reflects your certainty in the (species + disease) call jointly.\\n"
+                "- Use <0.5 when the image is blurry, ambiguous, or symptoms are non-specific.\\n"
+                "- Do not inflate confidence just because a specific hint was provided.\\n\\n"
+                "Respond with ONLY valid JSON, no markdown, no explanation, no reasoning text.\\n"
+                "Schema:\\n"
+                '{"is_plant_leaf": bool, "plant_species": str or null, "disease_name": str or null, "confidence": float 0-1, "severity": "mild"|"moderate"|"severe"|null, "organic_cure": str or null, "chemical_cure": str or null, "prevention_tips": str or null}'
+            )
+        else:
+            prompt = (
+                "You are a plant pathologist analyzing a single leaf image.\\n\\n"
+                "ANALYSIS STEPS (do this reasoning internally, do not output it):\\n"
+                "1. Confirm this is actually a leaf. If it's not a leaf (or not identifiable as one), set is_plant_leaf to false and every other field to null.\\n"
+                "2. Identify the species independently from leaf shape, margin, venation pattern, and texture. If you're not confident in the species, say so via a lower confidence score.\\n"
+                "3. Inspect the leaf surface for visible symptoms: discoloration, lesions, spots, mold, wilting, curling, chlorosis, necrosis, pest damage, etc.\\n"
+                "4. Decide health status:\\n"
+                "   - No visible symptoms -> disease_name MUST be \\"healthy\\".\\n"
+                "   - Visible symptoms -> name the disease based on the specific visual pattern you observe (e.g. \\"early blight\\", \\"powdery mildew\\", \\"leaf rust\\"), not a generic guess. If symptoms are visible but you can't confidently name a specific disease, use \\"unidentified leaf disease\\" rather than inventing a specific name.\\n"
+                "5. Only fill organic_cure, chemical_cure, and prevention_tips if disease_name is a real (non-null, non-\\"healthy\\") diagnosis you are reasonably confident in. If disease_name is \\"healthy\\" or \\"unidentified leaf disease\\", set these three fields to null instead of fabricating treatment advice.\\n\\n"
+                "CONFIDENCE CALIBRATION:\\n"
+                "- confidence reflects your certainty in the (species + disease) call jointly.\\n"
+                "- Use <0.5 when the image is blurry, ambiguous, or symptoms are non-specific.\\n\\n"
+                "Respond with ONLY valid JSON, no markdown, no explanation, no reasoning text.\\n"
+                "Schema:\\n"
+                '{"is_plant_leaf": bool, "plant_species": str or null, "disease_name": str or null, "confidence": float 0-1, "severity": "mild"|"moderate"|"severe"|null, "organic_cure": str or null, "chemical_cure": str or null, "prevention_tips": str or null}'
+            )'''
+
+pattern = r'if species_hint:\s+prompt = \(.*?\)\s+else:\s+prompt = \(.*?\)'
+new_content = re.sub(pattern, replacement.strip(), content, flags=re.DOTALL)
+
+with open('plant_core/diseases/ml_logic.py', 'w') as f:
+    f.write(new_content)