UI - LLM - Fix for settings (wtforms vs pydantic)

LLM - Smarter reasoning budget logic for gemini models
0.55.6
2026-05-31 22:11:05 +00:00 · 2026-05-25 18:17:28 +02:00 · 2026-05-25 18:03:11 +02:00 · 2026-05-25 17:59:18 +02:00 · 2026-05-25 17:57:41 +02:00 · 2026-05-25 16:51:45 +02:00
228 changed files with 58069 additions and 4990 deletions
@@ -0,0 +1,33 @@
+server {
+    listen 80;
+    server_name localhost;
+
+    # Test basic reverse proxy to changedetection.io
+    location / {
+        proxy_pass http://changedet-app:5000;
+        proxy_set_header Host $host;
+        proxy_set_header X-Real-IP $remote_addr;
+        proxy_set_header X-Forwarded-For $proxy_add_x_forwarded_for;
+        proxy_set_header X-Forwarded-Proto $scheme;
+
+        # WebSocket support
+        proxy_http_version 1.1;
+        proxy_set_header Upgrade $http_upgrade;
+        proxy_set_header Connection "upgrade";
+    }
+
+    # Test subpath deployment with X-Forwarded-Prefix
+    location /changedet-sub/ {
+        proxy_pass http://changedet-app:5000/;
+        proxy_set_header X-Forwarded-Prefix /changedet-sub;
+        proxy_set_header Host $host;
+        proxy_set_header X-Real-IP $remote_addr;
+        proxy_set_header X-Forwarded-For $proxy_add_x_forwarded_for;
+        proxy_set_header X-Forwarded-Proto $scheme;
+
+        # WebSocket support
+        proxy_http_version 1.1;
+        proxy_set_header Upgrade $http_upgrade;
+        proxy_set_header Connection "upgrade";
+    }
+}
@@ -66,27 +66,27 @@ jobs:
          echo ${{ github.ref }} > changedetectionio/tag.txt

      - name: Set up QEMU
-        uses: docker/setup-qemu-action@v3
+        uses: docker/setup-qemu-action@v4
        with:
          image: tonistiigi/binfmt:latest
          platforms: all

      - name: Login to GitHub Container Registry
-        uses: docker/login-action@v3
+        uses: docker/login-action@v4
        with:
          registry: ghcr.io
          username: ${{ github.actor }}
          password: ${{ secrets.GITHUB_TOKEN }}

      - name: Login to Docker Hub Container Registry
-        uses: docker/login-action@v3
+        uses: docker/login-action@v4
        with:
          username: ${{ secrets.DOCKER_HUB_USERNAME }}
          password: ${{ secrets.DOCKER_HUB_ACCESS_TOKEN }}

      - name: Set up Docker Buildx
        id: buildx
-        uses: docker/setup-buildx-action@v3
+        uses: docker/setup-buildx-action@v4
        with:
          install: true
          version: latest
@@ -95,7 +95,7 @@ jobs:
      # master branch -> :dev container tag
      - name: Docker meta :dev
        if: ${{ github.ref == 'refs/heads/master' && github.event_name != 'release' }}
-        uses: docker/metadata-action@v5
+        uses: docker/metadata-action@v6
        id: meta_dev
        with:
          images: |
@@ -103,11 +103,19 @@ jobs:
            ghcr.io/${{ github.repository }}
          tags: |
            type=raw,value=dev
+          labels: |
+            org.opencontainers.image.created=${{ github.event.release.published_at }}
+            org.opencontainers.image.description=Website, webpage change detection, monitoring and notifications.
+            org.opencontainers.image.documentation=https://changedetection.io
+            org.opencontainers.image.revision=${{ github.sha }}
+            org.opencontainers.image.source=https://github.com/dgtlmoon/changedetection.io
+            org.opencontainers.image.title=changedetection.io
+            org.opencontainers.image.url=https://changedetection.io

      - name: Build and push :dev
        id: docker_build
        if: ${{ github.ref == 'refs/heads/master' && github.event_name != 'release' }}
-        uses: docker/build-push-action@v6
+        uses: docker/build-push-action@v7
        with:
          context: ./
          file: ./Dockerfile
@@ -128,10 +136,10 @@ jobs:
          echo "Release tag: ${{ github.event.release.tag_name }}"
          echo "Github ref: ${{ github.ref }}"
          echo "Github ref name: ${{ github.ref_name }}"
-          
+
      - name: Docker meta :tag
        if: github.event_name == 'release' && startsWith(github.event.release.tag_name, '0.')
-        uses: docker/metadata-action@v5
+        uses: docker/metadata-action@v6
        id: meta
        with:
            images: |
@@ -142,11 +150,20 @@ jobs:
                type=semver,pattern={{major}}.{{minor}},value=${{ github.event.release.tag_name }}
                type=semver,pattern={{major}},value=${{ github.event.release.tag_name }}
                type=raw,value=latest
+            labels: |
+              org.opencontainers.image.created=${{ github.event.release.published_at }}
+              org.opencontainers.image.description=Website, webpage change detection, monitoring and notifications.
+              org.opencontainers.image.documentation=https://changedetection.io
+              org.opencontainers.image.revision=${{ github.sha }}
+              org.opencontainers.image.source=https://github.com/dgtlmoon/changedetection.io
+              org.opencontainers.image.title=changedetection.io
+              org.opencontainers.image.url=https://changedetection.io
+              org.opencontainers.image.version=${{ github.event.release.tag_name }}

      - name: Build and push :tag
        id: docker_build_tag_release
        if: github.event_name == 'release' && startsWith(github.event.release.tag_name, '0.')
-        uses: docker/build-push-action@v6
+        uses: docker/build-push-action@v7
        with:
          context: ./
          file: ./Dockerfile
@@ -21,7 +21,7 @@ jobs:
    - name: Build a binary wheel and a source tarball
      run: python3 -m build
    - name: Store the distribution packages
-      uses: actions/upload-artifact@v6
+      uses: actions/upload-artifact@v7
      with:
        name: python-package-distributions
        path: dist/
@@ -34,7 +34,7 @@ jobs:
    - build
    steps:
    - name: Download all the dists
-      uses: actions/download-artifact@v7
+      uses: actions/download-artifact@v8
      with:
        name: python-package-distributions
        path: dist/
@@ -93,7 +93,7 @@ jobs:

    steps:
    - name: Download all the dists
-      uses: actions/download-artifact@v7
+      uses: actions/download-artifact@v8
      with:
        name: python-package-distributions
        path: dist/
@@ -60,14 +60,14 @@ jobs:

        # Just test that the build works, some libraries won't compile on ARM/rPi etc
        - name: Set up QEMU
-          uses: docker/setup-qemu-action@v3
+          uses: docker/setup-qemu-action@v4
          with:
            image: tonistiigi/binfmt:latest
            platforms: all

        - name: Set up Docker Buildx
          id: buildx
-          uses: docker/setup-buildx-action@v3
+          uses: docker/setup-buildx-action@v4
          with:
            install: true
            version: latest
@@ -75,7 +75,7 @@ jobs:

        - name: Test that the docker containers can build (${{ matrix.platform }} - ${{ matrix.dockerfile }})
          id: docker_build
-          uses: docker/build-push-action@v6
+          uses: docker/build-push-action@v7
          # https://github.com/docker/build-push-action#customizing
          with:
            context: ./
@@ -11,8 +11,8 @@ jobs:
      - name: Lint with Ruff
        run: |
          pip install ruff
-          # Check for syntax errors and undefined names
-          ruff check . --select E9,F63,F7,F82
+          # Check for syntax errors and undefined names, and gettext misuse
+          ruff check . --select E9,F63,F7,F82,INT
          # Complete check with errors treated as warnings
          ruff check . --exit-zero
      - name: Validate OpenAPI spec
@@ -20,10 +20,93 @@ jobs:
          pip install openapi-spec-validator
          python3 -c "from openapi_spec_validator import validate_spec; import yaml; validate_spec(yaml.safe_load(open('docs/api-spec.yaml')))"

+  lint-translations:
+    runs-on: ubuntu-latest
+    steps:
+      - uses: actions/checkout@v6
+      - name: Check .po files with msgfmt
+        run: |
+          sudo apt-get install -y gettext
+          find changedetectionio/translations -name "*.po" | while read f; do
+            echo "Checking $f"
+            msgfmt --check-format -o /dev/null "$f"
+          done
+      - name: Lint .pot template with dennis
+        run: |
+          pip install "$(grep -E '^dennis ?>=' requirements.txt)"
+          dennis-cmd lint --strict changedetectionio/translations/messages.pot
+      - name: Lint .po files with dennis
+        run: |
+          dennis-cmd lint --strict --excluderules=W302 changedetectionio/translations/*/LC_MESSAGES/messages.po
+        # W302 (unchanged) is excluded due to high false-positive rate in this codebase:
+        # many msgstrs intentionally match msgid (units like "AI", "LLM", and proper nouns).
+      - name: Check translation catalog is up-to-date
+        run: |
+          pip install "$(grep -E '^babel==' requirements.txt)"
+          python setup.py extract_messages
+          python setup.py update_catalog
+          python setup.py compile_catalog
+          # Ignore POT-Creation-Date timestamp lines — they change on every extract_messages run
+          if git diff changedetectionio/translations | grep -v 'POT-Creation-Date' | grep -qE '^[+-][^+-]'; then
+            echo "ERROR: Translation catalog is out of sync. Run: python setup.py extract_messages && python setup.py update_catalog && python setup.py compile_catalog"
+            git diff --stat changedetectionio/translations
+            exit 1
+          fi
+
+  lint-template-i18n:
+    runs-on: ubuntu-latest
+    steps:
+      - uses: actions/checkout@v6
+      - name: Check for fragmented gettext calls in templates
+        run: |
+          python3 << 'PYEOF'
+          import re, sys
+          from pathlib import Path
+
+          # Detects adjacent {{ _(...) }} calls on the same line separated only by HTML
+          # tags, whitespace, or non-translating Jinja2 variables — the anti-pattern of
+          # splitting a single sentence across multiple msgids.
+          # See https://github.com/dgtlmoon/changedetection.io/issues/4074 for background.
+          #
+          # The correct fix is to consolidate fragments into one entire-sentence msgid,
+          # injecting dynamic values via %(name)s kwargs — per the GNU gettext manual
+          # sections "Entire sentences" and "No string concatenation". See PR #4076 for
+          # worked examples of each consolidation pattern.
+          #
+          # BASELINE: this limit reflects pre-existing violations present when this check
+          # was introduced. It must only ever go DOWN. Each time you fix a template, lower
+          # the limit by the number of lines fixed so the improvement is locked in.
+          # When the count reaches 0, replace the baseline check with a hard sys.exit(1).
+          BASELINE_LIMIT = 44
+
+          FRAGMENT_RE = re.compile(
+              r'\{\{[^{}]*\b_\s*\([^)]*\)[^{}]*\}\}'
+              r'(?:\s*(?:<[^>]+>|\{\{(?![^}]*\b_\s*\()[^}]*\}\})\s*)+'
+              r'\{\{[^{}]*\b_\s*\([^)]*\)[^{}]*\}\}'
+          )
+
+          violations = []
+          for f in sorted(Path('changedetectionio').rglob('*.html')):
+              for lineno, line in enumerate(f.read_text().splitlines(), 1):
+                  if FRAGMENT_RE.search(line):
+                      violations.append(f"{f}:{lineno}: {line.strip()[:120]}")
+
+          count = len(violations)
+          print(f"Fragmented i18n calls found: {count} (limit: {BASELINE_LIMIT})")
+          for v in violations:
+              print(v)
+
+          if count > BASELINE_LIMIT:
+              print(f"\nERROR: {count} fragmented gettext calls exceed the baseline of {BASELINE_LIMIT}.")
+              print("Consolidate adjacent _() calls into a single entire-sentence msgid.")
+              print("See https://github.com/dgtlmoon/changedetection.io/issues/4074 for patterns.")
+              sys.exit(1)
+          PYEOF
+
  test-application-3-10:
    # Only run on push to master (including PR merges)
    if: github.event_name == 'push' && github.ref == 'refs/heads/master'
-    needs: lint-code
+    needs: [lint-code, lint-translations, lint-template-i18n]
    uses: ./.github/workflows/test-stack-reusable-workflow.yml
    with:
      python-version: '3.10'
@@ -31,7 +114,7 @@ jobs:

  test-application-3-11:
    # Always run
-    needs: lint-code
+    needs: [lint-code, lint-translations, lint-template-i18n]
    uses: ./.github/workflows/test-stack-reusable-workflow.yml
    with:
      python-version: '3.11'
@@ -39,7 +122,7 @@ jobs:
  test-application-3-12:
    # Only run on push to master (including PR merges)
    if: github.event_name == 'push' && github.ref == 'refs/heads/master'
-    needs: lint-code
+    needs: [lint-code, lint-translations, lint-template-i18n]
    uses: ./.github/workflows/test-stack-reusable-workflow.yml
    with:
      python-version: '3.12'
@@ -48,8 +131,17 @@ jobs:
  test-application-3-13:
    # Only run on push to master (including PR merges)
    if: github.event_name == 'push' && github.ref == 'refs/heads/master'
-    needs: lint-code
+    needs: [lint-code, lint-translations, lint-template-i18n]
    uses: ./.github/workflows/test-stack-reusable-workflow.yml
    with:
      python-version: '3.13'
-      skip-pypuppeteer: true
+      skip-pypuppeteer: true
+
+
+  test-application-3-14:
+    #if: github.event_name == 'push' && github.ref == 'refs/heads/master'
+    needs: [lint-code, lint-translations, lint-template-i18n]
+    uses: ./.github/workflows/test-stack-reusable-workflow.yml
+    with:
+      python-version: '3.14'
+      skip-pypuppeteer: false
@@ -42,10 +42,10 @@ jobs:
        run: echo "date=$(date +'%Y-%m-%d')" >> $GITHUB_OUTPUT

      - name: Set up Docker Buildx
-        uses: docker/setup-buildx-action@v3
+        uses: docker/setup-buildx-action@v4

      - name: Build changedetection.io container for testing under Python ${{ env.PYTHON_VERSION }}
-        uses: docker/build-push-action@v6
+        uses: docker/build-push-action@v7
        with:
          context: ./
          file: ./Dockerfile
@@ -71,7 +71,7 @@ jobs:
          docker save test-changedetectionio -o /tmp/test-changedetectionio.tar

      - name: Upload Docker image artifact
-        uses: actions/upload-artifact@v6
+        uses: actions/upload-artifact@v7
        with:
          name: test-changedetectionio-${{ env.PYTHON_VERSION }}
          path: /tmp/test-changedetectionio.tar
@@ -88,7 +88,7 @@ jobs:
      - uses: actions/checkout@v6

      - name: Download Docker image artifact
-        uses: actions/download-artifact@v7
+        uses: actions/download-artifact@v8
        with:
          name: test-changedetectionio-${{ env.PYTHON_VERSION }}
          path: /tmp
@@ -99,11 +99,7 @@ jobs:

      - name: Run Unit Tests
        run: |
-          docker run test-changedetectionio bash -c 'python3 -m unittest changedetectionio.tests.unit.test_notification_diff'
-          docker run test-changedetectionio bash -c 'python3 -m unittest changedetectionio.tests.unit.test_watch_model'
-          docker run test-changedetectionio bash -c 'python3 -m unittest changedetectionio.tests.unit.test_jinja2_security'
-          docker run test-changedetectionio bash -c 'python3 -m unittest changedetectionio.tests.unit.test_semver'
-          docker run test-changedetectionio bash -c 'python3 -m unittest changedetectionio.tests.unit.test_html_to_text'         
+          docker run test-changedetectionio bash -c 'cd changedetectionio;pytest tests/unit/ tests/llm/'

  # Basic pytest tests with ancillary services
  basic-tests:
@@ -116,7 +112,7 @@ jobs:
      - uses: actions/checkout@v6

      - name: Download Docker image artifact
-        uses: actions/download-artifact@v7
+        uses: actions/download-artifact@v8
        with:
          name: test-changedetectionio-${{ env.PYTHON_VERSION }}
          path: /tmp
@@ -165,14 +161,14 @@ jobs:

      - name: Store test artifacts
        if: always()
-        uses: actions/upload-artifact@v6
+        uses: actions/upload-artifact@v7
        with:
          name: test-cdio-basic-tests-output-py${{ env.PYTHON_VERSION }}
          path: output-logs

      - name: Store CLI test output
        if: always()
-        uses: actions/upload-artifact@v6
+        uses: actions/upload-artifact@v7
        with:
          name: test-cdio-cli-opts-output-py${{ env.PYTHON_VERSION }}
          path: cli-opts-output.txt
@@ -188,7 +184,7 @@ jobs:
      - uses: actions/checkout@v6

      - name: Download Docker image artifact
-        uses: actions/download-artifact@v7
+        uses: actions/download-artifact@v8
        with:
          name: test-changedetectionio-${{ env.PYTHON_VERSION }}
          path: /tmp
@@ -230,7 +226,7 @@ jobs:
      - uses: actions/checkout@v6

      - name: Download Docker image artifact
-        uses: actions/download-artifact@v7
+        uses: actions/download-artifact@v8
        with:
          name: test-changedetectionio-${{ env.PYTHON_VERSION }}
          path: /tmp
@@ -270,7 +266,7 @@ jobs:
      - uses: actions/checkout@v6

      - name: Download Docker image artifact
-        uses: actions/download-artifact@v7
+        uses: actions/download-artifact@v8
        with:
          name: test-changedetectionio-${{ env.PYTHON_VERSION }}
          path: /tmp
@@ -306,7 +302,7 @@ jobs:
      - uses: actions/checkout@v6

      - name: Download Docker image artifact
-        uses: actions/download-artifact@v7
+        uses: actions/download-artifact@v8
        with:
          name: test-changedetectionio-${{ env.PYTHON_VERSION }}
          path: /tmp
@@ -324,6 +320,175 @@ jobs:
        run: |
          docker run --rm --network changedet-network test-changedetectionio bash -c 'cd changedetectionio;pytest tests/smtp/test_notification_smtp.py'

+  nginx-reverse-proxy:
+    runs-on: ubuntu-latest
+    needs: build
+    timeout-minutes: 10
+    env:
+      PYTHON_VERSION: ${{ inputs.python-version }}
+    steps:
+      - uses: actions/checkout@v6
+
+      - name: Download Docker image artifact
+        uses: actions/download-artifact@v8
+        with:
+          name: test-changedetectionio-${{ env.PYTHON_VERSION }}
+          path: /tmp
+
+      - name: Load Docker image
+        run: |
+          docker load -i /tmp/test-changedetectionio.tar
+
+      - name: Spin up services
+        run: |
+          docker network create changedet-network
+
+          # Start changedetection.io container with X-Forwarded headers support
+          docker run --name changedet-app --hostname changedet-app --network changedet-network \
+            -e USE_X_SETTINGS=true \
+            -d test-changedetectionio
+          sleep 3
+
+      - name: Start nginx reverse proxy
+        run: |
+          # Start nginx with our test configuration
+          docker run --name nginx-proxy --network changedet-network -d -p 8080:80 --rm \
+            -v ${{ github.workspace }}/.github/nginx-reverse-proxy-test.conf:/etc/nginx/conf.d/default.conf:ro \
+            nginx:alpine
+          sleep 2
+
+      - name: Test reverse proxy - root path
+        run: |
+          echo "=== Testing nginx reverse proxy at root path ==="
+          curl --retry-connrefused --retry 6 -s http://localhost:8080/ > /tmp/nginx-test-root.html
+
+          # Check for changedetection.io UI elements
+          if grep -q "checkbox-uuid" /tmp/nginx-test-root.html; then
+            echo "✓ Found checkbox-uuid in response"
+          else
+            echo "ERROR: checkbox-uuid not found in response"
+            cat /tmp/nginx-test-root.html
+            exit 1
+          fi
+
+          # Check for watchlist content
+          if grep -q -i "watch" /tmp/nginx-test-root.html; then
+            echo "✓ Found watch/watchlist content in response"
+          else
+            echo "ERROR: watchlist content not found"
+            cat /tmp/nginx-test-root.html
+            exit 1
+          fi
+
+          echo "✓ Root path reverse proxy working correctly"
+
+      - name: Test reverse proxy - subpath with X-Forwarded-Prefix
+        run: |
+          echo "=== Testing nginx reverse proxy at subpath /changedet-sub/ ==="
+          curl --retry-connrefused --retry 6 -s http://localhost:8080/changedet-sub/ > /tmp/nginx-test-subpath.html
+
+          # Check for changedetection.io UI elements
+          if grep -q "checkbox-uuid" /tmp/nginx-test-subpath.html; then
+            echo "✓ Found checkbox-uuid in subpath response"
+          else
+            echo "ERROR: checkbox-uuid not found in subpath response"
+            cat /tmp/nginx-test-subpath.html
+            exit 1
+          fi
+
+          echo "✓ Subpath reverse proxy working correctly"
+
+      - name: Test API through reverse proxy subpath
+        run: |
+          echo "=== Testing API endpoints through nginx subpath /changedet-sub/ ==="
+
+          # Extract API key from the changedetection.io datastore
+          API_KEY=$(docker exec changedet-app cat /datastore/changedetection.json | grep -o '"api_access_token": *"[^"]*"' | cut -d'"' -f4)
+
+          if [ -z "$API_KEY" ]; then
+            echo "ERROR: Could not extract API key from datastore"
+            docker exec changedet-app cat /datastore/changedetection.json
+            exit 1
+          fi
+
+          echo "✓ Extracted API key: ${API_KEY:0:8}..."
+
+          # Create a watch via API through nginx proxy subpath
+          echo "Creating watch via POST to /changedet-sub/api/v1/watch"
+          RESPONSE=$(curl -s -w "\n%{http_code}" -X POST "http://localhost:8080/changedet-sub/api/v1/watch" \
+            -H "x-api-key: ${API_KEY}" \
+            -H "Content-Type: application/json" \
+            -d '{
+              "url": "https://example.com/test-nginx-proxy",
+              "tag": "nginx-test"
+            }')
+
+          HTTP_CODE=$(echo "$RESPONSE" | tail -n1)
+          BODY=$(echo "$RESPONSE" | head -n-1)
+
+          if [ "$HTTP_CODE" != "201" ]; then
+            echo "ERROR: Expected HTTP 201, got $HTTP_CODE"
+            echo "Response: $BODY"
+            exit 1
+          fi
+
+          echo "✓ Watch created successfully (HTTP 201)"
+
+          # Extract the watch UUID from response
+          WATCH_UUID=$(echo "$BODY" | grep -o '"uuid": *"[^"]*"' | cut -d'"' -f4)
+          echo "✓ Watch UUID: $WATCH_UUID"
+
+          # Update the watch via PUT through nginx proxy subpath
+          echo "Updating watch via PUT to /changedet-sub/api/v1/watch/${WATCH_UUID}"
+          RESPONSE=$(curl -s -w "\n%{http_code}" -X PUT "http://localhost:8080/changedet-sub/api/v1/watch/${WATCH_UUID}" \
+            -H "x-api-key: ${API_KEY}" \
+            -H "Content-Type: application/json" \
+            -d '{
+              "paused": true
+            }')
+
+          HTTP_CODE=$(echo "$RESPONSE" | tail -n1)
+          BODY=$(echo "$RESPONSE" | head -n-1)
+
+          if [ "$HTTP_CODE" != "200" ]; then
+            echo "ERROR: Expected HTTP 200, got $HTTP_CODE"
+            echo "Response: $BODY"
+            exit 1
+          fi
+
+          if echo "$BODY" | grep -q 'OK'; then
+            echo "✓ Watch updated successfully (HTTP 200, response: OK)"
+          else
+            echo "ERROR: Expected response 'OK', got: $BODY"
+            echo "Response: $BODY"
+            exit 1
+          fi
+
+          # Verify the watch is paused via GET
+          echo "Verifying watch is paused via GET"
+          RESPONSE=$(curl -s "http://localhost:8080/changedet-sub/api/v1/watch/${WATCH_UUID}" \
+            -H "x-api-key: ${API_KEY}")
+
+          if echo "$RESPONSE" | grep -q '"paused": *true'; then
+            echo "✓ Watch is paused as expected"
+          else
+            echo "ERROR: Watch paused state not confirmed"
+            echo "Response: $RESPONSE"
+            exit 1
+          fi
+
+          echo "✓ API tests through nginx subpath completed successfully"
+
+      - name: Cleanup nginx test
+        if: always()
+        run: |
+          docker logs nginx-proxy || true
+          docker logs changedet-app || true
+          docker stop nginx-proxy changedet-app || true
+          docker rm nginx-proxy changedet-app || true
+
+
+
  # Proxy tests
  proxy-tests:
    runs-on: ubuntu-latest
@@ -335,7 +500,7 @@ jobs:
      - uses: actions/checkout@v6

      - name: Download Docker image artifact
-        uses: actions/download-artifact@v7
+        uses: actions/download-artifact@v8
        with:
          name: test-changedetectionio-${{ env.PYTHON_VERSION }}
          path: /tmp
@@ -375,7 +540,7 @@ jobs:
      - uses: actions/checkout@v6

      - name: Download Docker image artifact
-        uses: actions/download-artifact@v7
+        uses: actions/download-artifact@v8
        with:
          name: test-changedetectionio-${{ env.PYTHON_VERSION }}
          path: /tmp
@@ -405,7 +570,7 @@ jobs:
      - uses: actions/checkout@v6

      - name: Download Docker image artifact
-        uses: actions/download-artifact@v7
+        uses: actions/download-artifact@v8
        with:
          name: test-changedetectionio-${{ env.PYTHON_VERSION }}
          path: /tmp
@@ -418,6 +583,10 @@ jobs:
        run: |
          docker run  -e EXTRA_PACKAGES=changedetection.io-osint-processor test-changedetectionio bash -c 'cd changedetectionio;pytest -vvv -s  tests/plugins/test_processor.py::test_check_plugin_processor'

+      - name: Plugin get_html_head_extras hook injects into base.html
+        run: |
+          docker run test-changedetectionio bash -c 'cd changedetectionio;pytest -vvv -s tests/plugins/test_html_head_extras.py'
+
  # Container startup tests
  container-tests:
    runs-on: ubuntu-latest
@@ -429,7 +598,7 @@ jobs:
      - uses: actions/checkout@v6

      - name: Download Docker image artifact
-        uses: actions/download-artifact@v7
+        uses: actions/download-artifact@v8
        with:
          name: test-changedetectionio-${{ env.PYTHON_VERSION }}
          path: /tmp
@@ -474,7 +643,7 @@ jobs:
      - uses: actions/checkout@v6

      - name: Download Docker image artifact
-        uses: actions/download-artifact@v7
+        uses: actions/download-artifact@v8
        with:
          name: test-changedetectionio-${{ env.PYTHON_VERSION }}
          path: /tmp
@@ -537,7 +706,19 @@ jobs:
      - name: Check upgrade works without error
        run: |
          echo "=== Testing upgrade path from 0.49.1 to ${{ github.ref_name }} (${{ github.sha }}) ==="
-
+          sudo apt-get update && sudo apt-get install -y --no-install-recommends \
+              g++ \
+              gcc \
+              libc-dev \
+              libffi-dev \
+              libjpeg-dev \
+              libssl-dev \
+              libxslt-dev \
+              make \
+              patch \
+              pkg-config \
+              zlib1g-dev
+          
          # Checkout old version and create datastore
          git checkout 0.49.1
          python3 -m venv .venv
@@ -546,7 +727,7 @@ jobs:
          pip install 'pyOpenSSL>=23.2.0'

          echo "=== Running version 0.49.1 to create datastore ==="
-          python3 ./changedetection.py -C -d /tmp/data &
+          ALLOW_IANA_RESTRICTED_ADDRESSES=true python3 ./changedetection.py -C -d /tmp/data &
          APP_PID=$!

          # Wait for app to be ready
@@ -594,7 +775,7 @@ jobs:
          pip install -r requirements.txt

          echo "=== Running current version (commit ${{ github.sha }}) with old datastore (testing mode) ==="
-          TESTING_SHUTDOWN_AFTER_DATASTORE_LOAD=1 python3 ./changedetection.py -d /tmp/data > /tmp/upgrade-test.log 2>&1
+          ALLOW_IANA_RESTRICTED_ADDRESSES=true TESTING_SHUTDOWN_AFTER_DATASTORE_LOAD=1 python3 ./changedetection.py -d /tmp/data > /tmp/upgrade-test.log 2>&1

          echo "=== Upgrade test output ==="
          cat /tmp/upgrade-test.log
@@ -602,7 +783,7 @@ jobs:

          # Now start the current version normally to verify the tag survived
          echo "=== Starting current version to verify tag exists after upgrade ==="
-          timeout 20 python3 ./changedetection.py -d /tmp/data > /tmp/ui-test.log 2>&1 &
+          ALLOW_IANA_RESTRICTED_ADDRESSES=true timeout 20 python3 ./changedetection.py -d /tmp/data > /tmp/ui-test.log 2>&1 &
          APP_PID=$!

          # Wait for app to be ready and fetch UI
@@ -651,7 +832,7 @@ jobs:

      - name: Upload upgrade test logs
        if: always()
-        uses: actions/upload-artifact@v6
+        uses: actions/upload-artifact@v7
        with:
          name: upgrade-test-logs-py${{ env.PYTHON_VERSION }}
          path: /tmp/upgrade-test.log
@@ -7,3 +7,19 @@ repos:
        args: [--fix]
      # Fomrat
      - id: ruff-format
+
+  - repo: local
+    hooks:
+      - id: dennis-lint-pot
+        name: dennis lint pot
+        language: system
+        entry: dennis-cmd lint --strict
+        files: ^changedetectionio/translations/messages\.pot$
+        pass_filenames: true
+
+      - id: dennis-lint-po
+        name: dennis lint po
+        language: system
+        entry: dennis-cmd lint --strict --excluderules=W302
+        files: ^changedetectionio/translations/\w+/LC_MESSAGES/messages\.po$
+        pass_filenames: true
@@ -20,10 +20,11 @@ exclude = [
 select = [
    "B", # flake8-bugbear
    "B9",
-    "C", 
+    "C",
    "E", # pycodestyle
    "F", # Pyflakes
    "I", # isort
+    "INT", # flake8-gettext
    "N", # pep8-naming
    "UP", # pyupgrade
    "W", # pycodestyle
@@ -43,6 +44,9 @@ ignore = [
 [lint.mccabe]
 max-complexity = 12

+[lint.flake8-gettext]
+extend-function-names = ["_l", "lazy_gettext", "pgettext", "npgettext"]
+
 [format]
 indent-style = "space"
 quote-style = "preserve"
@@ -22,6 +22,20 @@ Ideal for monitoring price changes, content edits, conditional changes and more.
 - Get started watching and receiving website change notifications straight away.
 - See our [tutorials and how-to page for more inspiration](https://changedetection.io/tutorials) 

+## AI-powered website change detection — smart alerts and plain-language summaries
+
+Stop drowning in noise. Connect any LLM (OpenAI, Gemini, Anthropic, Ollama and more) and go from _"something changed"_ to _"only the thing you care about changed"_.
+
+**AI change detection rules** — write a plain-English intent once: _"notify me only when the price drops below $50"_, _"alert me when the item comes back in stock"_, _"ignore navigation and footer changes"_. The AI evaluates every detected diff against your intent and silently suppresses everything irrelevant. Fewer false positives, zero noise.
+
+**AI change summaries** — instead of staring at a raw diff, your notification reads _"Price dropped from $89.99 to $67.00"_ or _"3 new products added to the listing"_. Works globally or per-watch, with full control over the prompt.
+
+Works with any model you already pay for — GPT-4o-mini and Gemini Flash handle this well at fractions of a cent per check. Or run it entirely locally with **Ollama**, **vLLM**, **LM Studio**, or any **OpenAI-compatible self-hosted endpoint** — pick the *OpenAI-compatible (vLLM, LM Studio, llama.cpp)* option in the provider dropdown and point it at your server's `/v1` URL. Powered by [LiteLLM](https://github.com/BerriAI/litellm), giving you seamless access to [100+ supported providers and models](https://docs.litellm.ai/docs/providers).
+
+[<img src="./docs/LLM-change-summary.jpeg" style="max-width:100%;" alt="AI-powered website change detection — plain language change summaries and smart alert rules"  title="AI website change detection with LLM change summaries and intelligent alert filtering" />](https://changedetection.io?src=github)
+
+_Note: Available in our subscription/hosted service from June 2026_
+
 ### Target specific parts of the webpage using the Visual Selector tool.

 Available when connected to a <a href="https://github.com/dgtlmoon/changedetection.io/wiki/Playwright-content-fetcher">playwright content fetcher</a> (included as part of our subscription service)
@@ -308,9 +322,27 @@ I offer commercial support, this software is depended on by network security, ae
 [release-link]: https://github.com/dgtlmoon/changedetection.io/releases
 [docker-link]: https://hub.docker.com/r/dgtlmoon/changedetection.io

-## Commercial Licencing

-If you are reselling this software either in part or full as part of any commercial arrangement, you must abide by our COMMERCIAL_LICENCE.md found in our code repository, please contact dgtlmoon@gmail.com and contact@changedetection.io .
+## Disclaimer
+
+**This software is provided "as-is", without warranty of any kind, express or implied, including but not limited to the warranties of merchantability, fitness for a particular purpose, and non-infringement. In no event shall the authors or copyright holders be liable for any claim, damages, or other liability, whether in an action of contract, tort, or otherwise, arising from, out of, or in connection with the software or the use or other dealings in the software.**
+
+### Website content monitoring
+
+You are solely responsible for ensuring that your use of this software complies with the terms of service, `robots.txt` directives, access policies, and all applicable laws of any website or service you choose to monitor. The authors and contributors of this software accept no liability whatsoever for how you choose to use it or for any consequences arising from that use.
+
+### AI / LLM features
+
+If you choose to enable AI / LLM features, content detected on monitored websites — including page diffs and extracted text — will be transmitted to a third-party AI provider of your choosing, outside of this installation. You are solely responsible for:
+
+- Ensuring such transmission is permitted by the terms of service of every website you monitor.
+- Compliance with all applicable data-protection and privacy laws (including but not limited to GDPR) with respect to any personal data that may appear in monitored content.
+- All API costs and charges levied by your chosen AI provider. This software has no visibility into or control over those charges.
+- Any consequences arising from acting on AI-generated output.
+
+**AI and LLM models are known to hallucinate** — producing plausible-sounding but factually incorrect, incomplete, or entirely fabricated output with apparent confidence. By design, LLMs may also omit or silently truncate relevant information during summarisation. **AI output must never be relied upon as complete or accurate.**
+
+By using this software, and in particular any AI / LLM features, you personally indemnify and hold harmless the author(s), contributor(s), and any associated parties from and against any and all claims, damages, losses, costs, and expenses (including reasonable legal fees) arising out of or in connection with your use of this software.

 ## Third-party licenses

@@ -320,4 +352,6 @@ changedetectionio.html_tools.elementpath_tostring: Copyright (c), 2018-2021, SIS

 Recognition of fantastic contributors to the project

+<sub>Developer note: see [translation guide](changedetectionio/translations/README.md) for i18n template patterns and workflow.</sub>
+
 - Constantin Hong https://github.com/Constantin1489
@@ -1,5 +1,6 @@
 [python: **.py]
-keywords = _:1,_l:1,gettext:1
+keywords = _ _l gettext pgettext:1c,2

 [jinja2: **/templates/**.html]
 encoding = utf-8
+keywords = _ _l gettext pgettext:1c,2
@@ -2,7 +2,7 @@

 # Read more https://github.com/dgtlmoon/changedetection.io/wiki
 # Semver means never use .01, or 00. Should be .1.
-__version__ = '0.53.1'
+__version__ = '0.55.6'

 from changedetectionio.strtobool import strtobool
 from json.decoder import JSONDecodeError
@@ -61,8 +61,22 @@ import time
 # ==============================================================================

 import multiprocessing
+import os
 import sys

+# Limit glibc malloc arena count to prevent RSS growth from concurrent requests.
+# Default: glibc creates up to 8×CPU_cores arenas. Each concurrent thread/connection
+# can trigger a new arena, and freed memory stays mapped in those arenas as RSS forever.
+# With MALLOC_ARENA_MAX=2, at most 2 arenas are used; freed pages return to the OS faster.
+# Must be set before worker threads start; env var is read lazily by glibc on first arena creation.
+if 'MALLOC_ARENA_MAX' not in os.environ:
+    os.environ['MALLOC_ARENA_MAX'] = '2'
+    try:
+        import ctypes as _ctypes
+        _ctypes.CDLL('libc.so.6').mallopt(-8, 2)  # M_ARENA_MAX = -8
+    except Exception:
+        pass
+
 # Set spawn as global default (safety net - all our code uses explicit contexts anyway)
 # Skip in tests to avoid breaking pytest-flask's LiveServer fixture (uses unpicklable local functions)
 if 'pytest' not in sys.modules:
@@ -386,8 +400,11 @@ def main():
        datastore.data['settings']['application']['all_paused'] = all_paused
        logger.info(f"Setting all watches paused: {all_paused}")

+    # Register built-in restock plugins (deferred here to avoid circular imports at module load time)
+    from changedetectionio.pluggy_interface import inject_datastore_into_plugins, register_builtin_restock_plugins
+    register_builtin_restock_plugins()
+
    # Inject datastore into plugins that need access to settings
-    from changedetectionio.pluggy_interface import inject_datastore_into_plugins
    inject_datastore_into_plugins(datastore)

    # Step 1: Add URLs with their options (if provided via -u flags)
@@ -610,12 +627,14 @@ def main():

    @app.context_processor
    def inject_template_globals():
-        return dict(right_sticky="v{}".format(datastore.data['version_tag']),
+        from changedetectionio.llm.evaluator import get_llm_config as _get_llm_config
+        return dict(right_sticky="v"+__version__,
                    new_version_available=app.config['NEW_VERSION_AVAILABLE'],
                    has_password=datastore.data['settings']['application']['password'] != False,
                    socket_io_enabled=datastore.data['settings']['application'].get('ui', {}).get('socket_io_enabled', True),
                    all_paused=datastore.data['settings']['application'].get('all_paused', False),
-                    all_muted=datastore.data['settings']['application'].get('all_muted', False)
+                    all_muted=datastore.data['settings']['application'].get('all_muted', False),
+                    llm_configured=bool(_get_llm_config(datastore)),
                    )

    # Monitored websites will not receive a Referer header when a user clicks on an outgoing link.
@@ -0,0 +1,21 @@
+import functools
+from flask import make_response
+from flask_restful import Resource
+
+
+@functools.cache
+def _get_spec_yaml():
+    """Build and cache the merged spec as a YAML string (only serialized once per process)."""
+    import yaml
+    from changedetectionio.api import build_merged_spec_dict
+    return yaml.dump(build_merged_spec_dict(), default_flow_style=False, allow_unicode=True)
+
+
+class Spec(Resource):
+    def get(self):
+        """Return the merged OpenAPI spec including all registered processor extensions."""
+        return make_response(
+            _get_spec_yaml(),
+            200,
+            {'Content-Type': 'application/yaml'}
+        )
@@ -7,7 +7,7 @@ import threading
 from flask import request
 from . import auth

-from . import validate_openapi_request
+from . import validate_openapi_request, strip_internal_api_fields


 class Tag(Resource):
@@ -17,7 +17,7 @@ class Tag(Resource):
        self.update_q = kwargs['update_q']

    # Get information about a single tag
-    # curl http://localhost:5000/api/v1/tag/<string:uuid>
+    # curl http://localhost:5000/api/v1/tag/<uuid_str:uuid>
    @auth.check_token
    @validate_openapi_request('getTag')
    def get(self, uuid):
@@ -85,7 +85,8 @@ class Tag(Resource):
        # Create clean tag dict without Watch-specific fields
        clean_tag = {k: v for k, v in tag.items() if k not in watch_only_fields}

-        return clean_tag
+        # Never expose `__`-prefixed transient/internal fields
+        return strip_internal_api_fields(clean_tag)

    @auth.check_token
    @validate_openapi_request('deleteTag')
@@ -97,17 +98,6 @@ class Tag(Resource):
        # Delete the tag, and any tag reference
        del self.datastore.data['settings']['application']['tags'][uuid]

-        # Delete tag.json file if it exists
-        import os
-        tag_dir = os.path.join(self.datastore.datastore_path, uuid)
-        tag_json = os.path.join(tag_dir, "tag.json")
-        if os.path.exists(tag_json):
-            try:
-                os.unlink(tag_json)
-                logger.info(f"Deleted tag.json for tag {uuid}")
-            except Exception as e:
-                logger.error(f"Failed to delete tag.json for tag {uuid}: {e}")
-
        # Remove tag from all watches
        for watch_uuid, watch in self.datastore.data['watching'].items():
            if watch.get('tags') and uuid in watch['tags']:
@@ -124,8 +114,9 @@ class Tag(Resource):
        if not tag:
            abort(404, message='No tag exists with the UUID of {}'.format(uuid))

-        # Make a mutable copy of request.json for modification
-        json_data = dict(request.json)
+        # Make a mutable copy of request.json for modification.
+        # Silently discard `__`-prefixed transient/internal keys (not part of the public schema).
+        json_data = strip_internal_api_fields(dict(request.json))

        # Validate notification_urls if provided
        if 'notification_urls' in json_data:
@@ -173,7 +164,8 @@ class Tag(Resource):
    def post(self):
        """Create a single tag/group."""

-        json_data = request.get_json()
+        # Silently discard `__`-prefixed transient/internal keys (not part of the public schema).
+        json_data = strip_internal_api_fields(request.get_json())
        title = json_data.get("title",'').strip()

        # Validate that only valid fields are provided
@@ -188,6 +180,13 @@ class Tag(Resource):

        new_uuid = self.datastore.add_tag(title=title)
        if new_uuid:
+            # Apply any extra fields (e.g. processor_config_restock_diff) beyond just title
+            extra = {k: v for k, v in json_data.items() if k != 'title'}
+            if extra:
+                tag = self.datastore.data['settings']['application']['tags'].get(new_uuid)
+                if tag:
+                    tag.update(extra)
+                    tag.commit()
            return {'uuid': new_uuid}, 201
        else:
            return "Invalid or unsupported tag", 400
@@ -1,4 +1,5 @@
 import os
+import re
 import threading

 from changedetectionio.validate_url import is_safe_valid_url
@@ -12,7 +13,7 @@ from flask_restful import abort, Resource
 from loguru import logger
 import copy

-from . import validate_openapi_request, get_readonly_watch_fields
+from . import validate_openapi_request, get_readonly_watch_fields, strip_internal_api_fields
 from ..notification import valid_notification_formats
 from ..notification.handler import newline_re

@@ -57,7 +58,7 @@ class Watch(Resource):
        self.update_q = kwargs['update_q']

    # Get information about a single watch, excluding the history list (can be large)
-    # curl http://localhost:5000/api/v1/watch/<string:uuid>
+    # curl http://localhost:5000/api/v1/watch/<uuid_str:uuid>
    # @todo - version2 - ?muted and ?paused should be able to be called together, return the watch struct not "OK"
    # ?recheck=true
    @auth.check_token
@@ -103,9 +104,31 @@ class Watch(Resource):
        # attr .last_changed will check for the last written text snapshot on change
        watch['last_changed'] = watch_obj.last_changed
        watch['viewed'] = watch_obj.viewed
-        watch['link'] = watch_obj.link,
+        watch['link'] = watch_obj.link

-        return watch
+        # Resolved processor config: tag override wins over watch-level config (mirrors restock processor logic)
+        import json
+        _restock_path = os.path.join(watch_obj.data_dir, 'restock_diff.json') if watch_obj.data_dir else None
+        restock_config = {}
+        if _restock_path and os.path.isfile(_restock_path):
+            try:
+                with open(_restock_path, 'r', encoding='utf-8') as _f:
+                    restock_config = json.load(_f).get('restock_diff') or {}
+            except (json.JSONDecodeError, IOError) as e:
+                logger.warning(f"Failed to read restock_diff.json for watch {uuid}: {e}")
+        restock_source = 'watch'
+        tags = self.datastore.data['settings']['application'].get('tags', {})
+        for tag_uuid in (watch_obj.get('tags') or []):
+            tag = tags.get(tag_uuid, {})
+            if tag.get('overrides_watch'):
+                restock_config = dict(tag.get('processor_config_restock_diff') or {})
+                restock_source = f'tag:{tag_uuid}'
+                break
+        watch['processor_config_restock_diff'] = restock_config
+        watch['processor_config_restock_diff_source'] = restock_source
+
+        # Never expose `__`-prefixed transient/internal fields (e.g. __check_status)
+        return strip_internal_api_fields(watch)

    @auth.check_token
    @validate_openapi_request('deleteWatch')
@@ -166,8 +189,10 @@ class Watch(Resource):
        # Handle processor-config-* fields separately (save to JSON, not datastore)
        from changedetectionio import processors

-        # Make a mutable copy of request.json for modification
-        json_data = dict(request.json)
+        # Make a mutable copy of request.json for modification.
+        # Silently discard `__`-prefixed transient/internal keys — they are not part of the
+        # public schema and must never be writable (e.g. clients that round-trip GET → PUT).
+        json_data = strip_internal_api_fields(dict(request.json))

        # Extract and remove processor config fields from json_data
        processor_config_data = processors.extract_processor_config_from_form_data(json_data)
@@ -217,7 +242,7 @@ class WatchHistory(Resource):
        self.datastore = kwargs['datastore']

    # Get a list of available history for a watch by UUID
-    # curl http://localhost:5000/api/v1/watch/<string:uuid>/history
+    # curl http://localhost:5000/api/v1/watch/<uuid_str:uuid>/history
    @auth.check_token
    @validate_openapi_request('getWatchHistory')
    def get(self, uuid):
@@ -254,8 +279,28 @@ class WatchSingleHistory(Resource):
        if request.args.get('html'):
            content = watch.get_fetched_html(timestamp)
            if content:
+                # XSS mitigation (GHSA-cgj8-g98g-4p9x): this is an API endpoint, not a
+                # browser-rendered view. The bytes ARE HTML (that's what the caller asked
+                # for) but a programmatic client doesn't need text/html — and serving
+                # text/html lets attacker-planted <script> in a monitored site execute
+                # in our origin if someone opens the URL in a browser.
+                #
+                # text/plain + explicit utf-8 + nosniff = browser shows inert text,
+                # sniffing can't re-classify it as HTML, an absent charset can't be
+                # auto-detected as UTF-7 (an alternative XSS vector). API clients
+                # still get the raw bytes — they don't care about Content-Type.
                response = make_response(content, 200)
-                response.mimetype = "text/html"
+                response.headers['Content-Type'] = 'text/plain; charset=utf-8'
+                response.headers['X-Content-Type-Options'] = 'nosniff'
+                # Include the timestamp in the download name so downloading multiple
+                # snapshots doesn't collide. No extension — the stored bytes are
+                # "whatever the fetcher captured" (HTML, JSON, XML, text…), so
+                # claiming .html on the download would be a false content-type label
+                # for non-HTML watches. The user/curl can rename if needed.
+                # Strip to safe filename chars (timestamp is already validated as a
+                # watch.history key — this is defense in depth against header injection).
+                safe_ts = re.sub(r'[^0-9A-Za-z_-]', '', str(timestamp))[:32] or 'snapshot'
+                response.headers['Content-Disposition'] = f'attachment; filename="snapshot-{safe_ts}"'
            else:
                response = make_response("No content found", 404)
                response.mimetype = "text/plain"
@@ -338,7 +383,7 @@ class WatchHistoryDiff(Resource):
            word_diff = True

        # Get boolean diff preferences with defaults from DIFF_PREFERENCES_CONFIG
-        changes_only = strtobool(request.args.get('changesOnly', 'true'))
+        changes_only = strtobool(request.args.get('changesOnly', 'false'))
        ignore_whitespace = strtobool(request.args.get('ignoreWhitespace', 'false'))
        include_removed = strtobool(request.args.get('removed', 'true'))
        include_added = strtobool(request.args.get('added', 'true'))
@@ -349,7 +394,7 @@ class WatchHistoryDiff(Resource):
            previous_version_file_contents=from_version_file_contents,
            newest_version_file_contents=to_version_file_contents,
            ignore_junk=ignore_whitespace,
-            include_equal=changes_only,
+            include_equal=not changes_only,
            include_removed=include_removed,
            include_added=include_added,
            include_replaced=include_replaced,
@@ -422,7 +467,8 @@ class CreateWatch(Resource):
    def post(self):
        """Create a single watch."""

-        json_data = request.get_json()
+        # Silently discard `__`-prefixed transient/internal keys (not part of the public schema).
+        json_data = strip_internal_api_fields(request.get_json())
        url = json_data['url'].strip()

        if not is_safe_valid_url(url):
@@ -567,4 +613,4 @@ class CreateWatch(Resource):

                return {'status': f'OK, queueing {len(watches_to_queue)} watches in background'}, 202

-        return list, 200
+        return list, 200
@@ -3,29 +3,18 @@ from flask import request, abort
 from loguru import logger

@functools.cache
-def get_openapi_spec():
-    """Lazy load OpenAPI spec and dependencies only when validation is needed."""
-    import os
-    import yaml  # Lazy import - only loaded when API validation is actually used
-    from openapi_core import OpenAPI  # Lazy import - saves ~10.7 MB on startup
-
-    spec_path = os.path.join(os.path.dirname(__file__), '../../docs/api-spec.yaml')
-    if not os.path.exists(spec_path):
-        # Possibly for pip3 packages
-        spec_path = os.path.join(os.path.dirname(__file__), '../docs/api-spec.yaml')
-
-    with open(spec_path, 'r', encoding='utf-8') as f:
-        spec_dict = yaml.safe_load(f)
-    _openapi_spec = OpenAPI.from_dict(spec_dict)
-    return _openapi_spec
-
-@functools.cache
-def get_openapi_schema_dict():
+def build_merged_spec_dict():
    """
-    Get the raw OpenAPI spec dictionary for schema access.
+    Load the base OpenAPI spec and merge in any per-processor api.yaml extensions.

-    Used by Import endpoint to validate and convert query parameters.
-    Returns the YAML dict directly (not the OpenAPI object).
+    Each processor can provide an api.yaml file alongside its __init__.py that defines
+    additional schemas (e.g., processor_config_restock_diff). These are merged into
+    WatchBase.properties so the spec accurately reflects what the API accepts.
+
+    Plugin processors (via pluggy) are also supported - they just need an api.yaml
+    next to their processor module.
+
+    Returns the merged dict (cached - do not mutate the returned value).
    """
    import os
    import yaml
@@ -35,7 +24,59 @@ def get_openapi_schema_dict():
        spec_path = os.path.join(os.path.dirname(__file__), '../docs/api-spec.yaml')

    with open(spec_path, 'r', encoding='utf-8') as f:
-        return yaml.safe_load(f)
+        spec_dict = yaml.safe_load(f)
+
+    try:
+        from changedetectionio.processors import find_processors, get_parent_module
+        for module, proc_name in find_processors():
+            parent = get_parent_module(module)
+            if not parent or not hasattr(parent, '__file__'):
+                continue
+            api_yaml_path = os.path.join(os.path.dirname(parent.__file__), 'api.yaml')
+            if not os.path.exists(api_yaml_path):
+                continue
+            with open(api_yaml_path, 'r', encoding='utf-8') as f:
+                proc_spec = yaml.safe_load(f)
+            # Merge schemas
+            proc_schemas = proc_spec.get('components', {}).get('schemas', {})
+            spec_dict['components']['schemas'].update(proc_schemas)
+            # Inject processor_config_{name} into WatchBase if the schema is defined
+            schema_key = f'processor_config_{proc_name}'
+            if schema_key in proc_schemas:
+                spec_dict['components']['schemas']['WatchBase']['properties'][schema_key] = {
+                    '$ref': f'#/components/schemas/{schema_key}'
+                }
+            # Append x-code-samples from processor paths into existing path operations
+            for path, path_item in proc_spec.get('paths', {}).items():
+                if path not in spec_dict.get('paths', {}):
+                    continue
+                for method, operation in path_item.items():
+                    if method not in spec_dict['paths'][path]:
+                        continue
+                    if 'x-code-samples' in operation:
+                        existing = spec_dict['paths'][path][method].get('x-code-samples', [])
+                        spec_dict['paths'][path][method]['x-code-samples'] = existing + operation['x-code-samples']
+    except Exception as e:
+        logger.warning(f"Failed to merge processor API specs: {e}")
+
+    return spec_dict
+
+
+@functools.cache
+def get_openapi_spec():
+    """Lazy load OpenAPI spec and dependencies only when validation is needed."""
+    from openapi_core import OpenAPI  # Lazy import - saves ~10.7 MB on startup
+    return OpenAPI.from_dict(build_merged_spec_dict())
+
+@functools.cache
+def get_openapi_schema_dict():
+    """
+    Get the raw OpenAPI spec dictionary for schema access.
+
+    Used by Import endpoint to validate and convert query parameters.
+    Returns the merged YAML dict (not the OpenAPI object).
+    """
+    return build_merged_spec_dict()

@functools.cache
 def _resolve_schema_properties(schema_name):
@@ -92,6 +133,43 @@ def get_tag_schema_properties():
    """
    return _resolve_schema_properties('Tag')

+def strip_private_keys(data):
+    """
+    Remove `__`-prefixed keys from a watch/tag dict at the API boundary.
+
+    These are transient in-memory fields (e.g. `__check_status` set by the worker to
+    surface "Fetching page..." in the UI) and are not part of the public OpenAPI
+    contract. They must never appear in GET responses (otherwise a client that
+    round-trips GET → PUT trips the unknown-field validator), and must be silently
+    discarded from incoming PUT/POST payloads.
+
+    Returns a new dict; the input is not mutated.
+    """
+    if not isinstance(data, dict):
+        return data
+    return {k: v for k, v in data.items() if not (isinstance(k, str) and k.startswith('__'))}
+
+
+def strip_internal_api_fields(data):
+    """
+    Strip both `__`-prefixed keys AND system-managed fields that aren't in the public
+    OpenAPI spec (skip-cache hashes, LLM runtime state, processor-set status, etc.).
+
+    Use this at every public API boundary so GET responses and PUT/POST payloads agree
+    on what's part of the contract. The set of system-managed fields lives in
+    model/schema_utils.py:SYSTEM_MANAGED_NON_SPEC_FIELDS — extend it there, not here.
+
+    Returns a new dict; the input is not mutated.
+    """
+    if not isinstance(data, dict):
+        return data
+    from changedetectionio.model.schema_utils import SYSTEM_MANAGED_NON_SPEC_FIELDS
+    return {
+        k: v for k, v in data.items()
+        if not (isinstance(k, str) and (k.startswith('__') or k in SYSTEM_MANAGED_NON_SPEC_FIELDS))
+    }
+
+
 def validate_openapi_request(operation_id):
    """Decorator to validate incoming requests against OpenAPI spec."""
    def decorator(f):
@@ -103,6 +181,7 @@ def validate_openapi_request(operation_id):
                if request.method.upper() != 'GET':
                    # Lazy import - only loaded when actually validating a request
                    from openapi_core.contrib.flask import FlaskOpenAPIRequest
+                    from openapi_core.templating.paths.exceptions import ServerNotFound, PathNotFound, PathError

                    spec = get_openapi_spec()
                    openapi_request = FlaskOpenAPIRequest(request)
@@ -110,6 +189,16 @@ def validate_openapi_request(operation_id):
                    if result.errors:
                        error_details = []
                        for error in result.errors:
+                            # Skip path/server validation errors for reverse proxy compatibility
+                            # Flask routing already validates that endpoints exist (returns 404 if not).
+                            # OpenAPI validation here is primarily for request body schema validation.
+                            # When behind nginx/reverse proxy, URLs may have path prefixes that don't
+                            # match the OpenAPI server definitions, causing false positives.
+                            if isinstance(error, PathError):
+                                logger.debug(f"API Call - Skipping path/server validation (delegated to Flask): {error}")
+                                continue
+
+                            error_str = str(error)
                            # Extract detailed schema errors from __cause__
                            if hasattr(error, '__cause__') and hasattr(error.__cause__, 'schema_errors'):
                                for schema_error in error.__cause__.schema_errors:
@@ -117,9 +206,12 @@ def validate_openapi_request(operation_id):
                                    msg = schema_error.message if hasattr(schema_error, 'message') else str(schema_error)
                                    error_details.append(f"{field}: {msg}")
                            else:
-                                error_details.append(str(error))
+                                error_details.append(error_str)
+
+                        # Only raise if we have actual validation errors (not path/server issues)
+                        if error_details:
                            logger.error(f"API Call - Validation failed: {'; '.join(error_details)}")
-                        raise BadRequest(f"Validation failed: {'; '.join(error_details)}")
+                            raise BadRequest(f"Validation failed: {'; '.join(error_details)}")
            except BadRequest:
                # Re-raise BadRequest exceptions (validation failures)
                raise
@@ -136,5 +228,6 @@ from .Watch import Watch, WatchHistory, WatchSingleHistory, WatchHistoryDiff, Cr
 from .Tags import Tags, Tag
 from .Import import Import
 from .SystemInfo import SystemInfo
+from .Spec import Spec
 from .Notifications import Notifications

@@ -3,6 +3,16 @@ from functools import wraps
 from flask import current_app, redirect, request
 from loguru import logger

+# Endpoints exempt from auth when `shared_diff_access` is enabled.
+# Must be exact endpoint names — substring matching (GHSA-vwgh-2hvh-4xm5)
+# let the state-changing `/diff/<uuid>/extract` endpoints slip through
+# because their names share the `diff_history_page` prefix.
+SHARED_DIFF_READ_ONLY_ENDPOINTS = frozenset({
+    'ui.ui_diff.diff_history_page',
+    'ui.ui_diff.processor_asset',
+    'ui.ui_diff.download_patch',
+})
+
 def login_optionally_required(func):
    """
    If password authentication is enabled, verify the user is logged in.
@@ -20,7 +30,7 @@ def login_optionally_required(func):
        has_password_enabled = datastore.data['settings']['application'].get('password') or os.getenv("SALTED_PASS", False)

        # Permitted
-        if request.endpoint and 'diff_history_page' in request.endpoint and datastore.data['settings']['application'].get('shared_diff_access'):
+        if request.endpoint in SHARED_DIFF_READ_ONLY_ENDPOINTS and datastore.data['settings']['application'].get('shared_diff_access'):
            return func(*args, **kwargs)
        elif request.method in flask_login.config.EXEMPT_METHODS:
            return func(*args, **kwargs)
@@ -13,7 +13,7 @@ from loguru import logger
 BACKUP_FILENAME_FORMAT = "changedetection-backup-{}.zip"


-def create_backup(datastore_path, watches: dict):
+def create_backup(datastore_path, watches: dict, tags: dict = None):
    logger.debug("Creating backup...")
    import zipfile
    from pathlib import Path
@@ -40,10 +40,14 @@ def create_backup(datastore_path, watches: dict):
            zipObj.write(url_watches_json, arcname="url-watches.json")
            logger.debug("Added url-watches.json to backup")

-        # Add the flask app secret (if it exists)
-        secret_file = os.path.join(datastore_path, "secret.txt")
-        if os.path.isfile(secret_file):
-            zipObj.write(secret_file, arcname="secret.txt")
+        # Add tag data directories (each tag has its own {uuid}/tag.json)
+        for uuid, tag in (tags or {}).items():
+            for f in Path(tag.data_dir).glob('*'):
+                zipObj.write(f,
+                             arcname=os.path.join(f.parts[-2], f.parts[-1]),
+                             compress_type=zipfile.ZIP_DEFLATED,
+                             compresslevel=8)
+            logger.debug(f"Added tag '{tag.get('title')}' ({uuid}) to backup")

        # Add any data in the watch data directory.
        for uuid, w in watches.items():
@@ -88,24 +92,28 @@ def create_backup(datastore_path, watches: dict):


 def construct_blueprint(datastore: ChangeDetectionStore):
+    from .restore import construct_restore_blueprint
+
    backups_blueprint = Blueprint('backups', __name__, template_folder="templates")
+    backups_blueprint.register_blueprint(construct_restore_blueprint(datastore))
    backup_threads = []

-    @login_optionally_required
    @backups_blueprint.route("/request-backup", methods=['GET'])
+    @login_optionally_required
    def request_backup():
        if any(thread.is_alive() for thread in backup_threads):
            flash(gettext("A backup is already running, check back in a few minutes"), "error")
-            return redirect(url_for('backups.index'))
+            return redirect(url_for('backups.create'))

        if len(find_backups()) > int(os.getenv("MAX_NUMBER_BACKUPS", 100)):
            flash(gettext("Maximum number of backups reached, please remove some"), "error")
-            return redirect(url_for('backups.index'))
+            return redirect(url_for('backups.create'))

        # With immediate persistence, all data is already saved
        zip_thread = threading.Thread(
            target=create_backup,
            args=(datastore.datastore_path, datastore.data.get("watching")),
+            kwargs={'tags': datastore.data['settings']['application'].get('tags', {})},
            daemon=True,
            name="BackupCreator"
        )
@@ -113,7 +121,7 @@ def construct_blueprint(datastore: ChangeDetectionStore):
        backup_threads.append(zip_thread)
        flash(gettext("Backup building in background, check back in a few minutes."))

-        return redirect(url_for('backups.index'))
+        return redirect(url_for('backups.create'))

    def find_backups():
        backup_filepath = os.path.join(datastore.datastore_path, BACKUP_FILENAME_FORMAT.format("*"))
@@ -133,40 +141,43 @@ def construct_blueprint(datastore: ChangeDetectionStore):

        return backup_info

-    @login_optionally_required
    @backups_blueprint.route("/download/<string:filename>", methods=['GET'])
+    @login_optionally_required
    def download_backup(filename):
        import re
        filename = filename.strip()
-        backup_filename_regex = BACKUP_FILENAME_FORMAT.format("\d+")
-
-        full_path = os.path.join(os.path.abspath(datastore.datastore_path), filename)
-        if not full_path.startswith(os.path.abspath(datastore.datastore_path)):
-            abort(404)
+        backup_filename_regex = BACKUP_FILENAME_FORMAT.format(r"\d+")

+        # Resolve 'latest' before any validation so checks run against the real filename.
        if filename == 'latest':
            backups = find_backups()
+            if not backups:
+                abort(404)
            filename = backups[0]['filename']

        if not re.match(r"^" + backup_filename_regex + "$", filename):
            abort(400)  # Bad Request if the filename doesn't match the pattern

+        full_path = os.path.join(os.path.abspath(datastore.datastore_path), filename)
+        if not full_path.startswith(os.path.abspath(datastore.datastore_path) + os.sep):
+            abort(404)
+
        logger.debug(f"Backup download request for '{full_path}'")
        return send_from_directory(os.path.abspath(datastore.datastore_path), filename, as_attachment=True)

+    @backups_blueprint.route("/", methods=['GET'])
+    @backups_blueprint.route("/create", methods=['GET'])
    @login_optionally_required
-    @backups_blueprint.route("", methods=['GET'])
-    def index():
+    def create():
        backups = find_backups()
-        output = render_template("overview.html",
+        output = render_template("backup_create.html",
                                 available_backups=backups,
                                 backup_running=any(thread.is_alive() for thread in backup_threads)
                                 )
-
        return output

-    @login_optionally_required
    @backups_blueprint.route("/remove-backups", methods=['GET'])
+    @login_optionally_required
    def remove_backups():

        backup_filepath = os.path.join(datastore.datastore_path, BACKUP_FILENAME_FORMAT.format("*"))
@@ -176,6 +187,6 @@ def construct_blueprint(datastore: ChangeDetectionStore):

        flash(gettext("Backups were deleted."))

-        return redirect(url_for('backups.index'))
+        return redirect(url_for('backups.create'))

    return backups_blueprint
@@ -0,0 +1,248 @@
+import io
+import json
+import os
+import re
+import shutil
+import tempfile
+import threading
+import zipfile
+
+from flask import Blueprint, render_template, flash, url_for, redirect, request
+from flask_babel import gettext, lazy_gettext as _l
+from wtforms import Form, BooleanField, SubmitField
+from flask_wtf.file import FileField, FileAllowed
+from loguru import logger
+
+from changedetectionio.flask_app import login_optionally_required
+
+# Maximum size of the uploaded zip file. Override via env var MAX_RESTORE_UPLOAD_MB.
+_MAX_UPLOAD_BYTES = int(os.getenv("MAX_RESTORE_UPLOAD_MB", 256)) * 1024 * 1024
+# Maximum total uncompressed size of all entries (zip-bomb guard). Override via MAX_RESTORE_DECOMPRESSED_MB.
+_MAX_DECOMPRESSED_BYTES = int(os.getenv("MAX_RESTORE_DECOMPRESSED_MB", 1024)) * 1024 * 1024
+# Only top-level directories whose name is a valid UUID are treated as watch/tag entries.
+_UUID_RE = re.compile(
+    r'^[0-9a-f]{8}-[0-9a-f]{4}-4[0-9a-f]{3}-[89ab][0-9a-f]{3}-[0-9a-f]{12}$',
+    re.IGNORECASE,
+)
+
+
+class RestoreForm(Form):
+    zip_file = FileField(_l('Backup zip file'), validators=[
+        FileAllowed(['zip'], _l('Must be a .zip backup file!'))
+    ])
+    include_groups = BooleanField(_l('Include groups'), default=True)
+    include_groups_replace_existing = BooleanField(_l('Replace existing groups of the same UUID'), default=True)
+    include_watches = BooleanField(_l('Include watches'), default=True)
+    include_watches_replace_existing = BooleanField(_l('Replace existing watches of the same UUID'), default=True)
+    submit = SubmitField(_l('Restore backup'))
+
+
+def import_from_zip(zip_stream, datastore, include_groups, include_groups_replace, include_watches, include_watches_replace):
+    """
+    Extract and import watches and groups from a backup zip stream.
+
+    Mirrors the store's _load_watches / _load_tags loading pattern:
+      - UUID dirs with tag.json  → Tag.model + tag_obj.commit()
+      - UUID dirs with watch.json → rehydrate_entity + watch_obj.commit()
+
+    Returns a dict with counts: restored_groups, skipped_groups, restored_watches, skipped_watches.
+    Raises zipfile.BadZipFile if the stream is not a valid zip.
+    """
+    from changedetectionio.model import Tag
+
+    restored_groups = 0
+    skipped_groups = 0
+    restored_watches = 0
+    skipped_watches = 0
+
+    current_tags = datastore.data['settings']['application'].get('tags', {})
+    current_watches = datastore.data['watching']
+
+    with tempfile.TemporaryDirectory() as tmpdir:
+        logger.debug(f"Restore: extracting zip to {tmpdir}")
+        with zipfile.ZipFile(zip_stream, 'r') as zf:
+            total_uncompressed = sum(m.file_size for m in zf.infolist())
+            if total_uncompressed > _MAX_DECOMPRESSED_BYTES:
+                raise ValueError(
+                    f"Backup archive decompressed size ({total_uncompressed // (1024 * 1024)} MB) "
+                    f"exceeds the {_MAX_DECOMPRESSED_BYTES // (1024 * 1024)} MB limit"
+                )
+            resolved_dest = os.path.realpath(tmpdir)
+            for member in zf.infolist():
+                member_dest = os.path.realpath(os.path.join(resolved_dest, member.filename))
+                if not member_dest.startswith(resolved_dest + os.sep) and member_dest != resolved_dest:
+                    raise ValueError(f"Zip Slip path traversal detected in backup archive: {member.filename!r}")
+                zf.extract(member, tmpdir)
+        logger.debug("Restore: zip extracted, scanning UUID directories")
+
+        for entry in os.scandir(tmpdir):
+            if not entry.is_dir():
+                continue
+
+            uuid = entry.name
+            if not _UUID_RE.match(uuid):
+                logger.warning(f"Restore: skipping non-UUID directory {uuid!r}")
+                continue
+            tag_json_path = os.path.join(entry.path, 'tag.json')
+            watch_json_path = os.path.join(entry.path, 'watch.json')
+
+            # --- Tags (groups) ---
+            if include_groups and os.path.exists(tag_json_path):
+                if uuid in current_tags and not include_groups_replace:
+                    logger.debug(f"Restore: skipping existing group {uuid} (replace not requested)")
+                    skipped_groups += 1
+                    continue
+
+                try:
+                    with open(tag_json_path, 'r', encoding='utf-8') as f:
+                        tag_data = json.load(f)
+                except (json.JSONDecodeError, IOError) as e:
+                    logger.error(f"Restore: failed to read tag.json for {uuid}: {e}")
+                    continue
+
+                title = tag_data.get('title', uuid)
+                logger.debug(f"Restore: importing group '{title}' ({uuid})")
+
+                # Mirror _load_tags: set uuid and force processor
+                tag_data['uuid'] = uuid
+                tag_data['processor'] = 'restock_diff'
+
+                # Copy the UUID directory so data_dir exists for commit()
+                dst_dir = os.path.join(datastore.datastore_path, uuid)
+                if os.path.exists(dst_dir):
+                    shutil.rmtree(dst_dir)
+                shutil.copytree(entry.path, dst_dir)
+
+                tag_obj = Tag.model(
+                    datastore_path=datastore.datastore_path,
+                    __datastore=datastore.data,
+                    default=tag_data
+                )
+                current_tags[uuid] = tag_obj
+                tag_obj.commit()
+                restored_groups += 1
+                logger.success(f"Restore: group '{title}' ({uuid}) restored")
+
+            # --- Watches ---
+            elif include_watches and os.path.exists(watch_json_path):
+                if uuid in current_watches and not include_watches_replace:
+                    logger.debug(f"Restore: skipping existing watch {uuid} (replace not requested)")
+                    skipped_watches += 1
+                    continue
+
+                try:
+                    with open(watch_json_path, 'r', encoding='utf-8') as f:
+                        watch_data = json.load(f)
+                except (json.JSONDecodeError, IOError) as e:
+                    logger.error(f"Restore: failed to read watch.json for {uuid}: {e}")
+                    continue
+
+                url = watch_data.get('url', uuid)
+                logger.debug(f"Restore: importing watch '{url}' ({uuid})")
+
+                # Copy UUID directory first so data_dir and history files exist
+                dst_dir = os.path.join(datastore.datastore_path, uuid)
+                if os.path.exists(dst_dir):
+                    shutil.rmtree(dst_dir)
+                shutil.copytree(entry.path, dst_dir)
+
+                # Mirror _load_watches / rehydrate_entity
+                watch_data['uuid'] = uuid
+                watch_obj = datastore.rehydrate_entity(uuid, watch_data)
+                current_watches[uuid] = watch_obj
+                watch_obj.commit()
+                restored_watches += 1
+                logger.success(f"Restore: watch '{url}' ({uuid}) restored")
+
+        logger.debug(f"Restore: scan complete - groups {restored_groups} restored / {skipped_groups} skipped, "
+                     f"watches {restored_watches} restored / {skipped_watches} skipped")
+
+    # Persist changedetection.json (includes the updated tags dict)
+    logger.debug("Restore: committing datastore settings")
+    datastore.commit()
+
+    return {
+        'restored_groups': restored_groups,
+        'skipped_groups': skipped_groups,
+        'restored_watches': restored_watches,
+        'skipped_watches': skipped_watches,
+    }
+
+
+
+def construct_restore_blueprint(datastore):
+    restore_blueprint = Blueprint('restore', __name__, template_folder="templates")
+    restore_threads = []
+
+    @restore_blueprint.route("/restore", methods=['GET'])
+    @login_optionally_required
+    def restore():
+        form = RestoreForm()
+        return render_template("backup_restore.html",
+                               form=form,
+                               restore_running=any(t.is_alive() for t in restore_threads),
+                               max_upload_mb=_MAX_UPLOAD_BYTES // (1024 * 1024),
+                               max_decompressed_mb=_MAX_DECOMPRESSED_BYTES // (1024 * 1024))
+
+    @restore_blueprint.route("/restore/start", methods=['POST'])
+    @login_optionally_required
+    def backups_restore_start():
+        if any(t.is_alive() for t in restore_threads):
+            flash(gettext("A restore is already running, check back in a few minutes"), "error")
+            return redirect(url_for('backups.restore.restore'))
+
+        zip_file = request.files.get('zip_file')
+        if not zip_file or not zip_file.filename:
+            flash(gettext("No file uploaded"), "error")
+            return redirect(url_for('backups.restore.restore'))
+
+        if not zip_file.filename.lower().endswith('.zip'):
+            flash(gettext("File must be a .zip backup file"), "error")
+            return redirect(url_for('backups.restore.restore'))
+
+        # Reject oversized uploads before reading the stream into memory.
+        content_length = request.content_length
+        if content_length and content_length > _MAX_UPLOAD_BYTES:
+            flash(gettext("Backup file is too large (max %(mb)s MB)", mb=_MAX_UPLOAD_BYTES // (1024 * 1024)), "error")
+            return redirect(url_for('backups.restore.restore'))
+
+        # Read into memory now — the request stream is gone once we return.
+        # Read one byte beyond the limit so we can detect truncated-but-still-oversized streams.
+        try:
+            raw = zip_file.read(_MAX_UPLOAD_BYTES + 1)
+            if len(raw) > _MAX_UPLOAD_BYTES:
+                flash(gettext("Backup file is too large (max %(mb)s MB)", mb=_MAX_UPLOAD_BYTES // (1024 * 1024)), "error")
+                return redirect(url_for('backups.restore.restore'))
+            zip_bytes = io.BytesIO(raw)
+            with zipfile.ZipFile(zip_bytes):  # quick validity check before spawning
+                pass
+            zip_bytes.seek(0)
+        except zipfile.BadZipFile:
+            flash(gettext("Invalid or corrupted zip file"), "error")
+            return redirect(url_for('backups.restore.restore'))
+
+        include_groups = request.form.get('include_groups') == 'y'
+        include_groups_replace = request.form.get('include_groups_replace_existing') == 'y'
+        include_watches = request.form.get('include_watches') == 'y'
+        include_watches_replace = request.form.get('include_watches_replace_existing') == 'y'
+
+        restore_thread = threading.Thread(
+            target=import_from_zip,
+            kwargs={
+                'zip_stream': zip_bytes,
+                'datastore': datastore,
+                'include_groups': include_groups,
+                'include_groups_replace': include_groups_replace,
+                'include_watches': include_watches,
+                'include_watches_replace': include_watches_replace,
+            },
+            daemon=True,
+            name="BackupRestore"
+        )
+        restore_thread.start()
+        restore_threads[:] = [t for t in restore_threads if t.is_alive()]
+        restore_threads.append(restore_thread)
+        flash(gettext("Restore started in background, check back in a few minutes."))
+        return redirect(url_for('backups.restore.restore'))
+
+    return restore_blueprint
@@ -0,0 +1,49 @@
+{% extends 'base.html' %}
+{% block content %}
+    {% from '_helpers.html' import render_simple_field, render_field %}
+
+    <div class="edit-form">
+        <div class="tabs collapsable">
+            <ul>
+                <li class="tab active" id=""><a href="{{ url_for('backups.create') }}">{{ _('Create') }}</a></li>
+                <li class="tab"><a href="{{ url_for('backups.restore.restore') }}">{{ _('Restore') }}</a></li>
+            </ul>
+        </div>
+        <div class="box-wrap inner">
+            <div id="general">
+                {% if backup_running %}
+                    <p>
+                        <span class="spinner"></span>&nbsp;<strong>{{ _('A backup is running!') }}</strong>
+                    </p>
+                {% endif %}
+
+                <p>
+                    {{ _('Here you can download and request a new backup, when a backup is completed you will see it listed below.') }}
+                </p>
+                <br>
+                {% if available_backups %}
+                    <ul>
+                        {% for backup in available_backups %}
+                            <li>
+                                <a href="{{ url_for('backups.download_backup', filename=backup["filename"]) }}">{{ backup["filename"] }}</a> {{ backup["filesize"] }} {{ _('Mb') }}
+                            </li>
+                        {% endfor %}
+                    </ul>
+                {% else %}
+                    <p>
+                        <strong>{{ _('No backups found.') }}</strong>
+                    </p>
+                {% endif %}
+
+                <a class="pure-button pure-button-primary"
+                   href="{{ url_for('backups.request_backup') }}">{{ _('Create backup') }}</a>
+                {% if available_backups %}
+                    <a class="pure-button button-small button-error "
+                       href="{{ url_for('backups.remove_backups') }}">{{ _('Remove backups') }}</a>
+                {% endif %}
+
+            </div>
+
+        </div>
+    </div>
+{% endblock %}
@@ -0,0 +1,61 @@
+{% extends 'base.html' %}
+{% block content %}
+    {% from '_helpers.html' import render_field, render_checkbox_field %}
+
+    <div class="edit-form">
+        <div class="tabs collapsable">
+            <ul>
+                <li class="tab"><a href="{{ url_for('backups.create') }}">{{ _('Create') }}</a></li>
+                <li class="tab active"><a href="{{ url_for('backups.restore.restore') }}">{{ _('Restore') }}</a></li>
+            </ul>
+        </div>
+        <div class="box-wrap inner">
+            <div id="general">
+                {% if restore_running %}
+                    <p>
+                        <span class="spinner"></span>&nbsp;<strong>{{ _('A restore is running!') }}</strong>
+                    </p>
+                {% endif %}
+
+                <p>{{ _('Restore a backup. Must be a .zip backup file created on/after v0.53.1 (new database layout).') }}</p>
+                <p>{{ _('Note: This does not override the main application settings, only watches and groups.') }}</p>
+                <p class="pure-form-message">
+                    {{ _('Max upload size: %(upload)s MB, Max decompressed size: %(decomp)s MB', upload=max_upload_mb, decomp=max_decompressed_mb) }}
+                </p>
+
+                <form class="pure-form pure-form-stacked settings"
+                      action="{{ url_for('backups.restore.backups_restore_start') }}"
+                      method="POST"
+                      enctype="multipart/form-data">
+                    <input type="hidden" name="csrf_token" value="{{ csrf_token() }}">
+
+                    <div class="pure-control-group">
+                        {{ render_checkbox_field(form.include_groups) }}
+                        <span class="pure-form-message-inline">{{ _('Include all groups found in backup?') }}</span>
+                    </div>
+                    <div class="pure-control-group">
+                        {{ render_checkbox_field(form.include_groups_replace_existing) }}
+                        <span class="pure-form-message-inline">{{ _('Replace any existing groups of the same UUID?') }}</span>
+                    </div>
+
+                    <div class="pure-control-group">
+                        {{ render_checkbox_field(form.include_watches) }}
+                        <span class="pure-form-message-inline">{{ _('Include all watches found in backup?') }}</span>
+                    </div>
+                    <div class="pure-control-group">
+                        {{ render_checkbox_field(form.include_watches_replace_existing) }}
+                        <span class="pure-form-message-inline">{{ _('Replace any existing watches of the same UUID?') }}</span>
+                    </div>
+
+                    <div class="pure-control-group">
+                        {{ render_field(form.zip_file) }}
+                    </div>
+
+                    <div class="pure-controls">
+                        <button type="submit" class="pure-button pure-button-primary">{{ _('Restore backup') }}</button>
+                    </div>
+                </form>
+            </div>
+        </div>
+    </div>
+{% endblock %}
@@ -1,36 +0,0 @@
-{% extends 'base.html' %}
-{% block content %}
-    {% from '_helpers.html' import render_simple_field, render_field %}
-    <div class="edit-form">
-        <div class="box-wrap inner">
-            <h2>{{ _('Backups') }}</h2>
-            {% if backup_running %}
-                <p>
-                    <span class="spinner"></span>&nbsp;<strong>{{ _('A backup is running!') }}</strong>
-                </p>
-            {% endif %}
-            <p>
-                {{ _('Here you can download and request a new backup, when a backup is completed you will see it listed below.') }}
-            </p>
-            <br>
-                {% if available_backups %}
-                    <ul>
-                    {% for backup in available_backups %}
-                        <li><a href="{{ url_for('backups.download_backup', filename=backup["filename"]) }}">{{ backup["filename"] }}</a> {{  backup["filesize"] }} {{ _('Mb') }}</li>
-                    {% endfor %}
-                    </ul>
-                {% else %}
-                    <p>
-                    <strong>{{ _('No backups found.') }}</strong>
-                    </p>
-                {% endif %}
-
-            <a class="pure-button pure-button-primary" href="{{ url_for('backups.request_backup') }}">{{ _('Create backup') }}</a>
-            {% if available_backups %}
-                <a class="pure-button button-small button-error " href="{{ url_for('backups.remove_backups') }}">{{ _('Remove backups') }}</a>
-            {% endif %}
-        </div>
-    </div>
-
-
-{% endblock %}
@@ -102,6 +102,35 @@ def run_async_in_browser_loop(coro):
    else:
        raise RuntimeError("Browser steps event loop is not available")

+async def _close_session_resources(session_data, label=''):
+    """Close all browser resources for a session in the correct order.
+
+    browserstepper.cleanup() closes page+context but not the browser itself.
+    For CloakBrowser, browser.close() is what stops the local Chromium process via pw.stop().
+    For the default CDP path, playwright_context.stop() shuts down the playwright instance.
+    """
+    browserstepper = session_data.get('browserstepper')
+    if browserstepper:
+        try:
+            await browserstepper.cleanup()
+        except Exception as e:
+            logger.error(f"Error cleaning up browserstepper{label}: {e}")
+
+    browser = session_data.get('browser')
+    if browser:
+        try:
+            await asyncio.wait_for(browser.close(), timeout=5.0)
+        except Exception as e:
+            logger.warning(f"Error closing browser{label}: {e}")
+
+    playwright_context = session_data.get('playwright_context')
+    if playwright_context:
+        try:
+            await playwright_context.stop()
+        except Exception as e:
+            logger.warning(f"Error stopping playwright context{label}: {e}")
+
+
 def cleanup_expired_sessions():
    """Remove expired browsersteps sessions and cleanup their resources"""
    global browsersteps_sessions, browsersteps_watch_to_session
@@ -119,13 +148,10 @@ def cleanup_expired_sessions():
        logger.debug(f"Cleaning up expired browsersteps session {session_id}")
        session_data = browsersteps_sessions[session_id]

-        # Cleanup playwright resources asynchronously
-        browserstepper = session_data.get('browserstepper')
-        if browserstepper:
-            try:
-                run_async_in_browser_loop(browserstepper.cleanup())
-            except Exception as e:
-                logger.error(f"Error cleaning up session {session_id}: {e}")
+        try:
+            run_async_in_browser_loop(_close_session_resources(session_data, label=f" for session {session_id}"))
+        except Exception as e:
+            logger.error(f"Error cleaning up session {session_id}: {e}")

        # Remove from sessions dict
        del browsersteps_sessions[session_id]
@@ -152,12 +178,10 @@ def cleanup_session_for_watch(watch_uuid):

    session_data = browsersteps_sessions.get(session_id)
    if session_data:
-        browserstepper = session_data.get('browserstepper')
-        if browserstepper:
-            try:
-                run_async_in_browser_loop(browserstepper.cleanup())
-            except Exception as e:
-                logger.error(f"Error cleaning up session {session_id} for watch {watch_uuid}: {e}")
+        try:
+            run_async_in_browser_loop(_close_session_resources(session_data, label=f" for watch {watch_uuid}"))
+        except Exception as e:
+            logger.error(f"Error cleaning up session {session_id} for watch {watch_uuid}: {e}")

        # Remove from sessions dict
        del browsersteps_sessions[session_id]
@@ -178,64 +202,74 @@ def construct_blueprint(datastore: ChangeDetectionStore):
        import time
        from playwright.async_api import async_playwright

-        # We keep the playwright session open for many minutes
        keepalive_seconds = int(os.getenv('BROWSERSTEPS_MINUTES_KEEPALIVE', 10)) * 60
+        keepalive_ms = ((keepalive_seconds + 3) * 1000)

        browsersteps_start_session = {'start_time': time.time()}

-        # Create a new async playwright instance for browser steps
-        playwright_instance = async_playwright()
-        playwright_context = await playwright_instance.start()
-
-        keepalive_ms = ((keepalive_seconds + 3) * 1000)
-        base_url = os.getenv('PLAYWRIGHT_DRIVER_URL', '').strip('"')
-        a = "?" if not '?' in base_url else '&'
-        base_url += a + f"timeout={keepalive_ms}"
-
-        browser = await playwright_context.chromium.connect_over_cdp(base_url, timeout=keepalive_ms)
-        browsersteps_start_session['browser'] = browser
-        browsersteps_start_session['playwright_context'] = playwright_context
-
+        # Build proxy dict first — needed by both the CDP path and fetcher-specific launchers
        proxy_id = datastore.get_preferred_proxy_for_watch(uuid=watch_uuid)
        proxy = None
        if proxy_id:
-            proxy_url = datastore.proxy_list.get(proxy_id).get('url')
+            proxy_url = datastore.proxy_list.get(proxy_id, {}).get('url')
            if proxy_url:
-
-                # Playwright needs separate username and password values
                from urllib.parse import urlparse
                parsed = urlparse(proxy_url)
                proxy = {'server': proxy_url}
-
                if parsed.username:
                    proxy['username'] = parsed.username
-
                if parsed.password:
                    proxy['password'] = parsed.password
-
                logger.debug(f"Browser Steps: UUID {watch_uuid} selected proxy {proxy_url}")

-        # Tell Playwright to connect to Chrome and setup a new session via our stepper interface
+        # Resolve the fetcher class for this watch so we can ask it to launch its own browser
+        # if it supports that (e.g. CloakBrowser, which runs locally rather than via CDP)
+        watch = datastore.data['watching'][watch_uuid]
+        from changedetectionio import content_fetchers
+        fetcher_name = watch.get_fetch_backend or 'system'
+        if fetcher_name == 'system':
+            fetcher_name = datastore.data['settings']['application'].get('fetch_backend', 'html_requests')
+        fetcher_class = getattr(content_fetchers, fetcher_name, None)
+
+        browser = None
+        playwright_context = None
+
+        # If the fetcher has its own browser launch for the live steps UI, use it.
+        # get_browsersteps_browser(proxy, keepalive_ms) returns (browser, playwright_context_or_None)
+        # or None to fall back to the default CDP path.
+        if fetcher_class and hasattr(fetcher_class, 'get_browsersteps_browser'):
+            result = await fetcher_class.get_browsersteps_browser(proxy=proxy, keepalive_ms=keepalive_ms)
+            if result is not None:
+                browser, playwright_context = result
+                logger.debug(f"Browser Steps: using fetcher-specific browser for '{fetcher_name}'")
+
+        # Default: connect to the remote Playwright/sockpuppetbrowser via CDP
+        if browser is None:
+            playwright_instance = async_playwright()
+            playwright_context = await playwright_instance.start()
+            base_url = os.getenv('PLAYWRIGHT_DRIVER_URL', '').strip('"')
+            a = "?" if '?' not in base_url else '&'
+            base_url += a + f"timeout={keepalive_ms}"
+            browser = await playwright_context.chromium.connect_over_cdp(base_url, timeout=keepalive_ms)
+            logger.debug(f"Browser Steps: using CDP connection to {base_url}")
+
+        browsersteps_start_session['browser'] = browser
+        browsersteps_start_session['playwright_context'] = playwright_context
+
        browserstepper = browser_steps.browsersteps_live_ui(
            playwright_browser=browser,
            proxy=proxy,
-            start_url=datastore.data['watching'][watch_uuid].link,
-            headers=datastore.data['watching'][watch_uuid].get('headers')
+            start_url=watch.link,
+            headers=watch.get('headers')
        )
-        
-        # Initialize the async connection
        await browserstepper.connect(proxy=proxy)
-        
        browsersteps_start_session['browserstepper'] = browserstepper

-        # For test
-        #await browsersteps_start_session['browserstepper'].action_goto_url(value="http://example.com?time="+str(time.time()))
-
        return browsersteps_start_session


-    @login_optionally_required
    @browser_steps_blueprint.route("/browsersteps_start_session", methods=['GET'])
+    @login_optionally_required
    def browsersteps_start_session():
        # A new session was requested, return sessionID
        import uuid
@@ -270,8 +304,8 @@ def construct_blueprint(datastore: ChangeDetectionStore):
        logger.debug("Starting connection with playwright - done")
        return {'browsersteps_session_id': browsersteps_session_id}

-    @login_optionally_required
    @browser_steps_blueprint.route("/browsersteps_image", methods=['GET'])
+    @login_optionally_required
    def browser_steps_fetch_screenshot_image():
        from flask import (
            make_response,
@@ -296,8 +330,8 @@ def construct_blueprint(datastore: ChangeDetectionStore):
            return make_response('Unable to fetch image, is the URL correct? does the watch exist? does the step_type-n.jpeg exist?', 401)

    # A request for an action was received
-    @login_optionally_required
    @browser_steps_blueprint.route("/browsersteps_update", methods=['POST'])
+    @login_optionally_required
    def browsersteps_ui_update():
        import base64

@@ -379,6 +413,4 @@ def construct_blueprint(datastore: ChangeDetectionStore):

    return browser_steps_blueprint

-    return browser_steps_blueprint
-

@@ -40,12 +40,13 @@ def construct_blueprint(datastore: ChangeDetectionStore):
        contents = ''
        now = time.time()
        try:
+            import asyncio
            processor_module = importlib.import_module("changedetectionio.processors.text_json_diff.processor")
            update_handler = processor_module.perform_site_check(datastore=datastore,
                                                                 watch_uuid=uuid
                                                                 )

-            update_handler.call_browser(preferred_proxy_id=preferred_proxy)
+            asyncio.run(update_handler.call_browser(preferred_proxy_id=preferred_proxy))
        # title, size is len contents not len xfer
        except content_fetcher_exceptions.Non200ErrorCodeReceived as e:
            if e.status_code == 404:
@@ -94,13 +95,13 @@ def construct_blueprint(datastore: ChangeDetectionStore):
        return results

    @login_required
-    @check_proxies_blueprint.route("/<string:uuid>/status", methods=['GET'])
+    @check_proxies_blueprint.route("/<uuid_str:uuid>/status", methods=['GET'])
    def get_recheck_status(uuid):
        results = _recalc_check_status(uuid=uuid)
        return results

    @login_required
-    @check_proxies_blueprint.route("/<string:uuid>/start", methods=['GET'])
+    @check_proxies_blueprint.route("/<uuid_str:uuid>/start", methods=['GET'])
    def start_check(uuid):

        if not datastore.proxy_list:
@@ -75,7 +75,7 @@ class import_url_list(Importer):
                self.remaining_data = []
            self.remaining_data.append(url)

-        flash(gettext("{} Imported from list in {:.2f}s, {} Skipped.").format(good, time.time() - now, len(self.remaining_data)))
+        flash(gettext("{count} Imported from list in {duration}s, {skipped_count} Skipped.").format(count=good, duration=f"{time.time() - now:.2f}", skipped_count=len(self.remaining_data)))


 class import_distill_io_json(Importer):
@@ -136,7 +136,7 @@ class import_distill_io_json(Importer):
                    self.new_uuids.append(new_uuid)
                    good += 1

-        flash(gettext("{} Imported from Distill.io in {:.2f}s, {} Skipped.").format(len(self.new_uuids), time.time() - now, len(self.remaining_data)))
+        flash(gettext("{count} Imported from Distill.io in {duration}s, {skipped_count} Skipped.").format(count=len(self.new_uuids), duration=f"{time.time() - now:.2f}", skipped_count=len(self.remaining_data)))


 class import_xlsx_wachete(Importer):
@@ -160,8 +160,7 @@ class import_xlsx_wachete(Importer):
            flash(gettext("Unable to read export XLSX file, something wrong with the file?"), 'error')
            return

-        row_id = 2
-        for row in wb.active.iter_rows(min_row=row_id):
+        for row_id, row in enumerate(wb.active.iter_rows(min_row=2), start=2):
            try:
                extras = {}
                data = {}
@@ -212,10 +211,8 @@ class import_xlsx_wachete(Importer):
            except Exception as e:
                logger.error(e)
                flash(gettext("Error processing row number {}, check all cell data types are correct, row was skipped.").format(row_id), 'error')
-            else:
-                row_id += 1

-        flash(gettext("{} imported from Wachete .xlsx in {:.2f}s").format(len(self.new_uuids), time.time() - now))
+        flash(gettext("{count} imported from Wachete .xlsx in {duration}s").format(count=len(self.new_uuids), duration=f"{time.time() - now:.2f}"))


 class import_xlsx_custom(Importer):
@@ -241,10 +238,10 @@ class import_xlsx_custom(Importer):

        # @todo cehck atleast 2 rows, same in other method
        from changedetectionio.forms import validate_url
-        row_i = 1
+        row_i = 0

        try:
-            for row in wb.active.iter_rows():
+            for row_i, row in enumerate(wb.active.iter_rows(), start=1):
                url = None
                tags = None
                extras = {}
@@ -295,7 +292,5 @@ class import_xlsx_custom(Importer):
        except Exception as e:
            logger.error(e)
            flash(gettext("Error processing row number {}, check all cell data types are correct, row was skipped.").format(row_i), 'error')
-        else:
-            row_i += 1

-        flash(gettext("{} imported from custom .xlsx in {:.2f}s").format(len(self.new_uuids), time.time() - now))
+        flash(gettext("{count} imported from custom .xlsx in {duration}s").format(count=len(self.new_uuids), duration=f"{time.time() - now:.2f}"))
@@ -7,8 +7,9 @@
    <div class="tabs collapsable">
        <ul>
            <li class="tab" id=""><a href="#url-list">{{ _('URL List') }}</a></li>
-            <li class="tab"><a href="#distill-io">{{ _('Distill.io') }}</a></li>
+            <li class="tab"><a href="#distill-io">Distill.io</a></li>
            <li class="tab"><a href="#xlsx">{{ _('.XLSX & Wachete') }}</a></li>
+            <li class="tab"><a href="{{url_for('backups.restore.restore')}}">{{ _('Backup Restore') }}</a></li>
        </ul>
    </div>

@@ -16,10 +17,15 @@
        <form class="pure-form" action="{{url_for('imports.import_page')}}" method="POST" enctype="multipart/form-data">
            <input type="hidden" name="csrf_token" value="{{ csrf_token() }}">
            <div class="tab-pane-inner" id="url-list">
+
+                <p>
+                {{ _('Restoring changedetection.io backups is in the') }}<a href="{{ url_for('backups.restore.restore') }}"> {{ _('backups section') }}</a>.
+                <br>
+                </p>
                <div class="pure-control-group">
                        {{ _('Enter one URL per line, and optionally add tags for each URL after a space, delineated by comma (,):') }}
                        <br>
-                        <p><strong>{{ _('Example:') }}  </strong><code>https://example.com tag1, tag2, last tag</code></p>
+                        <p><strong>{{ _('Example') }}:  </strong><code>https://example.com tag1, tag2, last tag</code></p>
                        {{ _('URLs which do not pass validation will stay in the textarea.') }}
                </div>
                {{ render_field(form.processor, class="processor") }}
@@ -37,20 +43,16 @@
            </div>

            <div class="tab-pane-inner" id="distill-io">
-
-
-
                    <div class="pure-control-group">
                        {{ _('Copy and Paste your Distill.io watch \'export\' file, this should be a JSON file.') }}<br>
-                        {{ _('This is') }} <i>{{ _('experimental') }}</i>, {{ _('supported fields are') }} <code>name</code>, <code>uri</code>, <code>tags</code>, <code>config:selections</code>, {{ _('the rest (including') }} <code>schedule</code>) {{ _('are ignored.') }}
+                        {# TRANSLATORS: CJK fonts lack native italics; allow substitution with conventional local styling. dennis-ignore: W303 #}
+                        {{ _('This is <i>experimental</i>, supported fields are <code>name</code>, <code>uri</code>, <code>tags</code>, <code>config:selections</code>, the rest (including <code>schedule</code>) are ignored.')|safe }}
                        <br>
                        <p>
                        {{ _('How to export?') }} <a href="https://distill.io/docs/web-monitor/how-export-and-import-monitors/">https://distill.io/docs/web-monitor/how-export-and-import-monitors/</a><br>
                        {{ _('Be sure to set your default fetcher to Chrome if required.') }}<br>
                        </p>
                    </div>
-
-
                    <textarea name="distill-io" class="pure-input-1-2" style="width: 100%;
                                font-family:monospace;
                                white-space: pre;
@@ -102,7 +104,7 @@
                            {% for n in range(4) %}
                                <td><select name="custom_xlsx[col_type_{{n}}]">
                                    <option value="" style="color: #aaa"> -- {{ _('none') }} --</option>
-                                    <option value="url">{{ _('URL') }}</option>
+                                    <option value="url">URL</option>
                                    <option value="title">{{ _('Title') }}</option>
                                    <option value="include_filters">{{ _('CSS/xPath filter') }}</option>
                                    <option value="tag">{{ _('Group / Tag name(s)') }}</option>
@@ -114,6 +116,7 @@
                </div>
            </div>
            <button type="submit" class="pure-button pure-input-1-2 pure-button-primary">{{ _('Import') }}</button>
+
        </form>

    </div>
@@ -15,7 +15,7 @@ def construct_blueprint(datastore: ChangeDetectionStore, update_q: PriorityQueue
    price_data_follower_blueprint = Blueprint('price_data_follower', __name__)

    @login_required
-    @price_data_follower_blueprint.route("/<string:uuid>/accept", methods=['GET'])
+    @price_data_follower_blueprint.route("/<uuid_str:uuid>/accept", methods=['GET'])
    def accept(uuid):
        datastore.data['watching'][uuid]['track_ldjson_price_data'] = PRICE_DATA_TRACK_ACCEPT
        datastore.data['watching'][uuid]['processor'] = 'restock_diff'
@@ -25,7 +25,7 @@ def construct_blueprint(datastore: ChangeDetectionStore, update_q: PriorityQueue
        return redirect(url_for("watchlist.index"))

    @login_required
-    @price_data_follower_blueprint.route("/<string:uuid>/reject", methods=['GET'])
+    @price_data_follower_blueprint.route("/<uuid_str:uuid>/reject", methods=['GET'])
    def reject(uuid):
        datastore.data['watching'][uuid]['track_ldjson_price_data'] = PRICE_DATA_TRACK_REJECT
        datastore.data['watching'][uuid].commit()
@@ -9,11 +9,12 @@ def construct_single_watch_routes(rss_blueprint, datastore):
        datastore: The ChangeDetectionStore instance
    """

-    @rss_blueprint.route("/watch/<string:uuid>", methods=['GET'])
+    @rss_blueprint.route("/watch/<uuid_str:uuid>", methods=['GET'])
    def rss_single_watch(uuid):
        import time

-        from flask import make_response, request
+        from flask import make_response, request, Response
+        from flask_babel import lazy_gettext as _l
        from feedgen.feed import FeedGenerator
        from loguru import logger

@@ -42,12 +43,12 @@ def construct_single_watch_routes(rss_blueprint, datastore):
        # Get the watch by UUID
        watch = datastore.data['watching'].get(uuid)
        if not watch:
-            return f"Watch with UUID {uuid} not found", 404
+            return Response(_l("Watch with UUID %(uuid)s not found", uuid=uuid), status=404, mimetype='text/plain')

        # Check if watch has at least 2 history snapshots
        dates = list(watch.history.keys())
        if len(dates) < 2:
-            return f"Watch {uuid} does not have enough history snapshots to show changes (need at least 2)", 400
+            return Response(_l("Watch %(uuid)s does not have enough history snapshots to show changes (need at least 2)", uuid=uuid), status=400, mimetype='text/plain')

        # Get the number of diffs to include (default: 5)
        rss_diff_length = datastore.data['settings']['application'].get('rss_diff_length', 5)
@@ -7,7 +7,7 @@ def construct_tag_routes(rss_blueprint, datastore):
        datastore: The ChangeDetectionStore instance
    """

-    @rss_blueprint.route("/tag/<string:tag_uuid>", methods=['GET'])
+    @rss_blueprint.route("/tag/<uuid_str:tag_uuid>", methods=['GET'])
    def rss_tag_feed(tag_uuid):

        from flask import make_response, request, url_for
@@ -10,10 +10,15 @@ from flask_babel import gettext

 from changedetectionio.store import ChangeDetectionStore
 from changedetectionio.auth_decorator import login_optionally_required
+from changedetectionio.model.LLMSettings import LLMSettings


 def construct_blueprint(datastore: ChangeDetectionStore):
+    from changedetectionio.llm.evaluator import is_llm_features_disabled
    settings_blueprint = Blueprint('settings', __name__, template_folder="templates")
+    if not is_llm_features_disabled():
+        from changedetectionio.blueprint.settings.llm import construct_llm_blueprint
+        settings_blueprint.register_blueprint(construct_llm_blueprint(datastore), url_prefix='/llm')

    @settings_blueprint.route("", methods=['GET', "POST"])
    @login_optionally_required
@@ -27,6 +32,14 @@ def construct_blueprint(datastore: ChangeDetectionStore):


        default = deepcopy(datastore.data['settings'])
+
+        # api_key is intentionally blanked on GET — PasswordField never re-renders
+        # its value, and a blank submission preserves the stored key.
+        default['llm'] = LLMSettings.model_validate(
+            datastore.data['settings']['application'].get('llm') or {}
+        ).model_dump()
+        default['llm']['api_key'] = ''
+
        if datastore.proxy_list is not None:
            available_proxies = list(datastore.proxy_list.keys())
            # When enabled
@@ -76,6 +89,44 @@ def construct_blueprint(datastore: ChangeDetectionStore):

                datastore.data['settings']['application'].update(app_update)

+                # LLM config lives under settings.application.llm.* (post update_31).
+                # Hydrate the stored dict into LLMSettings, then merge form input over it.
+                # WTForms field names match LLMSettings field names exactly, so both sides
+                # of the merge use the same key shape.
+                existing_llm = LLMSettings.model_validate(
+                    datastore.data['settings']['application'].get('llm') or {}
+                )
+
+                llm_form_input = dict(form.data.get('llm') or {})
+
+                # Empty IntegerField submissions come back as None from WTForms;
+                # the schema declares those fields as strict `int`, so passing
+                # them through would fail validation. Treat None like the
+                # absent-key case: keep the stored value, don't merge.
+                llm_form_input = {k: v for k, v in llm_form_input.items() if v is not None}
+
+                # PasswordField never re-renders, so a blank submitted value means
+                # "keep stored key" — drop it from the merge.
+                if not (llm_form_input.get('api_key') or '').strip():
+                    llm_form_input.pop('api_key', None)
+
+                # Env-var overrides make these fields read-only in the UI — ignore form input.
+                if os.getenv('LLM_TOKEN_BUDGET_MONTH', '').strip():
+                    llm_form_input.pop('token_budget_month', None)
+                if os.getenv('LLM_MAX_INPUT_CHARS', '').strip():
+                    llm_form_input.pop('max_input_chars', None)
+
+                # System-managed counters must never come from the form.
+                for protected in LLMSettings.PROTECTED_FIELDS:
+                    llm_form_input.pop(protected, None)
+
+                merged = LLMSettings.model_validate({**existing_llm.model_dump(), **llm_form_input})
+
+                # Clearing the model field strips only the provider-connection fields.
+                # User toggles, budgets, prompts and system counters survive (matches /llm/clear).
+                exclude = set(LLMSettings.CONNECTION_FIELDS) if not merged.model.strip() else None
+                datastore.data['settings']['application']['llm'] = merged.model_dump(exclude=exclude)
+
                # Handle dynamic worker count adjustment
                old_worker_count = datastore.data['settings']['requests'].get('workers', 1)
                new_worker_count = form.data['requests'].get('workers', 1)
@@ -95,8 +146,8 @@ def construct_blueprint(datastore: ChangeDetectionStore):
                    # Check CPU core availability and warn if worker count is high
                    cpu_count = os.cpu_count()
                    if cpu_count and new_worker_count >= (cpu_count * 0.9):
-                        flash(gettext("Warning: Worker count ({}) is close to or exceeds available CPU cores ({})").format(
-                            new_worker_count, cpu_count), 'warning')
+                        flash(gettext("Warning: Worker count ({worker_count}) is close to or exceeds available CPU cores ({cpu_count})").format(
+                            worker_count=new_worker_count, cpu_count=cpu_count), 'warning')

                    result = worker_pool.adjust_async_worker_count(
                        new_count=new_worker_count,
@@ -164,9 +215,34 @@ def construct_blueprint(datastore: ChangeDetectionStore):
            # Instantiate the form with existing settings
            plugin_forms[plugin_id] = form_class(data=settings)

+        from changedetectionio.llm.evaluator import (
+            get_llm_config as _get_llm_cfg,
+            llm_configured_via_env,
+            get_global_token_budget_month,
+        )
+        llm_config = _get_llm_cfg(datastore) or {}
+        llm_env_configured = llm_configured_via_env()
+        llm_stored = datastore.data['settings']['application'].get('llm') or {}
+        llm_token_budget_month = get_global_token_budget_month(datastore)
+        llm_token_budget_month_env = get_global_token_budget_month()  # env var only, for readonly logic
+        _max_input_chars_env_str = os.getenv('LLM_MAX_INPUT_CHARS', '').strip()
+        llm_max_input_chars_env = int(_max_input_chars_env_str) if _max_input_chars_env_str.isdigit() else 0
+        from changedetectionio.llm.evaluator import _get_max_input_chars, _DEFAULT_MAX_INPUT_CHARS
+        llm_effective_max_input_chars = _get_max_input_chars(datastore)
+        # Cost display: only when user configured their own key (not hosted/operator-managed)
+        llm_show_costs = not llm_env_configured
+
        output = render_template("settings.html",
                                active_plugins=active_plugins,
                                api_key=datastore.data['settings']['application'].get('api_access_token'),
+                                llm_config=llm_config,
+                                llm_env_configured=llm_env_configured,
+                                llm_stored=llm_stored,
+                                llm_token_budget_month=llm_token_budget_month,
+                                llm_token_budget_month_env=llm_token_budget_month_env,
+                                llm_max_input_chars_env=llm_max_input_chars_env,
+                                llm_effective_max_input_chars=llm_effective_max_input_chars,
+                                llm_show_costs=llm_show_costs,
                                python_version=python_version,
                                uptime_seconds=uptime_seconds,
                                available_timezones=sorted(available_timezones()),
@@ -0,0 +1,269 @@
+import json
+import logging
+import os
+import re
+
+from flask import Blueprint, jsonify, redirect, url_for, flash
+from flask_babel import gettext
+from loguru import logger
+
+from changedetectionio.store import ChangeDetectionStore
+from changedetectionio.auth_decorator import login_optionally_required
+
+
+class _LiteLLMWarningCapture(logging.Handler):
+    """Capture warnings emitted on the 'LiteLLM' stdlib logger during a single call.
+
+    litellm.get_valid_models() catches HTTP/auth errors internally, logs a warning,
+    and returns []. Without capturing that warning we can't tell the user *why*
+    no models came back (bad key vs. offline vs. genuinely empty model list).
+    """
+    def __init__(self):
+        super().__init__(level=logging.WARNING)
+        self.messages = []
+
+    def emit(self, record):
+        try:
+            self.messages.append(record.getMessage())
+        except Exception:
+            pass
+
+
+def _humanize_litellm_error(raw: str) -> str:
+    # litellm warnings typically look like:
+    #   "Error getting valid models: Failed to get models: { 'error': { 'message': '...' } }"
+    # Pull the inner provider message when present; otherwise trim the boilerplate.
+    if not raw:
+        return raw
+    m = re.search(r'\{.*\}', raw, re.DOTALL)
+    if m:
+        try:
+            body = json.loads(m.group(0))
+            inner = (body.get('error') or {}).get('message') or body.get('message')
+            if inner:
+                return inner
+        except Exception:
+            pass
+    cleaned = re.sub(r'^Error getting valid models:\s*', '', raw)
+    cleaned = re.sub(r'^Failed to get models:\s*', '', cleaned).strip()
+    return cleaned[:500]
+
+
+def construct_llm_blueprint(datastore: ChangeDetectionStore):
+    llm_blueprint = Blueprint('llm', __name__)
+
+    @llm_blueprint.route("/models", methods=['GET'])
+    @login_optionally_required
+    def llm_get_models():
+        from flask import request
+        from changedetectionio.validate_url import is_llm_api_base_safe
+        provider = request.args.get('provider', '').strip()
+        api_key  = request.args.get('api_key',  '').strip()
+        api_base = request.args.get('api_base', '').strip()
+
+        logger.debug(f"LLM model list requested for provider={provider!r} api_base={api_base!r}")
+
+        if not provider:
+            logger.debug("LLM model list: no provider specified, returning 400")
+            return jsonify({'models': [], 'error': 'No provider specified'}), 400
+
+        ok, reason = is_llm_api_base_safe(api_base)
+        if not ok:
+            logger.warning(f"LLM model list refused: api_base failed SSRF check ({reason})")
+            return jsonify({'models': [], 'error': reason}), 400
+
+        # Credential-exfiltration guard (GHSA-g36r-fm2p-87xm).
+        # Only substitute the stored api_key when api_base matches the stored
+        # api_base. If the caller pointed at a different destination, refuse —
+        # otherwise a CSRF / unauthenticated request can ship the operator's
+        # long-lived provider key (sent as Authorization: Bearer …) to an
+        # attacker-controlled URL.
+        stored_llm     = datastore.data['settings']['application'].get('llm') or {}
+        stored_api_base = (stored_llm.get('api_base') or '').strip()
+        if not api_key:
+            if api_base == stored_api_base:
+                api_key = (stored_llm.get('api_key') or '')
+                logger.debug("LLM model list: no api_key in request, using stored key (api_base matches saved)")
+            elif api_base:
+                logger.warning("LLM model list refused: api_base differs from saved config but no api_key supplied")
+                return jsonify({'models': [], 'error': gettext(
+                    "api_key is required when api_base differs from the saved configuration. "
+                    "Refusing to send the stored API key to a different endpoint."
+                )}), 400
+
+        _PREFIXES = {'gemini': 'gemini/', 'ollama': 'ollama/', 'openrouter': 'openrouter/',
+                     'openai_compatible': 'openai/'}
+        # vLLM / LM Studio / llama.cpp speak OpenAI's wire format — route through litellm's
+        # 'openai' provider but keep the UI-level name distinct from cloud OpenAI.
+        _LITELLM_PROVIDER = {'openai_compatible': 'openai'}
+        prefix = _PREFIXES.get(provider, '')
+        litellm_provider = _LITELLM_PROVIDER.get(provider, provider)
+
+        try:
+            import litellm
+            logger.debug(f"LLM model list: calling litellm.get_valid_models provider={provider!r} (litellm={litellm_provider!r}) api_base={api_base!r}")
+
+            capture = _LiteLLMWarningCapture()
+            litellm_logger = logging.getLogger('LiteLLM')
+            litellm_logger.addHandler(capture)
+            try:
+                raw = litellm.get_valid_models(
+                    check_provider_endpoint=True,
+                    custom_llm_provider=litellm_provider,
+                    api_key=api_key or None,
+                    api_base=api_base or None,
+                ) or []
+            finally:
+                litellm_logger.removeHandler(capture)
+
+            models = sorted({(m if m.startswith(prefix) else prefix + m) for m in raw})
+
+            if not models and capture.messages:
+                err = _humanize_litellm_error(capture.messages[-1])
+                logger.debug(f"LLM model list: 0 models, surfacing captured litellm warning: {err!r}")
+                return jsonify({'models': [], 'error': err}), 400
+
+            logger.debug(f"LLM model list: got {len(models)} models for provider={provider!r}")
+            return jsonify({'models': models, 'error': None})
+        except Exception as e:
+            logger.error(f"LLM model list failed for provider={provider!r}: {e}")
+            logger.exception("LLM model list full traceback:")
+            return jsonify({'models': [], 'error': str(e)}), 400
+
+    @llm_blueprint.route("/test", methods=['GET'])
+    @login_optionally_required
+    def llm_test():
+        from flask import request
+        from changedetectionio.llm.client import completion
+        from changedetectionio.validate_url import is_llm_api_base_safe
+
+        # Pull stored config as the fallback, then override with anything the
+        # form-driven JS sent as query params. Lets users test config changes
+        # without first hitting Save (matching how /settings/llm/models works).
+        stored = datastore.data['settings']['application'].get('llm') or {}
+        # Keep the raw request-supplied values around so we can detect whether
+        # the caller explicitly steered api_base / api_key (credential-exfil guard below).
+        req_api_key  = (request.args.get('api_key')  or '').strip()
+        req_api_base = (request.args.get('api_base') or '').strip()
+        stored_api_base = (stored.get('api_base') or '').strip()
+        llm_cfg = {
+            'model':                   (request.args.get('model')                   or stored.get('model', '')).strip(),
+            'api_key':                 (req_api_key  or stored.get('api_key', '')).strip(),
+            'api_base':                (req_api_base or stored_api_base).strip(),
+            'provider_kind':           (request.args.get('provider_kind')           or stored.get('provider_kind', '')).strip(),
+            'local_token_multiplier':   request.args.get('local_token_multiplier')  or stored.get('local_token_multiplier'),
+        }
+        model    = llm_cfg['model']
+        api_base = llm_cfg['api_base']
+
+        logger.debug(
+            f"LLM connection test requested: model={model!r} api_base={api_base!r} "
+            f"provider_kind={llm_cfg['provider_kind']!r} "
+            f"source={'form' if request.args.get('model') else 'datastore'}"
+        )
+
+        if not model:
+            logger.error("LLM connection test failed: no model configured")
+            return jsonify({'ok': False, 'error': 'No model configured.'}), 400
+
+        ok, reason = is_llm_api_base_safe(api_base)
+        if not ok:
+            logger.warning(f"LLM connection test refused: api_base failed SSRF check ({reason})")
+            return jsonify({'ok': False, 'error': reason}), 400
+
+        # Credential-exfiltration guard (GHSA-g36r-fm2p-87xm).
+        # If the caller specified an api_base that differs from the saved one but
+        # did NOT supply a matching api_key, refuse to substitute the stored key.
+        # Otherwise a CSRF / unauthenticated request can route the operator's
+        # long-lived provider key to an attacker-controlled endpoint.
+        if req_api_base and req_api_base != stored_api_base and not req_api_key:
+            logger.warning("LLM connection test refused: api_base differs from saved config but no api_key supplied")
+            return jsonify({'ok': False, 'error': gettext(
+                "api_key is required when api_base differs from the saved configuration. "
+                "Refusing to send the stored API key to a different endpoint."
+            )}), 400
+
+        try:
+            logger.debug(f"LLM connection test: sending test prompt to model={model!r}")
+            # Reuse the same multiplier path the production calls use, so cloud providers
+            # stay on a small base cap (matching upstream's pre-existing behavior) and only
+            # reasoning-capable endpoints (Ollama, openai_compatible) opt into the extra
+            # headroom needed for chain-of-thought to complete.
+            # Timeout: omit the override so the test inherits DEFAULT_TIMEOUT (60s, tunable
+            # via LLM_TIMEOUT). A shorter test-only timeout falsely fails on cold-starting
+            # cloud reasoning models (e.g. ollama.com hosting qwen3.5:397b takes ~60s on
+            # first hit) even though the same call succeeds in production.
+            from changedetectionio.llm.evaluator import apply_local_token_multiplier, get_llm_settings
+            text, total_tokens, input_tokens, output_tokens = completion(
+                model=model,
+                messages=[{'role': 'user', 'content':
+                    'Respond with just the word: ready'}],
+                api_key=llm_cfg.get('api_key') or None,
+                api_base=api_base or None,
+                max_tokens=apply_local_token_multiplier(200, llm_cfg),
+                debug=get_llm_settings(datastore).debug,
+            )
+            reply = text.strip()
+            if not reply:
+                logger.warning(
+                    f"LLM connection test: model={model!r} responded but returned empty content "
+                    f"tokens={total_tokens} (in={input_tokens} out={output_tokens}) — "
+                    f"check finish_reason in client debug log above"
+                )
+                return jsonify({'ok': False, 'error': 'Model responded but returned empty content — check server logs.'}), 400
+
+            logger.success(
+                f"LLM connection test OK: model={model!r} "
+                f"tokens={total_tokens} (in={input_tokens} out={output_tokens}) "
+                f"reply={reply!r}"
+            )
+            return jsonify({'ok': True, 'text': reply, 'tokens': total_tokens})
+
+        except Exception as e:
+            logger.error(f"LLM connection test FAILED: model={model!r} api_base={api_base!r} error={e}")
+            logger.exception("LLM connection test full traceback:")
+            return jsonify({'ok': False, 'error': str(e)}), 400
+
+    # Both clear endpoints accept POST only — GET would let an attacker fire them via
+    # <img src="...">, wiping LLM configuration / cached summaries on a logged-in
+    # operator's browser (GHSA-g36r-fm2p-87xm). Flask-WTF CSRFProtect enforces a
+    # CSRF token on POST automatically; the template renders csrf_token() inside the
+    # surrounding <form>.
+    @llm_blueprint.route("/clear", methods=['POST'])
+    @login_optionally_required
+    def llm_clear():
+        from changedetectionio.model.LLMSettings import LLMSettings
+        logger.debug("LLM configuration cleared by user")
+        # Read existing config, write back a dict that omits the connection fields —
+        # so the saved dict no longer has model/api_key/api_base/etc.
+        # Toggles, prompts, budgets and counters survive.
+        settings = LLMSettings.model_validate(
+            datastore.data['settings']['application'].get('llm') or {}
+        )
+        datastore.data['settings']['application']['llm'] = settings.model_dump(
+            exclude=set(LLMSettings.CONNECTION_FIELDS)
+        )
+        datastore.commit()
+        flash(gettext("AI / LLM configuration removed."), 'notice')
+        return redirect(url_for('settings.settings_page') + '#ai')
+
+    @llm_blueprint.route("/clear-summary-cache", methods=['POST'])
+    @login_optionally_required
+    def llm_clear_summary_cache():
+        import glob
+        count = 0
+        for watch in datastore.data['watching'].values():
+            if not watch.data_dir:
+                continue
+            for f in glob.glob(os.path.join(watch.data_dir, 'change-summary-*.txt')):
+                try:
+                    os.remove(f)
+                    logger.info(f"LLM summary cache removed: {f}")
+                    count += 1
+                except OSError as e:
+                    logger.warning(f"Could not remove LLM summary cache file {f}: {e}")
+        logger.info(f"LLM summary cache cleared: {count} file(s) removed")
+        flash(gettext("AI summary cache cleared ({} file(s) removed).").format(count), 'notice')
+        return redirect(url_for('settings.settings_page') + '#ai')
+
+    return llm_blueprint
@@ -9,6 +9,7 @@
    const email_notification_prefix=JSON.parse('{{emailprefix|tojson}}');
 {% endif %}
 </script>
+<script src="{{url_for('static_content', group='js', filename='modal.js')}}"></script>
 <script src="{{url_for('static_content', group='js', filename='tabs.js')}}" defer></script>
 <script src="{{url_for('static_content', group='js', filename='plugins.js')}}" defer></script>
 <script src="{{url_for('static_content', group='js', filename='notifications.js')}}" defer></script>
@@ -23,9 +24,9 @@
            <li class="tab"><a href="#fetching">{{ _('Fetching') }}</a></li>
            <li class="tab"><a href="#filters">{{ _('Global Filters') }}</a></li>
            <li class="tab"><a href="#ui-options">{{ _('UI Options') }}</a></li>
-            <li class="tab"><a href="#api">{{ _('API') }}</a></li>
-            <li class="tab"><a href="#rss">{{ _('RSS') }}</a></li>
-            <li class="tab"><a href="{{ url_for('backups.index') }}">{{ _('Backups') }}</a></li>
+            <li class="tab"><a href="#api">API</a></li>
+            <li class="tab"><a href="#rss">RSS</a></li>
+            <li class="tab"><a href="{{ url_for('backups.create') }}">{{ _('Backups') }}</a></li>
            <li class="tab"><a href="#timedate">{{ _('Time & Date') }}</a></li>
            <li class="tab"><a href="#proxies">{{ _('CAPTCHA & Proxies') }}</a></li>
            {% if plugin_tabs %}
@@ -33,6 +34,9 @@
            <li class="tab"><a href="#plugin-{{ tab.plugin_id }}">{{ tab.tab_label }}</a></li>
                {% endfor %}
            {% endif %}
+            {% if not llm_features_disabled %}
+            <li class="tab"><a href="#ai">{{ _('AI / LLM') }}</a></li>
+            {% endif %}
            <li class="tab"><a href="#info">{{ _('Info') }}</a></li>
        </ul>
    </div>
@@ -56,7 +60,7 @@
                        {{ render_field(form.application.form.filter_failure_notification_threshold_attempts, class="filter_failure_notification_threshold_attempts") }}
                        <span class="pure-form-message-inline">{{ _('After this many consecutive times that the CSS/xPath filter is missing, send a notification') }}
                            <br>
-                        {{ _('Set to') }} <strong>0</strong> {{ _('to disable') }}
+                        {{ _('Set to <strong>0</strong> to disable')|safe }}
                        </span>
                    </div>
                    <div class="pure-control-group">
@@ -118,15 +122,15 @@
                <div class="pure-control-group inline-radio">
                    {{ render_field(form.application.form.fetch_backend, class="fetch-backend") }}
                    <span class="pure-form-message-inline">
-                        <p>{{ _('Use the') }} <strong>{{ _('Basic') }}</strong> {{ _('method (default) where your watched sites don\'t need Javascript to render.') }}</p>
-                        <p>{{ _('The') }} <strong>{{ _('Chrome/Javascript') }}</strong> {{ _('method requires a network connection to a running WebDriver+Chrome server, set by the ENV var') }} 'WEBDRIVER_URL'. </p>
+                        <p>{{ _('Use the <strong>Basic</strong> method (default) where your watched sites don\'t need Javascript to render.')|safe }}</p>
+                        <p>{{ _('The <strong>Chrome/Javascript</strong> method requires a network connection to a running WebDriver+Chrome server, set by the ENV var \'WEBDRIVER_URL\'.')|safe }}</p>
                    </span>
                </div>
                <fieldset class="pure-group" id="webdriver-override-options" data-visible-for="application-fetch_backend=html_webdriver">
                    <div class="pure-form-message-inline">
                        <strong>{{ _('If you\'re having trouble waiting for the page to be fully rendered (text missing etc), try increasing the \'wait\' time here.') }}</strong>
                        <br>
-                        {{ _('This will wait') }} <i>n</i> {{ _('seconds before extracting the text.') }}
+                        {{ _('This will wait <i>n</i> seconds before extracting the text.')|safe }}
                    </div>
                    <div class="pure-control-group">
                        {{ render_field(form.application.form.webdriver_delay) }}
@@ -154,9 +158,8 @@
                    </span>
                </div>
                <div class="pure-control-group">
-                                        <br>
-                    {{ _('Tip:') }} <a href="https://github.com/dgtlmoon/changedetection.io/wiki/Proxy-configuration#brightdata-proxy-support">{{ _('Connect using Bright Data and Oxylabs Proxies, find out more here.') }}</a>
-
+                <br>
+                    {{ _('Tip:') }} <a href="{{ url_for('settings.settings_page')}}#proxies">{{ _('Connect using Bright Data proxies, find out more here.') }}</a>
                </div>
            </div>

@@ -196,7 +199,7 @@ nav
                    <span class="pure-form-message-inline">{{ _('Note: This is applied globally in addition to the per-watch rules.') }}</span><br>
                    <span class="pure-form-message-inline">
                        <ul>
-                            <li>{{ _('Matching text will be') }} <strong>{{ _('ignored') }}</strong> {{ _('in the text snapshot (you can still see it but it wont trigger a change)') }}</li>
+                            <li>{{ _('Matching text will be ignored in the text snapshot (you can still see it but it wont trigger a change)') }}</li>
                            <li>{{ _('Note: This is applied globally in addition to the per-watch rules.') }}</li>
                            <li>{{ _('Each line processed separately, any line matching will be ignored (removed before creating the checksum)') }}</li>
                            <li>{{ _('Regular Expression support, wrap the entire line in forward slash') }} <code>/regex/</code></li>
@@ -264,7 +267,7 @@ nav
                    </div>
                    <div>
                       {{ render_field(form.application.form.rss_template_override) }}
-                        {{ show_token_placeholders(extra_notification_token_placeholder_info=extra_notification_token_placeholder_info, suffix="-rss") }}
+                        {{ show_token_placeholders(extra_notification_token_placeholder_info=extra_notification_token_placeholder_info, suffix="-rss", settings_application=settings_application) }}
                    </div>
                </div>
                <br>
@@ -352,7 +355,7 @@ nav
                    </div>
                </div>

-                <p><strong>{{ _('Tip') }}</strong>: {{ _('"Residential" and "Mobile" proxy type can be more successfull than "Data Center" for blocked websites.') }}</p>
+                <p><strong>{{ _('Tip') }}</strong>: {{ _('"Residential" and "Mobile" proxy type can be more successful than "Data Center" for blocked websites.') }}</p>

                <div class="pure-control-group" id="extra-proxies-setting">
                {{ render_fieldlist_with_inline_errors(form.requests.form.extra_proxies) }}
@@ -393,6 +396,9 @@ nav
            </div>
                {% endfor %}
            {% endif %}
+            {% if not llm_features_disabled %}
+            {% include 'settings_llm_tab.html' %}
+            {% endif %}
            <div class="tab-pane-inner" id="info">
                <p><strong>{{ _('Uptime:') }}</strong> {{ uptime_seconds|format_duration }}</p>
                <p><strong>{{ _('Python version:') }}</strong> {{ python_version }}</p>
@@ -0,0 +1,611 @@
+{% from '_helpers.html' import render_field %}
+{% from '_stab.html' import stab_shell, stab_pane %}
+{#
+  AI / LLM settings tab content — included from settings.html.
+  Requires template context: form, llm_config, llm_env_configured
+#}
+<div class="tab-pane-inner" id="ai">
+<script src="{{ url_for('static_content', group='js', filename='sub-tabs.js') }}"></script>
+
+{# TRANSLATORS: 'Usage' here means token consumption/cost stats for the AI provider, not a how-to guide #}
+{% set _usage_label = pgettext('AI usage stats', 'Usage') %}
+{% call stab_shell('ai-settings', [
+    {'id': 'overview',  'label': _('Overview'),  'icon': '✦'},
+    {'id': 'provider',  'label': _('Provider'),  'icon': '⚙'},
+    {'id': 'prompts',   'label': _('Prompts'),   'icon': '≡'},
+    {'id': 'behaviour', 'label': _('Behaviour'), 'icon': '⚑'},
+    {'id': 'usage',     'label': _usage_label,   'icon': '$'},
+]) %}
+
+  {# ── Overview ──────────────────────────────────────────────────────────── #}
+  {% call stab_pane('overview') %}
+  <div class="stab-overview-hero">
+    <h3><span class="stab-overview-glyph">✦</span> {{ _('AI-powered change monitoring') }}</h3>
+    <p>{{ _('Connect an LLM to move from "something changed" to "only the thing you care about changed".') }}</p>
+  </div>
+
+  <div class="stab-overview-features">
+    <div class="stab-overview-feature">
+      <div class="stab-overview-icon">⊞</div>
+      <div class="stab-overview-text">
+        <strong>{{ _('Intent filtering') }}</strong>
+        <p>{{ _('Each watch or tag can carry a plain-text intent — %(ex1)s or %(ex2)s. On every detected change the AI evaluates the diff against it and suppresses irrelevant noise.', ex1='<strong>"notify me only when the price drops"</strong>', ex2='<strong>"alert when the item goes out of stock"</strong>') | safe }}</p>
+        <p><small>{{ _('Tip: intent evaluation benefits from a capable model — recommended %(local)s locally, or %(gpt)s / %(gemini)s. Very small models (≤3B) may misjudge numeric comparisons.',
+              local='<code>qwen2.5:7b</code>',
+              gpt='<code>gpt-4o-mini</code>',
+              gemini='<code>gemini-2.0-flash</code>') | safe }}</small></p>
+      </div>
+    </div>
+    <div class="stab-overview-feature">
+      <div class="stab-overview-icon">≡</div>
+      <div class="stab-overview-text">
+        <strong>{{ _('AI Change Summary') }}</strong>
+        <p>{{ _('Instead of raw diffs, receive plain-language summaries in notifications — %(ex1)s or %(ex2)s. Set a global default prompt here, or override per watch or tag.', ex1='<strong>"Price dropped from $89 to $67"</strong>', ex2='<strong>"3 new items added to the listing"</strong>') | safe }}</p>
+      </div>
+    </div>
+    <div class="stab-overview-feature">
+      <div class="stab-overview-icon">≈</div>
+      <div class="stab-overview-text">
+        <strong>{{ _('Minimal cost') }}</strong>
+        <p>{{ _('The AI sees only a unified diff of what changed — never full page HTML. Low-cost models like %(gpt)s or %(gemini)s handle this well, typically fractions of a cent per check.',
+              gpt='<a href="https://platform.openai.com/api-keys" target="_blank" rel="noopener">gpt-4o-mini</a>',
+              gemini='<a href="https://aistudio.google.com/apikey" target="_blank" rel="noopener">Gemini Flash</a>') | safe }}</p>
+      </div>
+    </div>
+  </div>
+
+  <div class="stab-overview-cta">
+    {% if llm_config and llm_config.get('model') %}
+    <span class="stab-configured-badge">&#10003; {{ _('AI / LLM configured:') }} {{ llm_config.get('model') }}</span>
+    {% else %}
+    <button type="button" class="pure-button pure-button-primary" data-stab-goto="provider">
+      ⚙ {{ _('Configure AI Provider') }} &rarr;
+    </button>
+    {% endif %}
+  </div>
+  {% endcall %}
+
+  {# ── Provider ──────────────────────────────────────────────────────────── #}
+  {% call stab_pane('provider') %}
+  <p class="stab-section-title">{{ _('AI Provider') }}</p>
+
+  <div class="pure-control-group">
+    <label></label>
+    {{ form.llm.form.enabled() }}
+    <label for="{{ form.llm.form.enabled.id }}" style="display:inline; font-weight:normal;">
+      {{ form.llm.form.enabled.label.text }}
+    </label>
+    <span class="pure-form-message-inline">
+      {{ _('Master switch — when off, all AI lookups are skipped even if a provider is configured below.') }}
+    </span>
+  </div>
+
+  {% if not llm_env_configured and not (llm_config and llm_config.get('model')) %}
+  <div class="stab-overview-disclaimer">
+    <div class="stab-disclaimer-icon">⚠</div>
+    <div class="stab-disclaimer-body">
+      <strong>{{ _('Third-party data transfer — please read') }}</strong>
+      <p>{{ _('When AI features are active, change data from the websites you monitor — including page diffs and extracted text — is sent to an external AI provider of your choice.') }}</p>
+      <ul>
+        <li>{{ _('You are solely responsible for ensuring this complies with the terms of service of each website you monitor.') }}</li>
+        <li>{{ _("You are solely responsible for compliance with applicable data-protection laws (e.g. GDPR) regarding any personal data that may appear in monitored content.") }}</li>
+        <li>{{ _('API costs charged by your chosen provider are your own responsibility; this software has no visibility into or control over those charges.') }}</li>
+        <li>{{ _('AI / LLM models are known to hallucinate — producing plausible-sounding but factually incorrect or entirely fabricated output with apparent confidence — and by design may omit or truncate relevant data during summarisation. AI output must never be relied upon as complete or accurate. This software is provided as-is with no warranty of any kind.') }}</li>
+        <li>{{ _('By enabling AI features you personally indemnify and hold harmless the creator(s) and contributor(s) of this software from any claims, damages, or liability arising from this data transfer or your use of AI features.') }}</li>
+      </ul>
+      <label class="stab-disclaimer-check">
+        <input type="checkbox" id="llm-disclaimer-accept" onchange="llmDisclaimerToggle(this)">
+        <span>{{ _('I have read and understood the above. I accept full responsibility and indemnify the creator(s) of this software.') }}</span>
+      </label>
+    </div>
+  </div>
+  <div id="llm-provider-fields" style="display:none">
+  {% endif %}
+
+  {% if llm_env_configured %}
+  <div class="inline-warning" style="margin-bottom: 1em;">
+    <img class="inline-warning-icon" src="{{ url_for('static_content', group='images', filename='notice.svg') }}" alt="{{ _('Note') }}">
+    {{ _('AI / LLM is configured via environment variables (<code>LLM_MODEL=%(model)s</code>%(api_base)s). Remove the <code>LLM_MODEL</code> environment variable to configure via this form instead.',
+         model=llm_config.get('model', '')|e,
+         api_base=(', <code>LLM_API_BASE=' ~ (llm_config.get('api_base')|e) ~ '</code>') if llm_config.get('api_base') else '') | safe }}
+  </div>
+  {% else %}
+
+  <div class="pure-control-group">
+    <label for="llm-provider">{{ _('Provider') }}</label>
+    <select id="llm-provider" onchange="llmOnProviderChange(this.value)">
+      <option value="">— {{ _('select a provider') }} —</option>
+        <option value="anthropic">Anthropic</option>
+        <option value="gemini">Google (Gemini)</option>
+        <option value="ollama">Ollama</option>
+        <option value="openai">OpenAI</option>
+        <option value="openai_compatible">{{ _('OpenAI-compatible (vLLM, LM Studio, llama.cpp)') }}</option>
+        <option value="openrouter">OpenRouter (200+ models)</option>
+    </select>
+  </div>
+
+  <div class="pure-control-group">
+    {{ render_field(form.llm.form.api_key) }}
+    <span class="pure-form-message-inline" id="llm-key-hint"></span>
+  </div>
+  <div class="pure-control-group" id="llm-base-group" style="display:none">
+    {{ render_field(form.llm.form.api_base) }}
+    <span class="pure-form-message-inline">{{ _('Only needed for Ollama or custom/self-hosted endpoints. Leave blank for cloud providers.') }}</span>
+  </div>
+
+  {# Hidden field carrying the dropdown selection so the backend knows when to apply
+     reasoning-friendly token caps (Ollama and OpenAI-compatible endpoints, which commonly
+     serve reasoning models that need headroom for chain-of-thought to complete). #}
+  {{ form.llm.form.provider_kind() }}
+
+  <div class="pure-control-group" id="llm-local-advanced-group" style="display:none">
+    <label for="{{ form.llm.form.local_token_multiplier.id }}">{{ form.llm.form.local_token_multiplier.label.text }}</label>
+    {{ form.llm.form.local_token_multiplier() }}
+    <span class="pure-form-message-inline">
+      {{ _('Reasoning models (Qwen3, DeepSeek-R1, Gemma 3, etc.) emit chain-of-thought before the final answer. This multiplier scales every <code>max_tokens</code> cap for this endpoint to leave reasoning room. Defaults to %(default)s; raise it if responses come back truncated or empty, lower it (down to 1x) if you want tighter limits on a paid endpoint. Applied to Ollama and OpenAI-compatible endpoints — other cloud providers (OpenAI, Anthropic, Gemini) keep their original tight caps.', default='5x') | safe }}
+    </span>
+  </div>
+
+  <div class="pure-control-group" id="llm-fetch-group" style="display:none">
+    <label></label>
+    <button type="button" id="llm-fetch-btn" class="pure-button button-xsmall" onclick="llmFetchModels()"
+            style="background:#27ae60;color:#fff;border:none;">
+      &#8635; {{ _('Load available models') }}
+    </button>
+    <span id="llm-fetch-status" style="margin-left:.6em;font-size:.85em;color:#888;"></span>
+  </div>
+
+  <div class="pure-control-group" id="llm-model-select-group" style="display:none">
+    <label for="llm-model-select">{{ _('Available models') }}</label>
+    <select id="llm-model-select" class="pure-input-1-2" onchange="llmOnModelPick(this.value)">
+      <option value="">— {{ _('choose a model') }} —</option>
+    </select>
+  </div>
+
+  <div class="pure-control-group">
+    {{ render_field(form.llm.form.model,
+                    placeholder=_("Enter API key and click 'Load available models'")) }}
+  </div>
+
+  {% if llm_config and llm_config.get('model') %}
+  <div class="pure-control-group">
+    <label></label>
+    <span style="color:#4a7c59;font-weight:bold;">
+      &#10003; {{ _('AI / LLM configured:') }} {{ llm_config.get('model') }}
+    </span>
+    &nbsp;
+    {# data-method="POST" tells modal.js to POST with the CSRF token instead of
+       navigating — GET previously allowed <img>-based CSRF wipe (GHSA-g36r-fm2p-87xm).
+       Stays as <a> because we're inside the outer settings <form> — nested forms are
+       invalid HTML, so modal.js builds a body-level hidden form for the POST. #}
+    <a href="{{ url_for('settings.llm.llm_clear') }}"
+       class="pure-button button-xsmall"
+       style="background:#c0392b;color:#fff;"
+       data-method="POST"
+       data-requires-confirm
+       data-confirm-type="danger"
+       data-confirm-title="{{ _('Remove AI / LLM configuration?') }}"
+       data-confirm-message="<p>{{ _('This will remove your saved AI provider, model, and API key.') }}</p>"
+       data-confirm-button="{{ _('Remove') }}"
+       data-cancel-button="{{ _('Cancel') }}">
+      &#10005; {{ _('Remove') }}
+    </a>
+    &nbsp;
+    <button type="button" id="llm-test-btn" class="pure-button button-xsmall" onclick="llmRunTest()"
+            style="background:#2980b9;color:#fff;border:none;">
+      &#9654; {{ _('Test connection') }}
+    </button>
+  </div>
+  <div id="llm-test-result" style="display:none; margin-top:0.6em; padding:0.6em 0.85em; border-radius:5px; font-size:0.88em; line-height:1.45;"></div>
+  {% endif %}
+
+  <p class="pure-form-message-inline" style="margin-top:0.5em;">
+    {{ _("Your API key is stored locally and sent only to your chosen provider. On each detected change, the watch's diff and extracted text are sent to the LLM — no full page HTML.") }}
+  </p>
+
+  <div class="pure-control-group" style="margin-top:1.2em; padding-top:1em; border-top:1px solid rgba(128,128,128,0.15);">
+    <label style="color:#888; font-size:0.85em;">{{ _('Cache') }}</label>
+    {# See comment above on data-method="POST"+modal.js (GHSA-g36r-fm2p-87xm). #}
+    <a href="{{ url_for('settings.llm.llm_clear_summary_cache') }}"
+       class="pure-button button-xsmall"
+       style="background:#7f8c8d;color:#fff;"
+       data-method="POST"
+       data-requires-confirm
+       data-confirm-type="warning"
+       data-confirm-title="{{ _('Clear all summary cache?') }}"
+       data-confirm-message="<p>{{ _('This will remove all cached AI change summaries across all watches.') }}</p><p>{{ _('They will be regenerated on the next check.') }}</p>"
+       data-confirm-button="{{ _('Clear cache') }}"
+       data-cancel-button="{{ _('Cancel') }}">
+      &#10005; {{ _('Clear all summary cache') }}
+    </a>
+    <span class="pure-form-message-inline">{{ _('Removes all cached AI change summaries across all watches. They will be regenerated on the next check.') }}</span>
+  </div>
+
+  <div class="pure-control-group">
+    <label></label>
+    {{ form.llm.form.debug() }}
+    <label for="{{ form.llm.form.debug.id }}" style="display:inline; font-weight:normal;">
+      {{ form.llm.form.debug.label.text }}
+    </label>
+    <span class="pure-form-message-inline">
+      {{ _('Enables litellm verbose output (routed through loguru). Useful when diagnosing provider errors or empty responses. Leave off in production — generates a lot of log volume.') }}
+    </span>
+  </div>
+  {% endif %}{# llm_env_configured #}
+
+  {% if not llm_env_configured and not (llm_config and llm_config.get('model')) %}
+  </div>{# llm-provider-fields #}
+  {% endif %}
+  {% endcall %}
+
+  {# ── Prompts ───────────────────────────────────────────────────────────── #}
+  {% call stab_pane('prompts') %}
+  <p class="stab-section-title">{{ _('Default AI Change Summary') }}</p>
+
+  <div class="pure-control-group">
+    {{ render_field(form.llm.form.change_summary_default) }}
+    <span class="pure-form-message-inline">
+      {{ _('Used for all watches unless overridden by the watch or its tag/group.') }}
+      &nbsp;<a href="#" class="pure-button button-small" onclick="var t=document.getElementById('llm-change_summary_default'); if(!t.value && t.placeholder) t.value=t.placeholder; return false;">{{ _('Modify default prompt') }}</a>
+    </span>
+  </div>
+
+  {% endcall %}
+
+  {# ── Behaviour ─────────────────────────────────────────────────────────── #}
+  {% call stab_pane('behaviour') %}
+  <p class="stab-section-title">{{ _('Behaviour') }}</p>
+
+  {% if llm_config and llm_config.get('model') %}
+  <div class="pure-control-group">
+    <label></label>
+    {{ form.llm.form.override_diff_with_summary() }}
+    <label for="{{ form.llm.form.override_diff_with_summary.id }}" style="display:inline; font-weight:normal;">
+      {{ form.llm.form.override_diff_with_summary.label.text }}
+    </label>
+    <span class="pure-form-message-inline">
+      {{ _('When enabled, the <code>%(diff)s</code> notification token shows the AI summary instead of the raw diff. Use <code>%(raw_diff)s</code> to always get the original.',
+           diff='{{diff}}', raw_diff='{{raw_diff}}') | safe }}
+    </span>
+  </div>
+
+  <div class="pure-control-group">
+    <label></label>
+    {{ form.llm.form.restock_use_fallback_extract() }}
+    <label for="{{ form.llm.form.restock_use_fallback_extract.id }}" style="display:inline; font-weight:normal;">
+      {{ form.llm.form.restock_use_fallback_extract.label.text }}
+    </label>
+    <span class="pure-form-message-inline">
+      {{ _('When enabled, the AI will be used as a last resort to extract price and stock status from product pages where no structured metadata (JSON-LD, microdata, OpenGraph) is found.') }}
+    </span>
+  </div>
+
+  <div class="pure-control-group">
+    <label for="{{ form.llm.form.thinking_budget.id }}">{{ form.llm.form.thinking_budget.label.text }}</label>
+    {{ form.llm.form.thinking_budget() }}
+    <span class="pure-form-message-inline">{{ _('For Gemini 2.5+ models only. Thinking tokens improve reasoning quality but count against the output budget. Set to Off if summaries are being cut short.') }}</span>
+  </div>
+
+  <div class="pure-control-group">
+    <label for="{{ form.llm.form.max_summary_tokens.id }}">{{ form.llm.form.max_summary_tokens.label.text }}</label>
+    {{ form.llm.form.max_summary_tokens() }}
+    <span class="pure-form-message-inline">{{ _('Upper limit on tokens the AI may use when writing a change summary. Higher values allow longer summaries but cost more.') }}</span>
+  </div>
+
+  <div class="pure-control-group">
+    <label>{{ form.llm.form.budget_action.label.text }}</label>
+    <div>
+      {% for subfield in form.llm.form.budget_action %}
+      <label class="pure-radio" style="display:block; font-weight:normal; margin-bottom:0.3em;">
+        {{ subfield() }} {{ subfield.label.text }}
+      </label>
+      {% endfor %}
+    </div>
+  </div>
+  {% else %}
+  <p class="pure-form-message-inline" style="margin-top:0.5em;">
+    {{ _('Configure a provider first to unlock behaviour settings.') }}
+  </p>
+  {% endif %}
+  {% endcall %}
+
+  {# ── Usage ─────────────────────────────────────────────────────────────── #}
+  {% call stab_pane('usage') %}
+  <p class="stab-section-title">{{ _('Token & Cost Tracking') }}</p>
+
+  {% if llm_stored.get('tokens_total_cumulative') or llm_stored.get('tokens_this_month') %}
+
+  <div class="llm-usage-grid">
+    <div class="llm-stat-card">
+      <div class="llm-stat-label">{{ _('This month') }}</div>
+      <div class="llm-stat-value">{{ '{:,}'.format(llm_stored.get('tokens_this_month', 0)) }}</div>
+      <div class="llm-stat-sub">{{ _('tokens') }}{% if llm_show_costs and llm_stored.get('cost_usd_this_month') %} &nbsp;·&nbsp; ≈&thinsp;${{ '%.4f'|format(llm_stored.get('cost_usd_this_month', 0)) }}{% endif %}</div>
+      {% if llm_token_budget_month %}
+      {% set pct = (llm_stored.get('tokens_this_month', 0) / llm_token_budget_month * 100)|int %}
+      <div class="llm-stat-bar-wrap">
+        <div class="llm-stat-bar-fill {% if pct >= 100 %}bar-over{% elif pct >= 80 %}bar-warn{% else %}bar-ok{% endif %}"
+             style="width:{{ [pct, 100]|min }}%"></div>
+      </div>
+      <div class="llm-stat-budget-text">{{ _('%(percent)s%% of %(budget)s', percent=pct, budget='{:,}'.format(llm_token_budget_month)) }}</div>
+      {% endif %}
+    </div>
+
+    <div class="llm-stat-card">
+      <div class="llm-stat-label">{{ _('All-time total') }}</div>
+      <div class="llm-stat-value">{{ '{:,}'.format(llm_stored.get('tokens_total_cumulative', 0)) }}</div>
+      <div class="llm-stat-sub">{{ _('tokens') }}{% if llm_show_costs and llm_stored.get('cost_usd_total_cumulative') %} &nbsp;·&nbsp; ≈&thinsp;${{ '%.4f'|format(llm_stored.get('cost_usd_total_cumulative', 0)) }}{% endif %}</div>
+    </div>
+  </div>
+
+  {% if llm_token_budget_month and llm_stored.get('tokens_this_month', 0) >= llm_token_budget_month %}
+  <p class="llm-budget-alert">&#9888; {{ _('Monthly token budget reached. AI summarisation is paused until next month.') }}</p>
+  {% endif %}
+
+  <div class="llm-usage-settings">
+    <div class="llm-usage-row">
+      <span class="llm-usage-row-label">{{ _('Token budget this period') }}</span>
+      <span class="llm-usage-row-value">
+        {% if llm_token_budget_month_env %}
+          <strong>{{ '{:,}'.format(llm_token_budget_month_env) }}</strong>
+          <span class="llm-env-badge">{{ _('(set via <code>LLM_TOKEN_BUDGET_MONTH</code>)') | safe }}</span>
+          <input type="hidden" name="llm-token_budget_month" value="{{ llm_token_budget_month_env }}">
+        {% else %}
+          {{ form.llm.form.token_budget_month(placeholder=_('0 = unlimited'), value=llm_stored.get('token_budget_month', 0) or '') }}
+          <span class="llm-field-hint">{{ _('tokens (0 = unlimited)') }}</span>
+        {% endif %}
+      </span>
+    </div>
+    {% if llm_stored.get('tokens_month_key') %}
+    <div class="llm-usage-row">
+      <span class="llm-usage-row-label">{{ _('Current billing period') }}</span>
+      <span class="llm-usage-row-value">{{ llm_stored.get('tokens_month_key') }}</span>
+    </div>
+    {% endif %}
+    <div class="llm-usage-row">
+      <span class="llm-usage-row-label">{{ _('Max input characters') }}</span>
+      <span class="llm-usage-row-value">
+        {% if llm_max_input_chars_env %}
+          {{ form.llm.form.max_input_chars(value=llm_max_input_chars_env, readonly=True, style="width:10em;opacity:0.6;cursor:not-allowed;") }}
+          <span class="llm-env-badge">{{ _('(set via <code>LLM_MAX_INPUT_CHARS</code>)') | safe }}</span>
+        {% else %}
+          {{ form.llm.form.max_input_chars(placeholder='100000', value=llm_stored.get('max_input_chars', 100000) or '') }}
+          <span class="llm-field-hint">{{ _('characters — currently enforcing: %(limit)s', limit='{:,}'.format(llm_effective_max_input_chars)) }}</span>
+        {% endif %}
+      </span>
+    </div>
+    <div class="llm-usage-row">
+      <span class="llm-usage-row-label">{{ _('Max tokens per watch per period') }}</span>
+      <span class="llm-usage-row-value">
+        {{ form.llm.form.max_tokens_per_count_period(placeholder=_('0 = unlimited'), value=llm_stored.get('max_tokens_per_count_period', 0) or '') }}
+        <span class="llm-field-hint">{{ _('tokens — skips AI evaluation on a watch once its usage within the current period (monthly) hits this cap (0 = unlimited)') }}</span>
+      </span>
+    </div>
+  </div>
+
+  {% else %}
+  <p class="llm-no-usage">{{ _('No AI usage recorded yet.') }}</p>
+
+  <div class="llm-usage-settings">
+    <div class="llm-usage-row">
+      <span class="llm-usage-row-label">{{ _('Token budget') }}</span>
+      <span class="llm-usage-row-value">
+        {% if llm_token_budget_month_env %}
+          <strong>{{ '{:,}'.format(llm_token_budget_month_env) }}</strong>
+          <span class="llm-env-badge">{{ _('(set via <code>LLM_TOKEN_BUDGET_MONTH</code>)') | safe }}</span>
+          <input type="hidden" name="llm-token_budget_month" value="{{ llm_token_budget_month_env }}">
+        {% else %}
+          {{ form.llm.form.token_budget_month(placeholder=_('0 = unlimited'), value=llm_stored.get('token_budget_month', 0) or '') }}
+          <span class="llm-field-hint">{{ _('tokens per month (0 = unlimited)') }}</span>
+        {% endif %}
+      </span>
+    </div>
+    <div class="llm-usage-row">
+      <span class="llm-usage-row-label">{{ _('Max input characters') }}</span>
+      <span class="llm-usage-row-value">
+        {% if llm_max_input_chars_env %}
+          {{ form.llm.form.max_input_chars(value=llm_max_input_chars_env, readonly=True, style="width:10em;opacity:0.6;cursor:not-allowed;") }}
+          <span class="llm-env-badge">{{ _('(set via <code>LLM_MAX_INPUT_CHARS</code>)') | safe }}</span>
+        {% else %}
+          {{ form.llm.form.max_input_chars(placeholder='100000', value=llm_stored.get('max_input_chars', 100000) or '') }}
+          <span class="llm-field-hint">{{ _('characters — currently enforcing: %(limit)s', limit='{:,}'.format(llm_effective_max_input_chars)) }}</span>
+        {% endif %}
+      </span>
+    </div>
+    <div class="llm-usage-row">
+      <span class="llm-usage-row-label">{{ _('Max tokens per watch per period') }}</span>
+      <span class="llm-usage-row-value">
+        {{ form.llm.form.max_tokens_per_count_period(placeholder=_('0 = unlimited'), value=llm_stored.get('max_tokens_per_count_period', 0) or '') }}
+        <span class="llm-field-hint">{{ _('tokens — skips AI evaluation on a watch once its usage within the current period (monthly) hits this cap (0 = unlimited)') }}</span>
+      </span>
+    </div>
+  </div>
+  {% endif %}
+  {% endcall %}
+
+{% endcall %}{# stab_shell #}
+</div>
+
+<script>
+(function () {
+  const LIVE_PROVIDERS = ['openai', 'anthropic', 'gemini', 'ollama', 'openai_compatible', 'openrouter'];
+  const BASE_DEFAULTS  = { ollama: 'http://localhost:11434' };
+  const KEY_HINTS = {
+    openai:             '{{ _("platform.openai.com → API keys") }}',
+    anthropic:          '{{ _("console.anthropic.com → API keys") }}',
+    gemini:             '{{ _("aistudio.google.com → Get API key") }}',
+    ollama:             '{{ _("No API key needed for local Ollama") }}',
+    openai_compatible:  '{{ _("Bearer token for your self-hosted server (vLLM, LM Studio, etc.)") }}',
+    openrouter:         '{{ _("openrouter.ai → Keys") }}',
+  };
+
+  window.llmDisclaimerToggle = function (cb) {
+    const fields = document.getElementById('llm-provider-fields');
+    if (fields) fields.style.display = cb.checked ? '' : 'none';
+  };
+
+  window.llmOnProviderChange = function (provider) {
+    const fetchGroup    = document.getElementById('llm-fetch-group');
+    const baseGroup     = document.getElementById('llm-base-group');
+    const modelSelGrp   = document.getElementById('llm-model-select-group');
+    const localAdvGrp   = document.getElementById('llm-local-advanced-group');
+    const baseField     = document.querySelector('[name="llm-api_base"]');
+    const kindField     = document.querySelector('[name="llm-provider_kind"]');
+    const hint          = document.getElementById('llm-key-hint');
+
+    fetchGroup.style.display = LIVE_PROVIDERS.includes(provider) ? '' : 'none';
+
+    const needsBase = provider === 'ollama' || provider === 'openai_compatible';
+    baseGroup.style.display = needsBase ? '' : 'none';
+    if (BASE_DEFAULTS[provider] !== undefined) {
+      if (!baseField.value) baseField.value = BASE_DEFAULTS[provider];
+    }
+
+    // Persist the dropdown selection so the backend can branch on provider kind
+    // (self-hosted endpoints — 'ollama' and 'openai_compatible' — trigger the
+    // local-multiplier code path; cloud providers do not).
+    if (kindField) kindField.value = provider || '';
+
+    // Show the local-endpoint advanced settings (token multiplier) for self-hosted
+    // endpoints. Cloud providers get the original tight caps and don't see this
+    // section at all.
+    if (localAdvGrp) localAdvGrp.style.display = (provider === 'ollama' || provider === 'openai_compatible') ? '' : 'none';
+
+    hint.textContent = KEY_HINTS[provider] || '';
+    modelSelGrp.style.display = 'none';
+    document.getElementById('llm-fetch-status').textContent = '';
+  };
+
+  window.llmFetchModels = async function () {
+    const provider  = document.getElementById('llm-provider').value;
+    const apiKey    = document.querySelector('[name="llm-api_key"]').value.trim();
+    const apiBase   = document.querySelector('[name="llm-api_base"]').value.trim();
+    const btn       = document.getElementById('llm-fetch-btn');
+    const statusEl  = document.getElementById('llm-fetch-status');
+    const selGroup  = document.getElementById('llm-model-select-group');
+    const modelSel  = document.getElementById('llm-model-select');
+
+    if (!provider) { statusEl.textContent = '{{ _("Select a provider first.") }}'; return; }
+
+    btn.disabled = true;
+    btn.textContent = '⏳ {{ _("Loading…") }}';
+    statusEl.textContent = '';
+
+    const params = new URLSearchParams({ provider });
+    if (apiKey)  params.set('api_key',  apiKey);
+    if (apiBase) params.set('api_base', apiBase);
+
+    try {
+      const resp = await fetch('{{ url_for("settings.llm.llm_get_models") }}?' + params);
+      const data = await resp.json();
+
+      if (data.error) {
+        statusEl.style.color = '#c0392b';
+        statusEl.textContent = '✗ ' + data.error;
+        selGroup.style.display = 'none';
+        return;
+      }
+
+      if (!data.models || data.models.length === 0) {
+        statusEl.style.color = '#e67e22';
+        statusEl.textContent = '{{ _("No models returned by the provider.") }}';
+        selGroup.style.display = 'none';
+        return;
+      }
+
+      modelSel.innerHTML = '<option value="">{{ _("— choose a model —") }}</option>';
+      const currentModel = document.querySelector('[name="llm-model"]').value.trim();
+      for (const m of data.models) {
+        const opt = document.createElement('option');
+        opt.value = m;
+        opt.textContent = m;
+        if (m === currentModel) opt.selected = true;
+        modelSel.appendChild(opt);
+      }
+
+      selGroup.style.display = '';
+      statusEl.style.color = '#27ae60';
+      statusEl.textContent = '✓ ' + data.models.length + ' {{ _("models available with your key") }}';
+    } catch (e) {
+      statusEl.style.color = '#c0392b';
+      statusEl.textContent = '✗ {{ _("Request failed") }}: ' + e.message;
+    } finally {
+      btn.disabled = false;
+      btn.textContent = '↻ {{ _("Load available models") }}';
+    }
+  };
+
+  window.llmOnModelPick = function (value) {
+    if (value) document.querySelector('[name="llm-model"]').value = value;
+  };
+
+  window.llmRunTest = async function () {
+    const btn    = document.getElementById('llm-test-btn');
+    const result = document.getElementById('llm-test-result');
+    if (!btn || !result) return;
+
+    btn.disabled = true;
+    btn.textContent = '⏳ {{ _("Testing…") }}';
+    result.style.display = 'none';
+
+    // Send the form's current values so the user doesn't have to hit Save before
+    // testing a config change. Endpoint falls back to the stored datastore values
+    // for any field we don't send.
+    const params = new URLSearchParams();
+    const model   = (document.querySelector('[name="llm-model"]')    || {}).value || '';
+    const apiKey  = (document.querySelector('[name="llm-api_key"]')  || {}).value || '';
+    const apiBase = (document.querySelector('[name="llm-api_base"]') || {}).value || '';
+    const kind    = (document.querySelector('[name="llm-provider_kind"]') || {}).value || '';
+    const mult    = (document.querySelector('[name="llm-local_token_multiplier"]') || {}).value || '';
+    if (model.trim())   params.set('model',    model.trim());
+    if (apiKey.trim())  params.set('api_key',  apiKey.trim());
+    if (apiBase.trim()) params.set('api_base', apiBase.trim());
+    if (kind.trim())    params.set('provider_kind', kind.trim());
+    if (mult.trim())    params.set('local_token_multiplier', mult.trim());
+
+    try {
+      const resp = await fetch('{{ url_for("settings.llm.llm_test") }}?' + params);
+      const data = await resp.json();
+      if (data.ok) {
+        result.style.cssText = 'display:block; background:rgba(39,174,96,0.08); border:1px solid rgba(39,174,96,0.3); border-radius:5px; padding:0.6em 0.85em; font-size:0.88em; line-height:1.45;';
+        result.innerHTML = '<span style="color:#27ae60; font-weight:600;">&#10003; {{ _("Connected") }}</span>'
+          + (data.tokens ? ' <span style="opacity:0.55; font-size:0.9em;">(' + data.tokens + ' {{ _("tokens") }})</span>' : '')
+          + '<br><em style="opacity:0.75;">' + data.text.replace(/</g,'&lt;') + '</em>';
+      } else {
+        result.style.cssText = 'display:block; background:rgba(192,57,43,0.07); border:1px solid rgba(192,57,43,0.25); border-radius:5px; padding:0.6em 0.85em; font-size:0.88em; line-height:1.45;';
+        result.innerHTML = '<span style="color:#c0392b; font-weight:600;">&#10007; {{ _("Failed") }}</span><br><code style="font-size:0.92em; word-break:break-all;">' + (data.error || '').replace(/</g,'&lt;') + '</code>';
+      }
+    } catch (e) {
+      result.style.cssText = 'display:block; background:rgba(192,57,43,0.07); border:1px solid rgba(192,57,43,0.25); border-radius:5px; padding:0.6em 0.85em; font-size:0.88em;';
+      result.innerHTML = '<span style="color:#c0392b; font-weight:600;">&#10007; {{ _("Request failed") }}</span>: ' + e.message.replace(/</g,'&lt;');
+    } finally {
+      btn.disabled = false;
+      btn.textContent = '▶ {{ _("Test connection") }}';
+    }
+  };
+
+  // On page load: detect and pre-select provider from current model
+  (function detectCurrentProvider() {
+    const modelField = document.querySelector('[name="llm-model"]');
+    if (!modelField) return;
+    const m = modelField.value.trim();
+    if (!m) return;
+
+    let guessed = '';
+    if (m.startsWith('gemini/'))       guessed = 'gemini';
+    else if (m.startsWith('ollama/'))  guessed = 'ollama';
+    else if (m.startsWith('openrouter/')) guessed = 'openrouter';
+    else if (m.startsWith('openai/')) {
+      // openai/<model> + custom api_base = self-hosted OpenAI-compatible (vLLM etc.)
+      const baseField = document.querySelector('[name="llm-api_base"]');
+      guessed = (baseField && baseField.value.trim()) ? 'openai_compatible' : 'openai';
+    }
+    else if (m.startsWith('claude'))   guessed = 'anthropic';
+    else if (m.startsWith('gpt') || m.startsWith('o1') || m.startsWith('o3')) guessed = 'openai';
+
+    if (guessed) {
+      const sel = document.getElementById('llm-provider');
+      if (sel) { sel.value = guessed; llmOnProviderChange(guessed); }
+    }
+  })();
+}());
+</script>
@@ -5,6 +5,7 @@ from loguru import logger

 from changedetectionio.store import ChangeDetectionStore
 from changedetectionio.flask_app import login_optionally_required
+from changedetectionio.llm.evaluator import get_llm_config as _get_llm_config


 def construct_blueprint(datastore: ChangeDetectionStore):
@@ -22,11 +23,14 @@ def construct_blueprint(datastore: ChangeDetectionStore):

        tag_count = Counter(tag for watch in datastore.data['watching'].values() if watch.get('tags') for tag in watch['tags'])

+        from changedetectionio import processors
        output = render_template("groups-overview.html",
                                 app_rss_token=datastore.data['settings']['application'].get('rss_access_token'),
                                 available_tags=sorted_tags,
                                 form=add_form,
+                                 generate_tag_colors=processors.generate_processor_badge_colors,
                                 tag_count=tag_count,
+                                 wcag_text_color=processors.wcag_text_color,
                                 )

        return output
@@ -54,7 +58,7 @@ def construct_blueprint(datastore: ChangeDetectionStore):

        return redirect(url_for('tags.tags_overview_page'))

-    @tags_blueprint.route("/mute/<string:uuid>", methods=['GET'])
+    @tags_blueprint.route("/mute/<uuid_str:uuid>", methods=['GET'])
    @login_optionally_required
    def mute(uuid):
        tag = datastore.data['settings']['application']['tags'].get(uuid)
@@ -63,24 +67,13 @@ def construct_blueprint(datastore: ChangeDetectionStore):
            tag.commit()
        return redirect(url_for('tags.tags_overview_page'))

-    @tags_blueprint.route("/delete/<string:uuid>", methods=['GET'])
+    @tags_blueprint.route("/delete/<uuid_str:uuid>", methods=['GET'])
    @login_optionally_required
    def delete(uuid):
        # Delete the tag from settings immediately
        if datastore.data['settings']['application']['tags'].get(uuid):
            del datastore.data['settings']['application']['tags'][uuid]

-        # Delete tag.json file if it exists
-        import os
-        tag_dir = os.path.join(datastore.datastore_path, uuid)
-        tag_json = os.path.join(tag_dir, "tag.json")
-        if os.path.exists(tag_json):
-            try:
-                os.unlink(tag_json)
-                logger.info(f"Deleted tag.json for tag {uuid}")
-            except Exception as e:
-                logger.error(f"Failed to delete tag.json for tag {uuid}: {e}")
-
        # Remove tag from all watches in background thread to avoid blocking
        def remove_tag_background(tag_uuid):
            """Background thread to remove tag from watches - discarded after completion."""
@@ -101,7 +94,7 @@ def construct_blueprint(datastore: ChangeDetectionStore):
        flash(gettext("Tag deleted, removing from watches in background"))
        return redirect(url_for('tags.tags_overview_page'))

-    @tags_blueprint.route("/unlink/<string:uuid>", methods=['GET'])
+    @tags_blueprint.route("/unlink/<uuid_str:uuid>", methods=['GET'])
    @login_optionally_required
    def unlink(uuid):
        # Unlink tag from all watches in background thread to avoid blocking
@@ -127,19 +120,11 @@ def construct_blueprint(datastore: ChangeDetectionStore):
    @tags_blueprint.route("/delete_all", methods=['GET'])
    @login_optionally_required
    def delete_all():
-        # Delete all tag.json files
-        import os
-        for tag_uuid in list(datastore.data['settings']['application']['tags'].keys()):
-            tag_dir = os.path.join(datastore.datastore_path, tag_uuid)
-            tag_json = os.path.join(tag_dir, "tag.json")
-            if os.path.exists(tag_json):
-                try:
-                    os.unlink(tag_json)
-                except Exception as e:
-                    logger.error(f"Failed to delete tag.json for tag {tag_uuid}: {e}")

-        # Clear all tags from settings immediately
-        datastore.data['settings']['application']['tags'] = {}
+        for tag_uuid in list(datastore.data['settings']['application']['tags'].keys()):
+# TagsDict 'del' handler will remove the dir
+            del datastore.data['settings']['application']['tags'][tag_uuid]
+

        # Clear tags from all watches in background thread to avoid blocking
        def clear_all_tags_background():
@@ -160,7 +145,7 @@ def construct_blueprint(datastore: ChangeDetectionStore):
        flash(gettext("All tags deleted, clearing from watches in background"))
        return redirect(url_for('tags.tags_overview_page'))

-    @tags_blueprint.route("/edit/<string:uuid>", methods=['GET'])
+    @tags_blueprint.route("/edit/<uuid_str:uuid>", methods=['GET'])
    @login_optionally_required
    def form_tag_edit(uuid):
        from changedetectionio.blueprint.tags.form import group_restock_settings_form
@@ -179,11 +164,27 @@ def construct_blueprint(datastore: ChangeDetectionStore):
                                       default_system_settings = datastore.data['settings'],
                                       )

+        # Bridge API-stored processor_config_* values into the form's FormField sub-forms.
+        # The API stores processor_config_restock_diff in the tag dict; find the matching
+        # FormField by checking which one's sub-fields cover the config keys.
+        from wtforms.fields.form import FormField as WTFormField
+        for key, value in default.items():
+            if not key.startswith('processor_config_') or not isinstance(value, dict):
+                continue
+            for form_field in form:
+                if isinstance(form_field, WTFormField) and all(k in form_field.form._fields for k in value):
+                    for sub_key, sub_value in value.items():
+                        sub_field = form_field.form._fields.get(sub_key)
+                        if sub_field is not None:
+                            sub_field.data = sub_value
+                    break
+
        template_args = {
            'data': default,
            'form': form,
            'watch': default,
            'extra_notification_token_placeholder_info': datastore.get_unique_notification_token_placeholders_available(),
+            'llm_configured': bool(_get_llm_config(datastore)),
        }

        included_content = {}
@@ -212,9 +213,17 @@ def construct_blueprint(datastore: ChangeDetectionStore):
            template = env.from_string(template_str)
            included_content = template.render(**template_args)

+        # Watches whose URL currently matches this tag's pattern
+        matching_watches = {
+            w_uuid: watch
+            for w_uuid, watch in datastore.data['watching'].items()
+            if default.matches_url(watch.get('url', ''))
+        }
+
        output = render_template("edit-tag.html",
                                 extra_form_content=included_content,
                                 extra_tab_content=form.extra_tab_content() if form.extra_tab_content() else None,
+                                 matching_watches=matching_watches,
                                 settings_application=datastore.data['settings']['application'],
                                 **template_args
                                 )
@@ -222,7 +231,7 @@ def construct_blueprint(datastore: ChangeDetectionStore):
        return output


-    @tags_blueprint.route("/edit/<string:uuid>", methods=['POST'])
+    @tags_blueprint.route("/edit/<uuid_str:uuid>", methods=['POST'])
    @login_optionally_required
    def form_tag_edit_submit(uuid):
        from changedetectionio.blueprint.tags.form import group_restock_settings_form
@@ -255,7 +264,4 @@ def construct_blueprint(datastore: ChangeDetectionStore):
        return redirect(url_for('tags.tags_overview_page'))


-    @tags_blueprint.route("/delete/<string:uuid>", methods=['GET'])
-    def form_tag_delete(uuid):
-        return redirect(url_for('tags.tags_overview_page'))
    return tags_blueprint
@@ -2,20 +2,31 @@ from wtforms import (
    Form,
    StringField,
    SubmitField,
+    TextAreaField,
    validators,
 )
 from wtforms.fields.simple import BooleanField
+from flask_babel import lazy_gettext as _l

 from changedetectionio.processors.restock_diff.forms import processor_settings_form as restock_settings_form
+from changedetectionio.llm.ui_strings import LLM_INTENT_TAG_PLACEHOLDER
+from changedetectionio.llm.evaluator import DEFAULT_CHANGE_SUMMARY_PROMPT

 class group_restock_settings_form(restock_settings_form):
-    overrides_watch = BooleanField('Activate for individual watches in this tag/group?', default=False)
+    overrides_watch = BooleanField(_l('Activate for individual watches in this tag/group?'), default=False)
+    url_match_pattern = StringField(_l('Auto-apply to watches with URLs matching'),
+                                    render_kw={"placeholder": _l("e.g. *://example.com/* or github.com/myorg")})
+    tag_colour = StringField(_l('Tag colour'), default='')
+    llm_intent = TextAreaField('AI Change Intent',
+                               validators=[validators.Optional(), validators.Length(max=2000)],
+                               render_kw={"rows": "5", "placeholder": LLM_INTENT_TAG_PLACEHOLDER})
+
+    llm_change_summary = TextAreaField('AI Change Summary',
+                               validators=[validators.Optional(), validators.Length(max=2000)],
+                               render_kw={"rows": "5", "placeholder": DEFAULT_CHANGE_SUMMARY_PROMPT},
+                               default='')

 class SingleTag(Form):

-    name = StringField('Tag name', [validators.InputRequired()], render_kw={"placeholder": "Name"})
-    save_button = SubmitField('Save', render_kw={"class": "pure-button pure-button-primary"})
-
-
-
-
+    name = StringField(_l('Tag name'), [validators.InputRequired()], render_kw={"placeholder": _l("Name")})
+    save_button = SubmitField(_l('Save'), render_kw={"class": "pure-button pure-button-primary"})
@@ -17,6 +17,8 @@

 </script>

+<script src="{{url_for('static_content', group='js', filename='plugins.js')}}" defer></script>
+<script src="{{url_for('static_content', group='js', filename='global-settings.js')}}" defer></script>
 <script src="{{url_for('static_content', group='js', filename='watch-settings.js')}}" defer></script>
 <script src="{{url_for('static_content', group='js', filename='notifications.js')}}" defer></script>

@@ -25,6 +27,9 @@
    <div class="tabs collapsable">
        <ul>
            <li class="tab" id=""><a href="#general">{{ _('General') }}</a></li>
+            {% if llm_configured %}
+            <li class="tab"><a href="#ai-llm">{{ _('AI / LLM') }}</a></li>
+            {% endif %}
            <li class="tab"><a href="#filters-and-triggers">{{ _('Filters & Triggers') }}</a></li>
            {% if extra_tab_content %}
            <li class="tab"><a href="#extras_tab">{{ extra_tab_content }}</a></li>
@@ -43,11 +48,59 @@
                    <div class="pure-control-group">
                        {{ render_field(form.title, placeholder="https://...", required=true, class="m-d") }}
                    </div>
+                    <div class="pure-control-group">
+                        {{ render_field(form.url_match_pattern, class="m-d") }}
+                        <span class="pure-form-message-inline">{{ _('Automatically applies this tag to any watch whose URL matches. Supports wildcards: <code>*example.com*</code> or plain substring: <code>github.com/myorg</code>')|safe }}</span>
+                    </div>
+                    {% if matching_watches %}
+                    <div class="pure-control-group">
+                        <label>{{ _('Currently matching watches') }} ({{ matching_watches|length }})</label>
+                        <ul class="tag-url-match-list">
+                            {% for w_uuid, w in matching_watches.items() %}
+                            <li><a href="{{ url_for('ui.ui_edit.edit_page', uuid=w_uuid) }}">{{ w.label }}</a></li>
+                            {% endfor %}
+                        </ul>
+                    </div>
+                    {% endif %}
+                    <div class="pure-control-group">
+                        <label>{{ _('Tag colour') }}</label>
+                        <div style="display:flex; align-items:center; gap:0.75em;">
+                            <input type="checkbox" id="use_custom_colour"
+                                   {% if data.get('tag_colour') %}checked{% endif %}>
+                            <label for="use_custom_colour" style="margin:0">{{ _('Custom colour') }}</label>
+                            <input type="color" id="tag_colour_picker"
+                                   value="{{ data.get('tag_colour') or '#4f8ef7' }}"
+                                   {% if not data.get('tag_colour') %}disabled{% endif %}>
+                            <input type="hidden" name="tag_colour" id="tag_colour_hidden"
+                                   value="{{ data.get('tag_colour', '') }}">
+                        </div>
+                        <span class="pure-form-message-inline">{{ _('Leave unchecked to use the auto-generated colour based on the tag name.') }}</span>
+                    </div>
+                    <script>
+                    (function () {
+                        var cb = document.getElementById('use_custom_colour');
+                        var picker = document.getElementById('tag_colour_picker');
+                        var hidden = document.getElementById('tag_colour_hidden');
+                        picker.addEventListener('input', function () { hidden.value = this.value; });
+                        cb.addEventListener('change', function () {
+                            picker.disabled = !this.checked;
+                            hidden.value = this.checked ? picker.value : '';
+                        });
+                    })();
+                    </script>
                </fieldset>
            </div>

+            {% if llm_configured %}
+            <div class="tab-pane-inner" id="ai-llm">
+                {% include "edit/include_llm_intent.html" %}
+            </div>
+            {% endif %}
+
            <div class="tab-pane-inner" id="filters-and-triggers">
-                <p>{{ _('These settings are') }} <strong><i>{{ _('added') }}</i></strong> {{ _('to any existing watch configurations.') }}</p>
+                {# TRANSLATORS: CJK fonts lack native italics; allow substitution with conventional local styling. dennis-ignore: W303 #}
+                <p>{{ _('These settings are <strong><i>added</i></strong> to any existing watch configurations.')|safe }}</p>
+
                {% include "edit/include_subtract.html" %}
                <div class="text-filtering border-fieldset">
                    <h3>{{ _('Text filtering') }}</h3>
@@ -78,7 +131,7 @@
                        {% if has_default_notification_urls %}
                        <div class="inline-warning">
                            <img class="inline-warning-icon" src="{{url_for('static_content', group='images', filename='notice.svg')}}" alt="{{ _('Look out!') }}" title="{{ _('Lookout!') }}" >
-                            {{ _('There are') }} <a href="{{ url_for('settings.settings_page')}}#notifications">{{ _('system-wide notification URLs enabled') }}</a>, {{ _('this form will override notification settings for this watch only') }} &dash; {{ _('an empty Notification URL list here will still send notifications.') }}
+                            {{ _('There are <a href="%(url)s">system-wide notification URLs enabled</a>, this form will override notification settings for this watch only &dash; an empty Notification URL list here will still send notifications.', url=url_for('settings.settings_page') ~ '#notifications')|safe }}
                        </div>
                        {% endif %}
                        <a href="#notifications" id="notification-setting-reset-to-default" class="pure-button button-xsmall" style="right: 20px; top: 20px; position: absolute; background-color: #5f42dd; border-radius: 4px; font-size: 70%; color: #fff">{{ _('Use system defaults') }}</a>
@@ -3,6 +3,26 @@
 {% from '_helpers.html' import render_simple_field, render_field %}
 <script src="{{url_for('static_content', group='js', filename='jquery-3.6.0.min.js')}}"></script>
 <script src="{{url_for('static_content', group='js', filename='modal.js')}}"></script>
+<style>
+{%- for uuid, tag in available_tags -%}
+{%- if tag and tag.title -%}
+{%- set class_name = tag.title|sanitize_tag_class -%}
+{%- if tag.get('tag_colour') -%}
+.watch-tag-list.tag-{{ class_name }} { background-color: {{ tag.tag_colour }}; color: {{ wcag_text_color(tag.tag_colour) }}; }
+{%- else -%}
+{%- set colors = generate_tag_colors(tag.title) -%}
+.watch-tag-list.tag-{{ class_name }} {
+  background-color: {{ colors['light']['bg'] }};
+  color: {{ colors['light']['color'] }};
+}
+html[data-darkmode="true"] .watch-tag-list.tag-{{ class_name }} {
+  background-color: {{ colors['dark']['bg'] }};
+  color: {{ colors['dark']['color'] }};
+}
+{%- endif -%}
+{%- endif -%}
+{%- endfor -%}
+</style>

 <div class="box">
    <form class="pure-form" action="{{ url_for('tags.form_tag_add') }}" method="POST" id="new-watch-form">
@@ -45,10 +65,10 @@
            {% for uuid, tag in available_tags  %}
            <tr id="{{ uuid }}" class="{{ loop.cycle('pure-table-odd', 'pure-table-even') }}">
                <td class="watch-controls">
-                    <a class="link-mute state-{{'on' if tag.notification_muted else 'off'}}" href="{{url_for('tags.mute', uuid=tag.uuid)}}"><img src="{{url_for('static_content', group='images', filename='bell-off.svg')}}" alt="Mute notifications" title="Mute notifications" class="icon icon-mute" ></a>
+                    <a class="link-mute state-{{'on' if tag.notification_muted else 'off'}}" href="{{url_for('tags.mute', uuid=tag.uuid)}}"><img src="{{url_for('static_content', group='images', filename='bell-off.svg')}}" alt="{{ _('Mute notifications') }}" title="{{ _('Mute notifications') }}" class="icon icon-mute" ></a>
                </td>
                <td>{{ "{:,}".format(tag_count[uuid]) if uuid in tag_count else 0 }}</td>
-                <td class="title-col inline"> <a href="{{url_for('watchlist.index', tag=uuid) }}">{{ tag.title }}</a></td>
+                <td class="title-col inline"> <a href="{{url_for('watchlist.index', tag=uuid) }}" class="watch-tag-list tag-{{ tag.title|sanitize_tag_class }}">{{ tag.title }}</a></td>
                <td>
                    <a class="pure-button pure-button-primary" href="{{ url_for('tags.form_tag_edit', uuid=uuid) }}">{{ _('Edit') }}</a>
                    <a href="{{ url_for('ui.form_watch_checknow', tag=uuid) }}" class="pure-button pure-button-primary" >{{ _('Recheck') }}</a>
@@ -141,7 +141,7 @@ def construct_blueprint(datastore: ChangeDetectionStore, update_q, worker_pool,
    # Import the login decorator
    from changedetectionio.auth_decorator import login_optionally_required

-    @ui_blueprint.route("/clear_history/<string:uuid>", methods=['GET'])
+    @ui_blueprint.route("/clear_history/<uuid_str:uuid>", methods=['GET'])
    @login_optionally_required
    def clear_watch_history(uuid):
        try:
@@ -156,9 +156,9 @@ def construct_blueprint(datastore: ChangeDetectionStore, update_q, worker_pool,
    @login_optionally_required
    def clear_all_history():
        if request.method == 'POST':
-            confirmtext = request.form.get('confirmtext')
+            confirmtext = request.form.get('confirmtext', '')

-            if confirmtext == 'clear':
+            if confirmtext.strip().lower() == gettext('clear').strip().lower():
                # Run in background thread to avoid blocking
                def clear_history_background():
                    # Capture UUIDs first to avoid race conditions
@@ -194,9 +194,9 @@ def construct_blueprint(datastore: ChangeDetectionStore, update_q, worker_pool,
        tag_limit = request.args.get('tag')
        now = int(time.time())

-        # Mark watches as viewed in background thread to avoid blocking
-        def mark_viewed_background():
-            """Background thread to mark watches as viewed - discarded after completion."""
+        # Mark watches as viewed - use background thread only for large watch counts
+        def mark_viewed_impl():
+            """Mark watches as viewed - can run synchronously or in background thread."""
            marked_count = 0
            try:
                for watch_uuid, watch in datastore.data['watching'].items():
@@ -209,15 +209,21 @@ def construct_blueprint(datastore: ChangeDetectionStore, update_q, worker_pool,
                    datastore.set_last_viewed(watch_uuid, now)
                    marked_count += 1

-                logger.info(f"Background marking complete: {marked_count} watches marked as viewed")
+                logger.info(f"Marking complete: {marked_count} watches marked as viewed")
            except Exception as e:
-                logger.error(f"Error in background mark as viewed: {e}")
+                logger.error(f"Error marking as viewed: {e}")

-        # Start background thread and return immediately
-        thread = threading.Thread(target=mark_viewed_background, daemon=True)
-        thread.start()
+        # For small watch counts (< 10), run synchronously to avoid race conditions in tests
+        # For larger counts, use background thread to avoid blocking the UI
+        watch_count = len(datastore.data['watching'])
+        if watch_count < 10:
+            # Run synchronously for small watch counts
+            mark_viewed_impl()
+        else:
+            # Start background thread for large watch counts
+            thread = threading.Thread(target=mark_viewed_impl, daemon=True)
+            thread.start()

-        flash(gettext("Marking watches as viewed in background..."))
        return redirect(url_for('watchlist.index', tag=tag_limit))

    @ui_blueprint.route("/delete", methods=['GET'])
@@ -301,8 +307,8 @@ def construct_blueprint(datastore: ChangeDetectionStore, update_q, worker_pool,
                # Provide feedback about skipped watches
                skipped_count = len(watches_to_queue) - len(watches_to_queue_filtered)
                if skipped_count > 0:
-                    flash(gettext("Queued {} watches for rechecking ({} already queued or running).").format(
-                        len(watches_to_queue_filtered), skipped_count))
+                    flash(gettext("Queued {count} watches for rechecking ({skipped_count} already queued or running).").format(
+                        count=len(watches_to_queue_filtered), skipped_count=skipped_count))
                else:
                    if len(watches_to_queue_filtered) == 1:
                        flash(gettext("Queued 1 watch for rechecking."))
@@ -360,7 +366,7 @@ def construct_blueprint(datastore: ChangeDetectionStore, update_q, worker_pool,
        return redirect(url_for('watchlist.index'))


-    @ui_blueprint.route("/share-url/<string:uuid>", methods=['GET'])
+    @ui_blueprint.route("/share-url/<uuid_str:uuid>", methods=['GET'])
    @login_optionally_required
    def form_share_put_watch(uuid):
        """Given a watch UUID, upload the info and return a share-link
@@ -17,6 +17,34 @@ from changedetectionio.store import ChangeDetectionStore
 from changedetectionio.auth_decorator import login_optionally_required


+def _clean_litellm_error(exc) -> str:
+    """Return a short, human-readable error string from a litellm exception.
+
+    litellm embeds the raw provider JSON in str(exc), which can be hundreds of
+    characters of verbose quota detail.  We try to pull just the provider's
+    'message' field; failing that we return the first non-empty line with the
+    'litellm.XxxError:' class prefix stripped.
+    """
+    import json, re
+    raw = str(exc)
+    # Try to parse the embedded JSON block (starts at first '{')
+    brace = raw.find('{')
+    if brace >= 0:
+        try:
+            payload = json.loads(raw[brace:])
+            msg = (payload.get('error') or {}).get('message') or ''
+            if msg:
+                # Take only the first sentence / line — provider messages can be long
+                return msg.split('\n')[0].split('. ')[0].strip() + '.'
+        except Exception:
+            pass
+    # Fallback: strip the "litellm.XxxError: litellm.XxxError: providerException - " prefix
+    first_line = raw.split('\n')[0]
+    first_line = re.sub(r'^(litellm\.\w+:\s*)+', '', first_line)
+    first_line = re.sub(r'\w+Exception\s*-\s*', '', first_line).strip()
+    return first_line or raw.split('\n')[0]
+
+
 def construct_blueprint(datastore: ChangeDetectionStore):
    diff_blueprint = Blueprint('ui_diff', __name__, template_folder="../ui/templates")

@@ -66,7 +94,7 @@ def construct_blueprint(datastore: ChangeDetectionStore):

        return Markup(result)

-    @diff_blueprint.route("/diff/<string:uuid>", methods=['GET'])
+    @diff_blueprint.route("/diff/<uuid_str:uuid>", methods=['GET'])
    @login_optionally_required
    def diff_history_page(uuid):
        """
@@ -128,7 +156,176 @@ def construct_blueprint(datastore: ChangeDetectionStore):
            redirect=redirect
        )

-    @diff_blueprint.route("/diff/<string:uuid>/extract", methods=['GET'])
+    @diff_blueprint.route("/diff/<uuid_str:uuid>/llm-summary/prompt", methods=['GET'])
+    @login_optionally_required
+    def diff_llm_summary_prompt(uuid):
+        """Return the effective LLM summary prompt for a watch immediately (no LLM call)."""
+        from flask import jsonify
+        watch = datastore.data['watching'].get(uuid)
+        if not watch:
+            return jsonify({'prompt': ''}), 404
+        try:
+            from changedetectionio.llm.evaluator import get_effective_summary_prompt
+            prompt = get_effective_summary_prompt(watch, datastore)
+        except Exception:
+            prompt = ''
+        return jsonify({'prompt': prompt})
+
+    @diff_blueprint.route("/diff/<uuid_str:uuid>/llm-summary", methods=['GET'])
+    @login_optionally_required
+    def diff_llm_summary(uuid):
+        """
+        Generate (or return cached) an AI summary of the diff between two snapshots.
+        Called via AJAX from the diff page when no cached summary exists.
+        Returns JSON: {"summary": "...", "error": null} or {"summary": null, "error": "..."}
+        """
+        import difflib
+        from flask import jsonify
+
+        try:
+            watch = datastore.data['watching'][uuid]
+        except KeyError:
+            return jsonify({'summary': None, 'error': 'Watch not found'}), 404
+
+        llm_cfg = datastore.data.get('settings', {}).get('application', {}).get('llm', {})
+        if not llm_cfg.get('model'):
+            return jsonify({'summary': None, 'error': 'LLM not configured'}), 400
+
+        dates = list(watch.history.keys())
+        if len(dates) < 2:
+            return jsonify({'summary': None, 'error': 'Not enough history'}), 400
+
+        best_from = watch.get_from_version_based_on_last_viewed
+        from_version      = request.args.get('from_version', best_from if best_from else dates[-2])
+        to_version        = request.args.get('to_version', dates[-1])
+        from changedetectionio.llm.evaluator import DiffPrefs
+        prefs             = DiffPrefs.from_request_args(request.args)
+        all_changes       = prefs.all_changes
+        ignore_whitespace = prefs.ignore_whitespace
+        show_removed      = prefs.show_removed
+        show_added        = prefs.show_added
+
+        def _prep(text):
+            """Optionally normalise whitespace on each line before diffing."""
+            if not ignore_whitespace:
+                return text.splitlines()
+            return [' '.join(line.split()) for line in text.splitlines()]
+
+        def _make_unified_diff(a_text, b_text):
+            lines = list(difflib.unified_diff(_prep(a_text), _prep(b_text), lineterm='', n=3))
+            return '\n'.join(lines[2:]) if len(lines) > 2 else '\n'.join(lines)
+
+        def _apply_filters(diff_text):
+            """Strip +/- lines the user has hidden in the UI so the LLM matches what they see."""
+            if show_removed and show_added:
+                return diff_text
+            out = []
+            for line in diff_text.splitlines():
+                if line.startswith('-') and not show_removed:
+                    continue
+                if line.startswith('+') and not show_added:
+                    continue
+                out.append(line)
+            return '\n'.join(out)
+
+        try:
+            from_text = watch.get_history_snapshot(timestamp=from_version)
+            to_text = watch.get_history_snapshot(timestamp=to_version)
+        except Exception as e:
+            return jsonify({'summary': None, 'error': f'Could not read snapshots: {e}'}), 500
+
+        if all_changes:
+            # Build sequential diffs for every intermediate snapshot between from and to
+            # so the LLM sees the full timeline of changes, not just start→end
+            sorted_dates = sorted(dates)
+            try:
+                start_idx = sorted_dates.index(from_version)
+                end_idx   = sorted_dates.index(to_version)
+            except ValueError:
+                start_idx, end_idx = 0, len(sorted_dates) - 1
+
+            steps = sorted_dates[start_idx:end_idx + 1]
+            segments = []
+            for i in range(len(steps) - 1):
+                a_ts, b_ts = steps[i], steps[i + 1]
+                try:
+                    a_text = watch.get_history_snapshot(timestamp=a_ts) or ''
+                    b_text = watch.get_history_snapshot(timestamp=b_ts) or ''
+                except Exception:
+                    continue
+                seg = _apply_filters(_make_unified_diff(a_text, b_text))
+                if seg.strip():
+                    segments.append(f'=== {a_ts} → {b_ts} ===\n{seg}')
+
+            diff_text = '\n\n'.join(segments) if segments else ''
+        else:
+            diff_text = _apply_filters(_make_unified_diff(from_text, to_text))
+
+        if not diff_text.strip():
+            return jsonify({'summary': None, 'error': 'No differences found'})
+
+        from changedetectionio.llm.evaluator import (
+            summarise_change, get_effective_summary_prompt, build_summary_cache_prompt,
+            is_global_token_budget_exceeded, get_global_token_budget_month,
+            LLMInputTooLargeError,
+        )
+
+        # Diff-pref flags + system prompt + active model are part of the cache key
+        # so prompt or model changes bust the cache.
+        from changedetectionio.llm.evaluator import get_llm_settings
+        _ls = get_llm_settings(datastore)
+        _max_summary_tokens = _ls.max_summary_tokens
+        _llm_model = _ls.model
+        cache_prompt = build_summary_cache_prompt(
+            effective_prompt=get_effective_summary_prompt(watch, datastore),
+            max_summary_tokens=_max_summary_tokens,
+            prefs=prefs,
+            model=_llm_model,
+        )
+
+        # Check cache — keyed by version pair + prompt hash (invalidates if prompt changes)
+        cached = watch.get_llm_diff_summary(from_version, to_version, prompt=cache_prompt)
+        if cached:
+            import time
+            datastore.set_last_viewed(uuid, int(time.time()))
+            return jsonify({'summary': cached, 'error': None, 'cached': True})
+
+        # Check global monthly token budget before making an LLM call
+        if is_global_token_budget_exceeded(datastore):
+            budget = get_global_token_budget_month(datastore)
+            llm_cfg = datastore.data.get('settings', {}).get('application', {}).get('llm', {})
+            used = llm_cfg.get('tokens_this_month', 0)
+            return jsonify({
+                'summary': None,
+                'error': gettext(
+                    'Monthly AI token budget of %(budget)s tokens reached (%(used)s used). Resets next month.',
+                    budget=f'{budget:,}',
+                    used=f'{used:,}',
+                ),
+                'budget_exceeded': True,
+            }), 429
+
+        try:
+            summary = summarise_change(watch, datastore, diff=diff_text, current_snapshot=to_text)
+        except LLMInputTooLargeError as e:
+            return jsonify({'summary': None, 'error': str(e)}), 400
+        except Exception as e:
+            logger.error(f"LLM summary generation failed for {uuid}: {e}")
+            return jsonify({'summary': None, 'error': _clean_litellm_error(e)}), 500
+
+        if not summary:
+            return jsonify({'summary': None, 'error': 'LLM returned empty summary'})
+
+        try:
+            watch.save_llm_diff_summary(summary, from_version, to_version, prompt=cache_prompt)
+        except Exception as e:
+            logger.warning(f"Could not cache llm summary for {uuid}: {e}")
+
+        import time
+        datastore.set_last_viewed(uuid, int(time.time()))
+        return jsonify({'summary': summary, 'error': None, 'cached': False})
+
+    @diff_blueprint.route("/diff/<uuid_str:uuid>/extract", methods=['GET'])
    @login_optionally_required
    def diff_history_page_extract_GET(uuid):
        """
@@ -182,7 +379,7 @@ def construct_blueprint(datastore: ChangeDetectionStore):
            redirect=redirect
        )

-    @diff_blueprint.route("/diff/<string:uuid>/extract", methods=['POST'])
+    @diff_blueprint.route("/diff/<uuid_str:uuid>/extract", methods=['POST'])
    @login_optionally_required
    def diff_history_page_extract_POST(uuid):
        """
@@ -238,7 +435,48 @@ def construct_blueprint(datastore: ChangeDetectionStore):
            redirect=redirect
        )

-    @diff_blueprint.route("/diff/<string:uuid>/processor-asset/<string:asset_name>", methods=['GET'])
+    @diff_blueprint.route("/diff/<uuid_str:uuid>/download-patch", methods=['GET'])
+    @login_optionally_required
+    def download_patch(uuid):
+        """
+        Generate and return a unified diff patch file between two snapshots.
+        Query params: from_version, to_version (timestamp strings from watch history).
+        Returns the patch as a downloadable .patch file — the same content fed to the LLM.
+        """
+        import difflib
+
+        try:
+            watch = datastore.data['watching'][uuid]
+        except KeyError:
+            return make_response('Watch not found', 404)
+
+        dates = list(watch.history.keys())
+        if len(dates) < 2:
+            return make_response('Not enough history', 400)
+
+        from_version = request.args.get('from_version', dates[-2])
+        to_version   = request.args.get('to_version',   dates[-1])
+
+        try:
+            from_text = watch.get_history_snapshot(timestamp=from_version)
+            to_text   = watch.get_history_snapshot(timestamp=to_version)
+        except Exception as e:
+            return make_response(f'Could not read snapshots: {e}', 500)
+
+        diff_lines = list(difflib.unified_diff(
+            from_text.splitlines(keepends=True),
+            to_text.splitlines(keepends=True),
+            fromfile=f'snapshot-{from_version}',
+            tofile=f'snapshot-{to_version}',
+            lineterm='',
+        ))
+        patch_text = ''.join(diff_lines) if diff_lines else '(no differences)\n'
+
+        response = make_response(patch_text)
+        response.headers['Content-Type'] = 'text/plain; charset=utf-8'
+        return response
+
+    @diff_blueprint.route("/diff/<uuid_str:uuid>/processor-asset/<string:asset_name>", methods=['GET'])
    @login_optionally_required
    def processor_asset(uuid, asset_name):
        """
@@ -10,17 +10,39 @@ from changedetectionio.store import ChangeDetectionStore
 from changedetectionio.auth_decorator import login_optionally_required
 from changedetectionio.time_handler import is_within_schedule
 from changedetectionio import worker_pool
+from changedetectionio.llm.evaluator import get_llm_config as _get_llm_config

 def construct_blueprint(datastore: ChangeDetectionStore, update_q, queuedWatchMetaData):
    edit_blueprint = Blueprint('ui_edit', __name__, template_folder="../ui/templates")
-    
+
+    def _resolve_llm_group_overrides(watch, datastore) -> dict:
+        """
+        For each LLM field (llm_intent, llm_change_summary): if the watch has no own
+        value but a linked tag does, return {'value': ..., 'group_name': ...} so the
+        edit template can render the textarea as readonly with a group-sourced placeholder.
+        Returns None for each field when the watch has its own value (editable).
+        """
+        result = {'llm_intent': None, 'llm_change_summary': None}
+        for field in ('llm_intent', 'llm_change_summary'):
+            if (watch.get(field) or '').strip():
+                continue  # watch has its own value — editable, no group override
+            for tag_uuid in watch.get('tags', []):
+                tag = datastore.data['settings']['application'].get('tags', {}).get(tag_uuid)
+                if tag and (tag.get(field) or '').strip():
+                    result[field] = {
+                        'value': tag.get(field).strip(),
+                        'group_name': tag.get('title', 'tag'),
+                    }
+                    break
+        return result
+
    def _watch_has_tag_options_set(watch):
        """This should be fixed better so that Tag is some proper Model, a tag is just a Watch also"""
        for tag_uuid, tag in datastore.data['settings']['application'].get('tags', {}).items():
            if tag_uuid in watch.get('tags', []) and (tag.get('include_filters') or tag.get('subtractive_selectors')):
                return True

-    @edit_blueprint.route("/edit/<string:uuid>", methods=['GET', 'POST'])
+    @edit_blueprint.route("/edit/<uuid_str:uuid>", methods=['GET', 'POST'])
    @login_optionally_required
    # https://stackoverflow.com/questions/42984453/wtforms-populate-form-with-data-if-data-exists
    # https://wtforms.readthedocs.io/en/3.0.x/forms/#wtforms.form.Form.populate_obj ?
@@ -117,19 +139,32 @@ def construct_blueprint(datastore: ChangeDetectionStore, update_q, queuedWatchMe
                processor_config = processor_instance.get_extra_watch_config(config_filename)

                if processor_config:
+                    from wtforms.fields.form import FormField
                    # Populate processor-config-* fields from JSON
                    for config_key, config_value in processor_config.items():
-                        field_name = f'processor_config_{config_key}'
-                        if hasattr(form, field_name):
-                            getattr(form, field_name).data = config_value
-                            logger.debug(f"Loaded processor config from {config_filename}: {field_name} = {config_value}")
+                        if not isinstance(config_value, dict):
+                            continue
+                        # Try exact API-named field first (e.g., processor_config_restock_diff)
+                        target_field = getattr(form, f'processor_config_{config_key}', None)
+                        # Fallback: find any FormField sub-form whose fields cover config_value keys
+                        if target_field is None:
+                            for form_field in form:
+                                if isinstance(form_field, FormField) and all(k in form_field.form._fields for k in config_value):
+                                    target_field = form_field
+                                    break
+                        if target_field is not None:
+                            for sub_key, sub_value in config_value.items():
+                                sub_field = target_field.form._fields.get(sub_key)
+                                if sub_field is not None:
+                                    sub_field.data = sub_value
+                                    logger.debug(f"Loaded processor config from {config_filename}: {sub_key} = {sub_value}")
            except Exception as e:
                logger.warning(f"Failed to load processor config: {e}")

        for p in datastore.extra_browsers:
            form.fetch_backend.choices.append(p)

-        form.fetch_backend.choices.append(("system", 'System settings default'))
+        form.fetch_backend.choices.append(("system", gettext('System settings default')))

        # form.browser_steps[0] can be assumed that we 'goto url' first

@@ -137,7 +172,7 @@ def construct_blueprint(datastore: ChangeDetectionStore, update_q, queuedWatchMe
            # @todo - Couldn't get setattr() etc dynamic addition working, so remove it instead
            del form.proxy
        else:
-            form.proxy.choices = [('', 'Default')]
+            form.proxy.choices = [('', gettext('Default'))]
            for p in datastore.proxy_list:
                form.proxy.choices.append(tuple((p, datastore.proxy_list[p]['label'])))

@@ -288,7 +323,7 @@ def construct_blueprint(datastore: ChangeDetectionStore, update_q, queuedWatchMe
                'extra_classes': ' '.join(c),
                'extra_notification_token_placeholder_info': datastore.get_unique_notification_token_placeholders_available(),
                'extra_processor_config': form.extra_tab_content(),
-                'extra_title': f" - Edit - {watch.label}",
+                'extra_title': f" - {gettext('Edit')} - {watch.label}",
                'form': form,
                'has_default_notification_urls': True if len(datastore.data['settings']['application']['notification_urls']) else False,
                'has_extra_headers_file': len(datastore.get_all_headers_in_textfile_for_watch(uuid=uuid)) > 0,
@@ -307,7 +342,15 @@ def construct_blueprint(datastore: ChangeDetectionStore, update_q, queuedWatchMe
                'using_global_webdriver_wait': not default['webdriver_delay'],
                'uuid': uuid,
                'watch': watch,
-                'capabilities': capabilities
+                'capabilities': capabilities,
+                'auto_applied_tags': {
+                    tag_uuid: tag
+                    for tag_uuid, tag in datastore.data['settings']['application']['tags'].items()
+                    if tag_uuid not in watch.get('tags', []) and tag.matches_url(watch.get('url', ''))
+                },
+                # LLM intent context
+                'llm_configured': bool(_get_llm_config(datastore)),
+                'llm_group_overrides': _resolve_llm_group_overrides(watch, datastore),
            }

            included_content = None
@@ -327,7 +370,7 @@ def construct_blueprint(datastore: ChangeDetectionStore, update_q, queuedWatchMe

        return output

-    @edit_blueprint.route("/edit/<string:uuid>/get-html", methods=['GET'])
+    @edit_blueprint.route("/edit/<uuid_str:uuid>/get-html", methods=['GET'])
    @login_optionally_required
    def watch_get_latest_html(uuid):
        from io import BytesIO
@@ -354,7 +397,7 @@ def construct_blueprint(datastore: ChangeDetectionStore, update_q, queuedWatchMe
        # Return a 500 error
        abort(500)

-    @edit_blueprint.route("/edit/<string:uuid>/get-data-package", methods=['GET'])
+    @edit_blueprint.route("/edit/<uuid_str:uuid>/get-data-package", methods=['GET'])
    @login_optionally_required
    def watch_get_data_package(uuid):
        """Download all data for a single watch as a zip file"""
@@ -405,7 +448,7 @@ def construct_blueprint(datastore: ChangeDetectionStore, update_q, queuedWatchMe
                        mimetype='application/zip')

    # Ajax callback
-    @edit_blueprint.route("/edit/<string:uuid>/preview-rendered", methods=['POST'])
+    @edit_blueprint.route("/edit/<uuid_str:uuid>/preview-rendered", methods=['POST'])
    @login_optionally_required
    def watch_get_preview_rendered(uuid):
        '''For when viewing the "preview" of the rendered text from inside of Edit'''
@@ -10,7 +10,8 @@ from changedetectionio import html_tools
 def construct_blueprint(datastore: ChangeDetectionStore):
    preview_blueprint = Blueprint('ui_preview', __name__, template_folder="../ui/templates")

-    @preview_blueprint.route("/preview/<string:uuid>", methods=['GET'])
+
+    @preview_blueprint.route("/preview/<uuid_str:uuid>", methods=['GET', 'POST'])
    @login_optionally_required
    def preview_page(uuid):
        """
@@ -59,12 +60,8 @@ def construct_blueprint(datastore: ChangeDetectionStore):
        versions = []
        timestamp = None

-        system_uses_webdriver = datastore.data['settings']['application']['fetch_backend'] == 'html_webdriver'
        extra_stylesheets = [url_for('static_content', group='styles', filename='diff.css')]
-
-        is_html_webdriver = False
-        if (watch.get('fetch_backend') == 'system' and system_uses_webdriver) or watch.get('fetch_backend') == 'html_webdriver' or watch.get('fetch_backend', '').startswith('extra_browser_'):
-            is_html_webdriver = True
+        is_html_webdriver = watch.fetcher_supports_screenshots

        triggered_line_numbers = []
        ignored_line_numbers = []
@@ -74,7 +71,9 @@ def construct_blueprint(datastore: ChangeDetectionStore):
            flash(gettext("Preview unavailable - No fetch/check completed or triggers not reached"), "error")
        else:
            # So prepare the latest preview or not
-            preferred_version = request.args.get('version')
+            preferred_version = request.values.get('version') if request.method == 'POST' else request.args.get('version')
+
+
            versions = list(watch.history.keys())
            timestamp = versions[-1]
            if preferred_version and preferred_version in versions:
@@ -108,7 +107,7 @@ def construct_blueprint(datastore: ChangeDetectionStore):
                                 current_diff_url=watch['url'],
                                 current_version=timestamp,
                                 extra_stylesheets=extra_stylesheets,
-                                 extra_title=f" - Diff - {watch.label} @ {timestamp}",
+                                 extra_title=f" - {gettext('Diff')} - {watch.label} @ {timestamp}",
                                 highlight_ignored_line_numbers=ignored_line_numbers,
                                 highlight_triggered_line_numbers=triggered_line_numbers,
                                 highlight_blocked_line_numbers=blocked_line_numbers,
@@ -125,7 +124,7 @@ def construct_blueprint(datastore: ChangeDetectionStore):

        return output

-    @preview_blueprint.route("/preview/<string:uuid>/processor-asset/<string:asset_name>", methods=['GET'])
+    @preview_blueprint.route("/preview/<uuid_str:uuid>/processor-asset/<string:asset_name>", methods=['GET'])
    @login_optionally_required
    def processor_asset(uuid, asset_name):
        """
@@ -25,7 +25,8 @@
        <fieldset class="diff-fieldset">
            {% if versions|length >= 1 %}
                <span style="white-space: nowrap;">
-                <label id="change-from" for="diff-from-version" class="from-to-label">{{ _('From') }}</label>
+                {# TRANSLATORS: 'From' labels the older snapshot version selector on the diff page #}
+                <label id="change-from" for="diff-from-version" class="from-to-label">{{ pgettext('diff version', 'From') }}</label>
                <select id="diff-from-version" name="from_version" class="needs-localtime">
                    {%- for version in versions|reverse -%}
                        <option value="{{ version }}" {% if version== from_version %} selected="" {% endif %}>
@@ -35,7 +36,8 @@
                </select>
                </span>
                <span style="white-space: nowrap;">
-                <label id="change-to" for="diff-to-version" class="from-to-label">{{ _('To') }}</label>
+                {# TRANSLATORS: 'To' labels the newer snapshot version selector on the diff page #}
+                <label id="change-to" for="diff-to-version" class="from-to-label">{{ pgettext('diff version', 'To') }}</label>
                <select id="diff-to-version" name="to_version" class="needs-localtime">
                    {%- for version in versions|reverse -%}
                        <option value="{{ version }}" {% if version== to_version %} selected="" {% endif %}>
@@ -76,6 +78,12 @@
                <label for="replaced" class="pure-checkbox" id="label-diff-replaced">
                <input type="checkbox" id="replaced"  name="replaced" {% if diff_prefs.replaced %}checked=""{% endif %}> {{ _('Replaced') }}</label>
            </span>
+            {%- if llm_configured -%}
+            <span>
+                <label for="llm_all_changes" class="pure-checkbox" id="label-diff-llm-all-changes" title="{{ _('Include all intermediate snapshots between the selected versions in the AI summary') }}">
+                <input type="checkbox" id="llm_all_changes" name="llm_all_changes" {% if diff_prefs.llm_all_changes %}checked=""{% endif %}> &#x2728; {{ _('AI: every change between versions') }}</label>
+            </span>
+            {%- endif -%}
        </fieldset>
        {%- if versions|length >= 2 -%}
            <div id="keyboard-nav">
@@ -124,9 +132,22 @@
        </div>
    {%- endif -%}
        {%- if password_enabled_and_share_is_off -%}
-            <div class="tip">{{ _('Pro-tip: You can enable') }} <strong>{{ _('"share access when password is enabled"') }}</strong> {{ _('from settings.') }}
+            <div class="tip">{{ _('Pro-tip: You can enable <strong>"share access when password is enabled"</strong> from settings.')|safe }}
            </div>
        {%- endif -%}
+        {%- if llm_configured -%}
+        <div id="llm-diff-summary-area"{% if not llm_diff_summary %} data-pending="1"{% endif %}>
+            <span class="llm-diff-summary-label">&#x2728; {{ _('AI Change Summary') }}</span>
+            {%- if llm_diff_summary -%}
+            <p class="llm-diff-summary-text">{{ llm_diff_summary }}</p>
+            {%- else -%}
+            <p class="llm-diff-summary-text llm-diff-summary-loading">{{ _('Generating summary…') }}</p>
+            {%- if llm_summary_prompt -%}
+            <p class="llm-diff-summary-prompt"><span class="llm-diff-summary-prompt-text">{{ llm_summary_prompt }}</span></p>
+            {%- endif -%}
+            {%- endif -%}
+        </div>
+        {%- endif -%}
        <div id="text-diff-heading-area"  style="user-select: none;">
            <div class="snapshot-age"><span>{{ from_version|format_timestamp_timeago }}</span>
                {%- if note -%}<span class="note"><strong>{{ note }}</strong></span>{%- endif -%}
@@ -136,6 +157,7 @@
        <pre id="difference" style="border-left: 2px solid #ddd;">{{ content| diff_unescape_difference_spans }}</pre>
    <div id="diff-visualiser-area-after" style="user-select: none;">
        <strong>{{ _('Tip:') }}</strong> {{ _('Highlight text to share or add to ignore lists.') }}
+        &nbsp;&mdash;&nbsp;<a href="{{ url_for('ui.ui_diff.download_patch', uuid=uuid, from_version=from_version, to_version=to_version) }}" target="_blank" rel="noopener" style="font-size:0.85em;">{{ _('Download difference patch') }}</a>
    </div>
    </div>

@@ -162,5 +184,58 @@
 </script>
 <script src="{{url_for('static_content', group='js', filename='diff-render.js')}}"></script>

+{% if llm_configured %}
+<script>
+$(function () {
+    var $area = $('#llm-diff-summary-area');
+    if (!$area.length || !$area.data('pending')) return;
+
+    var fromVersion = $('#diff-from-version').val();
+    var toVersion   = $('#diff-to-version').val();
+    var summaryUrl  = "{{ url_for('ui.ui_diff.diff_llm_summary', uuid=uuid) }}";
+
+    function showLlmError(msg) {
+        $area.find('.llm-diff-summary-text')
+             .removeClass('llm-diff-summary-loading')
+             .addClass('llm-error')
+             .text(msg);
+        $area.removeAttr('data-pending');
+    }
+
+    var llmAllChanges      = $('#llm_all_changes').is(':checked')    ? 1 : 0;
+    var ignoreWhitespace   = $('#ignoreWhitespace').is(':checked')   ? 1 : 0;
+    var showRemoved        = $('#removed').is(':checked')            ? 1 : 0;
+    var showAdded          = $('#added').is(':checked')              ? 1 : 0;
+    $.getJSON(summaryUrl, {
+        from_version:      fromVersion,
+        to_version:        toVersion,
+        all_changes:       llmAllChanges,
+        ignore_whitespace: ignoreWhitespace,
+        removed:           showRemoved,
+        added:             showAdded,
+    })
+        .done(function (data) {
+            if (data.summary) {
+                $area.find('.llm-diff-summary-text')
+                     .removeClass('llm-diff-summary-loading')
+                     .text(data.summary);
+                $area.removeAttr('data-pending');
+            } else if (data.error) {
+                showLlmError(data.error);
+            } else {
+                $area.remove();
+            }
+        })
+        .fail(function (xhr) {
+            var resp = xhr.responseJSON;
+            if (resp && resp.error) {
+                showLlmError(resp.error);
+            } else {
+                showLlmError('AI summary request failed (HTTP ' + xhr.status + ').');
+            }
+        });
+});
+</script>
+{% endif %}

 {% endblock %}
@@ -1,6 +1,6 @@
 {% extends 'base.html' %}
 {% block content %}
-{% from '_helpers.html' import render_field, render_checkbox_field, render_button, render_time_schedule_form, playwright_warning, only_playwright_type_watches_warning, highlight_trigger_ignored_explainer, render_conditions_fieldlist_of_formfields_as_table, render_ternary_field %}
+{% from '_helpers.html' import render_field, render_checkbox_field, render_button, render_time_schedule_form, only_playwright_type_watches_warning, highlight_trigger_ignored_explainer, render_conditions_fieldlist_of_formfields_as_table, render_ternary_field %}
 {% from '_common_fields.html' import render_common_settings_form %}
 <script src="{{url_for('static_content', group='js', filename='tabs.js')}}" defer></script>
 <script src="{{url_for('static_content', group='js', filename='vis.js')}}" defer></script>
@@ -57,6 +57,9 @@
            {% if capabilities.supports_visual_selector %}
            <li class="tab"><a id="visualselector-tab" href="#visualselector">{{ _('Visual Filter Selector') }}</a></li>
            {% endif %}
+            {% if not llm_features_disabled %}
+            <li class="tab"><a href="#ai-llm">{{ _('AI / LLM') }}</a></li>
+            {% endif %}
            {% if capabilities.supports_text_filters_and_triggers %}
            <li class="tab" id="filters-and-triggers-tab"><a href="#filters-and-triggers">{{ _('Filters & Triggers') }}</a></li>
            <li class="tab" id="conditions-tab"><a href="#conditions">{{ _('Conditions') }}</a></li>
@@ -81,6 +84,14 @@
                    <div class="pure-control-group">
                        {{ render_field(form.tags) }}
                        <span class="pure-form-message-inline">{{ _('Organisational tag/group name used in the main listing page') }}</span>
+                        {% if auto_applied_tags %}
+                        <span class="pure-form-message-inline">
+                            {{ _('Also automatically applied by URL pattern:') }}
+                            {% for tag_uuid, tag in auto_applied_tags.items() %}
+                            <a href="{{ url_for('tags.form_tag_edit', uuid=tag_uuid) }}" class="watch-tag-list tag-{{ tag.title|sanitize_tag_class }}">{{ tag.title }}</a>
+                            {% endfor %}
+                        </span>
+                        {% endif %}
                    </div>
                    <div class="pure-control-group inline-radio">
                        {{ render_field(form.processor) }}
@@ -133,8 +144,8 @@
                    <div class="pure-control-group inline-radio">
                        {{ render_field(form.fetch_backend, class="fetch-backend") }}
                        <span class="pure-form-message-inline">
-                            <p>{{ _('Use the') }} <strong>{{ _('Basic') }}</strong> {{ _('method (default) where your watched site doesn\'t need Javascript to render.') }}</p>
-                            <p>{{ _('The') }} <strong>{{ _('Chrome/Javascript') }}</strong> {{ _('method requires a network connection to a running WebDriver+Chrome server, set by the ENV var \'WEBDRIVER_URL\'.') }} </p>
+                            <p>{{ _('Use the <strong>Basic</strong> method (default) where your watched sites don\'t need Javascript to render.')|safe }}</p>
+                            <p>{{ _('The <strong>Chrome/Javascript</strong> method requires a network connection to a running WebDriver+Chrome server, set by the ENV var \'WEBDRIVER_URL\'.')|safe }}</p>
                            {{ _('Tip:') }} <a href="https://github.com/dgtlmoon/changedetection.io/wiki/Proxy-configuration#brightdata-proxy-support">{{ _('Connect using Bright Data and Oxylabs Proxies, find out more here.') }}</a>
                        </span>
                    </div>
@@ -155,7 +166,7 @@
                        <div class="pure-form-message-inline">
                            <strong>{{ _('If you\'re having trouble waiting for the page to be fully rendered (text missing etc), try increasing the \'wait\' time here.') }}</strong>
                            <br>
-                            {{ _('This will wait') }} <i>n</i> {{ _('seconds before extracting the text.') }}
+                            {{ _('This will wait <i>n</i> seconds before extracting the text.')|safe }}
                            {% if using_global_webdriver_wait %}
                            <br><strong>{{ _('Using the current global default settings') }}</strong>
                            {% endif %}
@@ -288,7 +299,7 @@ Math: {{ 1 + 1 }}") }}
                        {% if has_default_notification_urls %}
                        <div class="inline-warning">
                            <img class="inline-warning-icon" src="{{url_for('static_content', group='images', filename='notice.svg')}}" alt="{{ _('Look out!') }}" title="{{ _('Lookout!') }}" >
-                            {{ _('There are') }} <a href="{{ url_for('settings.settings_page')}}#notifications">{{ _('system-wide notification URLs enabled') }}</a>, {{ _('this form will override notification settings for this watch only') }} &dash; {{ _('an empty Notification URL list here will still send notifications.') }}
+                            {{ _('There are <a href="%(url)s">system-wide notification URLs enabled</a>, this form will override notification settings for this watch only &dash; an empty Notification URL list here will still send notifications.', url=url_for('settings.settings_page') ~ '#notifications')|safe }}
                        </div>
                        {% endif %}
                        <a href="#notifications" id="notification-setting-reset-to-default" class="pure-button button-xsmall" style="right: 20px; top: 20px; position: absolute; background-color: #5f42dd; border-radius: 4px; font-size: 70%; color: #fff">{{ _('Use system defaults') }}</a>
@@ -312,7 +323,13 @@ Math: {{ 1 + 1 }}") }}
                    </div>
                </div>
            </div>
+            {% if not llm_features_disabled %}
+            <div class="tab-pane-inner" id="ai-llm">
+                {% include "edit/include_llm_intent.html" %}
+            </div>
+            {% endif %}
            <div class="tab-pane-inner" id="filters-and-triggers">
+
                <span id="activate-text-preview" class="pure-button pure-button-primary button-xsmall">{{ _('Activate preview') }}</span>
              <div>
              <div id="edit-text-filter">
@@ -338,7 +355,7 @@ Math: {{ 1 + 1 }}") }}
                        {{ render_checkbox_field(form.filter_text_added) }}
                        {{ render_checkbox_field(form.filter_text_replaced) }}
                        {{ render_checkbox_field(form.filter_text_removed) }}
-                    <span class="pure-form-message-inline">{{ _('Note: Depending on the length and similarity of the text on each line, the algorithm may consider an') }} <strong>{{ _('addition') }}</strong> {{ _('instead of') }} <strong>{{ _('replacement') }}</strong> {{ _('for example.') }}</span><br>
+                    <span class="pure-form-message-inline">{{ _('Note: Depending on the length and similarity of the text on each line, the algorithm may consider an <strong>addition</strong> instead of <strong>replacement</strong> for example.')|safe }}</span><br>
                    <span class="pure-form-message-inline">&nbsp;{{ _('So it\'s always better to select') }} <strong>{{ _('Added') }}</strong>+<strong>{{ _('Replaced') }}</strong> {{ _('when you\'re interested in new content.') }}</span><br>
                    <span class="pure-form-message-inline">&nbsp;{{ _('When content is merely moved in a list, it will also trigger an') }} <strong>{{ _('addition') }}</strong>, {{ _('consider enabling') }} <code><strong>{{ _('Only trigger when unique lines appear') }}</strong></code></span>
                </fieldset>
@@ -352,7 +369,8 @@ Math: {{ 1 + 1 }}") }}
                </fieldset>
                <fieldset class="pure-control-group">
                    {{ render_checkbox_field(form.sort_text_alphabetically) }}
-                    <span class="pure-form-message-inline">{{ _('Helps reduce changes detected caused by sites shuffling lines around, combine with') }} <i>{{ _('check unique lines') }}</i> {{ _('below.') }}</span>
+                    {# TRANSLATORS: CJK fonts lack native italics; allow substitution with conventional local styling. dennis-ignore: W303 #}
+                    <span class="pure-form-message-inline">{{ _('Helps reduce changes detected caused by sites shuffling lines around, combine with <i>check unique lines</i> below.')|safe }}</span>
                </fieldset>
                <fieldset class="pure-control-group">
                    {{ render_checkbox_field(form.trim_text_whitespace) }}
@@ -366,7 +384,20 @@ Math: {{ 1 + 1 }}") }}
                    const preview_text_edit_filters_url="{{url_for('ui.ui_edit.watch_get_preview_rendered', uuid=uuid)}}";
                </script>
                <br>
-                {#<div id="text-preview-controls"><span id="text-preview-refresh" class="pure-button button-xsmall">Refresh</span></div>#}
+                {% if llm_configured %}
+                <div id="llm-preview-result" style="display:none; margin-bottom: 0.8em; padding: 0.8em 1.1em; border-radius: 4px; border-left: 4px solid #ccc; font-size: 0.9em;">
+                    <div style="font-size:0.75em; text-transform:uppercase; letter-spacing:0.06em; opacity:0.55; margin-bottom:0.35em;">{{ _('AI Intent preview') }}</div>
+                    <span class="llm-preview-verdict" style="font-weight: bold;"></span>
+                    <div class="llm-preview-answer" style="margin-top: 0.5em; white-space: pre-wrap; line-height: 1.5; font-style: italic;"></div>
+                </div>
+                <style>
+                    #llm-preview-result { transition: border-color 0.2s, background 0.2s; }
+                    #llm-preview-result[data-found="1"] { border-color: #2ecc71; background: rgba(46,204,113,0.07); }
+                    #llm-preview-result[data-found="1"] .llm-preview-verdict { color: #27ae60; }
+                    #llm-preview-result[data-found="0"] { border-color: #aaa; background: rgba(0,0,0,0.03); }
+                    #llm-preview-result[data-found="0"] .llm-preview-verdict { color: #888; }
+                </style>
+                {% endif %}
                <div class="minitabs-wrapper">
                  <div class="minitabs-content">
                      <div id="text-preview-inner" class="monospace-preview">
@@ -476,6 +507,16 @@ Math: {{ 1 + 1 }}") }}
                            <td>{{ _('Server type reply') }}</td>
                            <td>{{ watch.get('remote_server_reply') }}</td>
                        </tr>
+                        {% if not llm_features_disabled and settings_application.get('llm', {}).get('model') %}
+                        <tr>
+                            <td>{{ _('AI tokens (last check)') }}</td>
+                            <td>{{ "{:,}".format(watch.get('llm_last_tokens_used') or 0) }}</td>
+                        </tr>
+                        <tr>
+                            <td>{{ _('AI tokens (total)') }}</td>
+                            <td>{{ "{:,}".format(watch.get('llm_tokens_used_cumulative') or 0) }}</td>
+                        </tr>
+                        {% endif %}
                        </tbody>
                    </table>

@@ -17,7 +17,7 @@
    <script src="{{ url_for('static_content', group='js', filename='tabs.js') }}" defer></script>
    {% if versions|length >= 2 %}
        <div id="diff-form" style="text-align: center;">
-            <form class="pure-form " action="" method="POST">
+            <form class="pure-form "  action="{{url_for('ui.ui_preview.preview_page', uuid=uuid)}}" method="POST">
                <fieldset>
                    <label for="preview-version">{{ _('Select timestamp') }}</label> <select id="preview-version"
                                                                                 name="from_version"
@@ -28,6 +28,7 @@
                        </option>
                    {% endfor %}
                </select>
+                    <input type="hidden" name="csrf_token" value="{{ csrf_token() }}">
                    <button type="submit" class="pure-button pure-button-primary">{{ _('Go') }}</button>

                </fieldset>
@@ -26,7 +26,11 @@ def construct_blueprint(datastore: ChangeDetectionStore, update_q, queuedWatchMe
        add_paused = request.form.get('edit_and_watch_submit_button') != None
        from changedetectionio import processors
        processor = request.form.get('processor', processors.get_default_processor())
-        new_uuid = datastore.add_watch(url=url, tag=request.form.get('tags','').strip(), extras={'paused': add_paused, 'processor': processor})
+        llm_intent = request.form.get('llm_intent', '').strip()
+        extras = {'paused': add_paused, 'processor': processor}
+        if llm_intent:
+            extras['llm_intent'] = llm_intent
+        new_uuid = datastore.add_watch(url=url, tag=request.form.get('tags','').strip(), extras=extras)

        if new_uuid:
            if add_paused:
@@ -81,6 +81,12 @@ def construct_blueprint(datastore: ChangeDetectionStore, update_q, queuedWatchMe

        sorted_tags = sorted(datastore.data['settings']['application'].get('tags').items(), key=lambda x: x[1]['title'])

+        proxy_list = datastore.proxy_list
+
+        from changedetectionio.llm.evaluator import get_llm_config as _get_llm_config
+        from changedetectionio.llm.ui_strings import LLM_INTENT_WATCH_PLACEHOLDER
+        llm_configured = bool(_get_llm_config(datastore))
+
        output = render_template(
            "watch-overview.html",
            active_tag=active_tag,
@@ -88,11 +94,12 @@ def construct_blueprint(datastore: ChangeDetectionStore, update_q, queuedWatchMe
            app_rss_token=datastore.data['settings']['application'].get('rss_access_token'),
            datastore=datastore,
            errored_count=errored_count,
-            extra_classes='has-queue' if not update_q.empty() else '',
+            extra_classes=' '.join(filter(None, ['has-queue' if not update_q.empty() else '', 'llm-configured' if llm_configured else ''])),
            form=form,
            generate_tag_colors=processors.generate_processor_badge_colors,
+            wcag_text_color=processors.wcag_text_color,
            guid=datastore.data['app_guid'],
-            has_proxies=datastore.proxy_list,
+            has_proxies=proxy_list,
            hosted_sticky=os.getenv("SALTED_PASS", False) == False,
            now_time_server=round(time.time()),
            pagination=pagination,
@@ -107,9 +114,21 @@ def construct_blueprint(datastore: ChangeDetectionStore, update_q, queuedWatchMe
            system_default_fetcher=datastore.data['settings']['application'].get('fetch_backend'),
            tags=sorted_tags,
            unread_changes_count=datastore.unread_changes_count,
-            watches=sorted_watches
+            watches=sorted_watches,
+            llm_configured=llm_configured,
+            llm_intent_watch_placeholder=LLM_INTENT_WATCH_PLACEHOLDER,
        )

+        # Return freed template-building memory to the OS immediately.
+        # render_template allocates ~20MB of intermediate strings that are freed on return,
+        # but glibc keeps those pages mapped in its arenas as RSS. malloc_trim() forces
+        # glibc to release them, preventing RSS growth from concurrent Chrome connections.
+        try:
+            import ctypes
+            ctypes.CDLL('libc.so.6').malloc_trim(0)
+        except Exception:
+            pass
+
        if session.get('share-link'):
            del (session['share-link'])

@@ -71,6 +71,13 @@ document.addEventListener('DOMContentLoaded', function() {
 {%- for uuid, tag in tags -%}
 {%- if tag and tag.title -%}
 {%- set class_name = tag.title|sanitize_tag_class -%}
+{%- if tag.get('tag_colour') -%}
+.button-tag.tag-{{ class_name }},
+.watch-tag-list.tag-{{ class_name }} {
+  background-color: {{ tag.tag_colour }};
+  color: {{ wcag_text_color(tag.tag_colour) }};
+}
+{%- else -%}
 {%- set colors = generate_tag_colors(tag.title) -%}
 .button-tag.tag-{{ class_name }} {
  background-color: {{ colors['light']['bg'] }};
@@ -92,6 +99,7 @@ html[data-darkmode="true"] .watch-tag-list.tag-{{ class_name }} {
  color: {{ colors['dark']['color'] }};
 }
 {%- endif -%}
+{%- endif -%}
 {%- endfor -%}
 </style>
 <div class="box" id="form-quick-watch-add">
@@ -105,6 +113,16 @@ html[data-darkmode="true"] .watch-tag-list.tag-{{ class_name }} {
                    {{ render_nolabel_field(form.watch_submit_button, title=_("Watch this URL!") ) }}
                    {{ render_nolabel_field(form.edit_and_watch_submit_button, title=_("Edit first then Watch") ) }}
            </div>
+            {% if llm_configured %}
+            <div id="quick-watch-llm-intent" style="display:none; margin-top: 0.5em;">
+                <textarea name="llm_intent"
+                          id="quick_watch_llm_intent"
+                          rows="2"
+                          class="pure-input-1"
+                          placeholder="{{ _('AI — Notify when…') }} {{ llm_intent_watch_placeholder }}"
+                ></textarea>
+            </div>
+            {% endif %}
            <div id="watch-group-tag">
               {{ render_field(form.tags, value=active_tag.title if active_tag_uuid else '', placeholder=_("Watch group / tag"), class="transparent-field") }}
            </div>
@@ -118,6 +136,14 @@ html[data-darkmode="true"] .watch-tag-list.tag-{{ class_name }} {
        </span>
    </form>
 </div>
+{% if llm_configured %}
+<script>
+window.watchOverviewI18n = {
+    generatingSummary: {{ _('Generating summary…')|tojson }},
+    gotoHistory:       {{ _('Goto full history')|tojson }}
+};
+</script>
+{% endif %}
 <div class="box">
    <form class="pure-form" action="{{ url_for('ui.form_watch_list_checkbox_operations') }}" method="POST" id="watch-list-form">
    <input type="hidden" name="csrf_token" value="{{ csrf_token() }}" >
@@ -137,12 +163,13 @@ html[data-darkmode="true"] .watch-tag-list.tag-{{ class_name }} {
                       data-confirm-type="danger"
                       data-confirm-title="{{ _('Clear Histories') }}"
                       data-confirm-message="{{ _('<p>Are you sure you want to clear history for the selected items?</p><p>This action cannot be undone.</p>') }}"
+                       {# TRANSLATORS: Universally recognized; typically left as-is. dennis-ignore: W302 #}
                       data-confirm-button="{{ _('OK') }}"><i data-feather="trash-2" style="width: 14px; height: 14px; stroke: white; margin-right: 4px;"></i>{{ _('Clear/reset history') }}</button>
        <button class="pure-button button-secondary button-xsmall" style="background: #dd4242;" name="op" value="delete"
                       data-requires-confirm
                       data-confirm-type="danger"
                       data-confirm-title="{{ _('Delete Watches?') }}"
-                       data-confirm-message="{{ _('<p>Are you sure you want to delete the selected watches?</strong></p><p>This action cannot be undone.</p>') }}"
+                       data-confirm-message="{{ _('<p><strong>Are you sure you want to delete the selected watches?</strong></p><p>This action cannot be undone.</p>') }}"
                       data-confirm-button="{{ _('Delete') }}"><i data-feather="trash" style="width: 14px; height: 14px; stroke: white; margin-right: 4px;"></i>{{ _('Delete') }}</button>
    </div>

@@ -197,8 +224,8 @@ html[data-darkmode="true"] .watch-tag-list.tag-{{ class_name }} {
             {%- if any_has_restock_price_processor -%}
                <th>{{ _('Restock & Price') }}</th>
             {%- endif -%}
-                <th><a class="{{ 'active '+link_order if sort_attribute == 'last_checked' else 'inactive' }}" href="{{url_for('watchlist.index', sort='last_checked', order=link_order, tag=active_tag_uuid)}}"><span class="hide-on-mobile">{{ _('Last') }}</span> {{ _('Checked') }} <span class='arrow {{link_order}}'></span></a></th>
-                <th><a class="{{ 'active '+link_order if sort_attribute == 'last_changed' else 'inactive' }}" href="{{url_for('watchlist.index', sort='last_changed', order=link_order, tag=active_tag_uuid)}}"><span class="hide-on-mobile">{{ _('Last') }}</span> {{ _('Changed') }} <span class='arrow {{link_order}}'></span></a></th>
+                <th><a class="{{ 'active '+link_order if sort_attribute == 'last_checked' else 'inactive' }}" href="{{url_for('watchlist.index', sort='last_checked', order=link_order, tag=active_tag_uuid)}}"><span class="hide-on-mobile">{{ _('Last Checked') }}</span><span class="hide-on-desktop">{{ _('Checked') }}</span> <span class='arrow {{link_order}}'></span></a></th>
+                <th><a class="{{ 'active '+link_order if sort_attribute == 'last_changed' else 'inactive' }}" href="{{url_for('watchlist.index', sort='last_changed', order=link_order, tag=active_tag_uuid)}}"><span class="hide-on-mobile">{{ _('Last Changed') }}</span><span class="hide-on-desktop">{{ _('Changed') }}</span> <span class='arrow {{link_order}}'></span></a></th>
                <th class="empty-cell"></th>
            </tr>
            </thead>
@@ -213,12 +240,13 @@ html[data-darkmode="true"] .watch-tag-list.tag-{{ class_name }} {
                {%- set checking_now = is_checking_now(watch) -%}
                {%- set history_n = watch.history_n -%}
                {%- set favicon = watch.get_favicon_filename() -%}
+                {%- set error_texts = watch.compile_error_texts(has_proxies=has_proxies) -%}
                {%- set system_use_url_watchlist = datastore.data['settings']['application']['ui'].get('use_page_title_in_list')  -%}
                {#  Class settings mirrored in changedetectionio/static/js/realtime.js for the frontend #}
                {%- set row_classes = [
                    loop.cycle('pure-table-odd', 'pure-table-even'),
                    'processor-' ~ watch['processor'],
-                    'has-error' if watch.compile_error_texts()|length > 2 else '',
+                    'has-error' if error_texts|length > 2 else '',
                    'paused' if watch.paused is defined and watch.paused != False else '',
                    'unviewed' if watch.has_unviewed else '',
                    'has-restock-info' if watch.has_restock_info else 'no-restock-info',
@@ -236,10 +264,10 @@ html[data-darkmode="true"] .watch-tag-list.tag-{{ class_name }} {
                <td class="inline checkbox-uuid" ><div><input name="uuids"  type="checkbox" value="{{ watch.uuid}} " > <span class="counter-i">{{ loop.index+pagination.skip }}</span></div></td>
                <td class="inline watch-controls">
                    <div>
-                    <a class="ajax-op state-off pause-toggle" data-op="pause" href="{{url_for('watchlist.index', op='pause', uuid=watch.uuid, tag=active_tag_uuid)}}"><img src="{{url_for('static_content', group='images', filename='pause.svg')}}" alt="Pause checks" title="Pause checks" class="icon icon-pause" ></a>
-                    <a class="ajax-op state-on pause-toggle"  data-op="pause" style="display: none" href="{{url_for('watchlist.index', op='pause', uuid=watch.uuid, tag=active_tag_uuid)}}"><img src="{{url_for('static_content', group='images', filename='play.svg')}}" alt="UnPause checks" title="UnPause checks" class="icon icon-unpause" ></a>
-                    <a class="ajax-op state-off mute-toggle" data-op="mute" href="{{url_for('watchlist.index', op='mute', uuid=watch.uuid, tag=active_tag_uuid)}}"><img src="{{url_for('static_content', group='images', filename='bell-off.svg')}}" alt="Mute notification" title="Mute notification" class="icon icon-mute" ></a>
-                    <a class="ajax-op state-on mute-toggle" data-op="mute"  style="display: none" href="{{url_for('watchlist.index', op='mute', uuid=watch.uuid, tag=active_tag_uuid)}}"><img src="{{url_for('static_content', group='images', filename='bell-off.svg')}}" alt="UnMute notification" title="UnMute notification" class="icon icon-mute" ></a>
+                    <a class="ajax-op state-off pause-toggle" data-op="pause" href="{{url_for('watchlist.index', op='pause', uuid=watch.uuid, tag=active_tag_uuid)}}"><img src="{{url_for('static_content', group='images', filename='pause.svg')}}" alt="{{ _('Pause checks') }}" title="{{ _('Pause checks') }}" class="icon icon-pause" ></a>
+                    <a class="ajax-op state-on pause-toggle"  data-op="pause" style="display: none" href="{{url_for('watchlist.index', op='pause', uuid=watch.uuid, tag=active_tag_uuid)}}"><img src="{{url_for('static_content', group='images', filename='play.svg')}}" alt="{{ _('UnPause checks') }}" title="{{ _('UnPause checks') }}" class="icon icon-unpause" ></a>
+                    <a class="ajax-op state-off mute-toggle" data-op="mute" href="{{url_for('watchlist.index', op='mute', uuid=watch.uuid, tag=active_tag_uuid)}}"><img src="{{url_for('static_content', group='images', filename='bell-off.svg')}}" alt="{{ _('Mute notification') }}" title="{{ _('Mute notification') }}" class="icon icon-mute" ></a>
+                    <a class="ajax-op state-on mute-toggle" data-op="mute"  style="display: none" href="{{url_for('watchlist.index', op='mute', uuid=watch.uuid, tag=active_tag_uuid)}}"><img src="{{url_for('static_content', group='images', filename='bell-off.svg')}}" alt="{{ _('UnMute notification') }}" title="{{ _('UnMute notification') }}" class="icon icon-mute" ></a>
                    </div>
                </td>

@@ -271,7 +299,7 @@ html[data-darkmode="true"] .watch-tag-list.tag-{{ class_name }} {
                                {% endif %}
                               <a class="external" target="_blank" rel="noopener" href="{{ watch.link.replace('source:','') }}">&nbsp;</a>
                            </span>
-                            <div class="error-text" style="display:none;">{{ watch.compile_error_texts(has_proxies=datastore.proxy_list)|safe }}</div>
+                            <div class="error-text" style="display:none;">{{ error_texts|safe }}</div>
                            {%- if watch['processor'] == 'text_json_diff'  -%}
                                {%- if watch['has_ldjson_price_data'] and not watch['track_ldjson_price_data']  -%}
                                <div class="ldjson-price-track-offer">Switch to Restock & Price watch mode? <a href="{{url_for('price_data_follower.accept', uuid=watch.uuid)}}" class="pure-button button-xsmall">Yes</a> <a href="{{url_for('price_data_follower.reject', uuid=watch.uuid)}}" class="">No</a></div>
@@ -283,7 +311,7 @@ html[data-darkmode="true"] .watch-tag-list.tag-{{ class_name }} {
                            {%- endfor -%}
                        </div>
                    <div class="status-icons">
-                            <a class="link-spread" href="{{url_for('ui.form_share_put_watch', uuid=watch.uuid)}}"><img src="{{url_for('static_content', group='images', filename='spread.svg')}}" class="status-icon icon icon-spread" title="Create a link to share watch config with others" ></a>
+                            <a class="link-spread" href="{{url_for('ui.form_share_put_watch', uuid=watch.uuid)}}"><img src="{{url_for('static_content', group='images', filename='spread.svg')}}" class="status-icon icon icon-spread" title="{{ _('Create a link to share watch config with others') }}" ></a>
                            {%- set effective_fetcher = watch.get_fetch_backend if watch.get_fetch_backend != "system" else system_default_fetcher -%}
                            {%- if effective_fetcher and ("html_webdriver" in effective_fetcher or "html_" in effective_fetcher or "extra_browser_" in effective_fetcher) -%}
                                {{ effective_fetcher|fetcher_status_icons }}
@@ -304,11 +332,20 @@ html[data-darkmode="true"] .watch-tag-list.tag-{{ class_name }} {
                            </span>
                        {%- endif -%}

-                        {%- if watch.get('restock') and watch['restock']['price'] != None -%}
-                            {%- if watch['restock']['price'] != None -%}
-                                <span class="restock-label price" title="{{ _('Price') }}">
-                                {{ watch['restock']['price']|format_number_locale if watch['restock'].get('price') else '' }} {{ watch['restock'].get('currency','') }}
-                                </span>
+                        {%- if watch.get('restock') and watch['restock'].get('price') -%}
+                            {%- set restock = watch['restock'] -%}
+                            {%- set price = restock.get('price') -%}
+                            {%- set cur = restock.get('currency','') -%}
+
+                            {%- if price is not none and (price|string)|regex_search('\d') -%}
+                              <span class="restock-label price" title="{{ _('Price') }}">
+                              {# @todo: make parse_currency/parse_decimal aware of the locale of the actual web page and use that instead changedetectionio/processors/restock_diff/__init__.py #}
+                                {%- if price is number -%}{# It's a number so we can convert it to their locale' #}
+                                  {{ price|format_number_locale }} {{ cur }}<!-- as number -->
+                                {%- else -%}{# It's totally fine if it arrives as something else, the website might be something weird in this field #}
+                                  {{ price }} {{ cur }}<!-- as string -->
+                                {%- endif -%}
+                              </span>
                            {%- endif -%}
                        {%- elif not watch.has_restock_info -%}
                            <span class="restock-label error">{{ _('No information') }}</span>
@@ -317,13 +354,13 @@ html[data-darkmode="true"] .watch-tag-list.tag-{{ class_name }} {
                </td>
 {%- endif -%}
            {#last_checked becomes fetch-start-time#}
-                <td class="last-checked" data-timestamp="{{ watch.last_checked }}" data-fetchduration={{ watch.fetch_time }} data-eta_complete="{{ watch.last_checked+watch.fetch_time }}" >
+                <td class="last-checked" data-timestamp="{{ watch.last_checked }}" data-fetchduration={{ watch.fetch_time }} data-eta_complete="{{ watch.last_checked+watch.fetch_time }}" data-label="{{ _('Last Checked') }}">
                    <div class="spinner-wrapper" style="display:none;" >
-                        <span class="spinner"></span><span class="status-text">&nbsp;{{ _('Checking now') }}</span>
+                        <span class="spinner"></span><span class="status-text">&nbsp;{{ watch['__check_status'] or _('Checking now') }}</span>
                    </div>
                    <span class="innertext">{{watch|format_last_checked_time|safe}}</span>
                </td>
-                <td class="last-changed" data-timestamp="{{ watch.last_changed }}">{%- if watch.history_n >=2 and watch.last_changed >0 -%}
+                <td class="last-changed" data-timestamp="{{ watch.last_changed }}" data-label="{{ _('Last Changed') }}">{%- if watch.history_n >=2 and watch.last_changed >0 -%}
                    {{watch.last_changed|format_timestamp_timeago}}
                    {%- else -%}
                    {{ _('Not yet') }}
@@ -335,7 +372,7 @@ html[data-darkmode="true"] .watch-tag-list.tag-{{ class_name }} {
                    <a href="" class="already-in-queue-button recheck pure-button pure-button-primary" style="display: none;" disabled="disabled">{{ _('Queued') }}</a>
                    <a href="{{ url_for('ui.form_watch_checknow', uuid=watch.uuid, tag=request.args.get('tag')) }}" data-op='recheck' class="ajax-op recheck pure-button pure-button-primary">{{ _('Recheck') }}</a>
                    <a href="{{ url_for('ui.ui_edit.edit_page', uuid=watch.uuid, tag=active_tag_uuid)}}#general" class="pure-button pure-button-primary">{{ _('Edit') }}</a>
-                    <a href="{{ url_for('ui.ui_diff.diff_history_page', uuid=watch.uuid)}}" {{target_attr}} class="pure-button pure-button-primary history-link" style="display: none;">{{ _('History') }}</a>
+                    <a href="{{ url_for('ui.ui_diff.diff_history_page', uuid=watch.uuid)}}" {{target_attr}} class="pure-button pure-button-primary history-link ai-history-btn" style="display: none;" data-uuid="{{ watch.uuid }}" data-summary-url="{{ url_for('ui.ui_diff.diff_llm_summary', uuid=watch.uuid) }}"><span class="btn-label-history">{{ _('History') }}</span><span class="btn-label-summary">&#x2728; {{ _('Summary') }}</span></a>
                    <a href="{{ url_for('ui.ui_preview.preview_page', uuid=watch.uuid)}}" {{target_attr}} class="pure-button pure-button-primary preview-link" style="display: none;">{{ _('Preview') }}</a>
                    </div>
                </td>
@@ -1,4 +1,5 @@
 from json_logic.builtins import BUILTINS
+from flask_babel import lazy_gettext as _l

 from .exceptions import EmptyConditionRuleRowNotUsable
 from .pluggy_interface import plugin_manager  # Import the pluggy plugin manager
@@ -6,19 +7,19 @@ from . import default_plugin
 from loguru import logger
 # List of all supported JSON Logic operators
 operator_choices = [
-    (None, "Choose one - Operator"),
-    (">", "Greater Than"),
-    ("<", "Less Than"),
-    (">=", "Greater Than or Equal To"),
-    ("<=", "Less Than or Equal To"),
-    ("==", "Equals"),
-    ("!=", "Not Equals"),
-    ("in", "Contains"),
+    (None, _l("Choose one - Operator")),
+    (">", _l("Greater Than")),
+    ("<", _l("Less Than")),
+    (">=", _l("Greater Than or Equal To")),
+    ("<=", _l("Less Than or Equal To")),
+    ("==", _l("Equals")),
+    ("!=", _l("Not Equals")),
+    ("in", _l("Contains")),
 ]

 # Fields available in the rules
 field_choices = [
-    (None, "Choose one - Field"),
+    (None, _l("Choose one - Field")),
 ]

 # The data we will feed the JSON Rules to see if it passes the test/conditions or not
@@ -3,6 +3,7 @@ import re
 import pluggy
 from price_parser import Price
 from loguru import logger
+from flask_babel import lazy_gettext as _l

 hookimpl = pluggy.HookimplMarker("changedetectionio_conditions")

@@ -47,22 +48,22 @@ def register_operators():
@hookimpl
 def register_operator_choices():
    return [
-        ("!in", "Does NOT Contain"),
-        ("starts_with", "Text Starts With"),
-        ("ends_with", "Text Ends With"),
-        ("length_min", "Length minimum"),
-        ("length_max", "Length maximum"),
-        ("contains_regex", "Text Matches Regex"),
-        ("!contains_regex", "Text Does NOT Match Regex"),
+        ("!in", _l("Does NOT Contain")),
+        ("starts_with", _l("Text Starts With")),
+        ("ends_with", _l("Text Ends With")),
+        ("length_min", _l("Length minimum")),
+        ("length_max", _l("Length maximum")),
+        ("contains_regex", _l("Text Matches Regex")),
+        ("!contains_regex", _l("Text Does NOT Match Regex")),
    ]

@hookimpl
 def register_field_choices():
    return [
-        ("extracted_number", "Extracted number after 'Filters & Triggers'"),
+        ("extracted_number", _l("Extracted number after 'Filters & Triggers'")),
 #        ("meta_description", "Meta Description"),
 #        ("meta_keywords", "Meta Keywords"),
-        ("page_filtered_text", "Page text after 'Filters & Triggers'"),
+        ("page_filtered_text", _l("Page text after 'Filters & Triggers'")),
        #("page_title", "Page <title>"), # actual page title <title>
    ]

@@ -1,6 +1,7 @@
 # Condition Rule Form (for each rule row)
 from wtforms import Form, SelectField, StringField, validators
 from wtforms import validators
+from flask_babel import lazy_gettext as _l

 class ConditionFormRow(Form):

@@ -8,18 +9,18 @@ class ConditionFormRow(Form):
    from changedetectionio.conditions import plugin_manager
    from changedetectionio.conditions import operator_choices, field_choices
    field = SelectField(
-        "Field",
+        _l("Field"),
        choices=field_choices,
        validators=[validators.Optional()]
    )

    operator = SelectField(
-        "Operator",
+        _l("Operator"),
        choices=operator_choices,
        validators=[validators.Optional()]
    )

-    value = StringField("Value", validators=[validators.Optional()], render_kw={"placeholder": "A value"})
+    value = StringField(_l("Value"), validators=[validators.Optional()], render_kw={"placeholder": _l("A value")})

    def validate(self, extra_validators=None):
        # First, run the default validators
@@ -30,15 +31,15 @@ class ConditionFormRow(Form):
        # If any of the operator/field/value is set, then they must be all set
        if any(value not in ("", False, "None", None) for value in [self.operator.data, self.field.data, self.value.data]):
            if not self.operator.data or self.operator.data == 'None':
-                self.operator.errors.append("Operator is required.")
+                self.operator.errors.append(_l("Operator is required."))
                return False

            if not self.field.data or self.field.data == 'None':
-                self.field.errors.append("Field is required.")
+                self.field.errors.append(_l("Field is required."))
                return False

            if not self.value.data:
-                self.value.errors.append("Value is required.")
+                self.value.errors.append(_l("Value is required."))
                return False

        return True  # Only return True if all conditions pass
@@ -4,6 +4,7 @@ Provides metrics for measuring text similarity between snapshots.
 """
 import pluggy
 from loguru import logger
+from flask_babel import gettext as _, lazy_gettext as _l

 LEVENSHTEIN_MAX_LEN_FOR_EDIT_STATS=100000

@@ -53,8 +54,8 @@ def register_operator_choices():
@conditions_hookimpl
 def register_field_choices():
    return [
-        ("levenshtein_ratio", "Levenshtein - Text similarity ratio"),
-        ("levenshtein_distance", "Levenshtein - Text change distance"),
+        ("levenshtein_ratio", _l("Levenshtein - Text similarity ratio")),
+        ("levenshtein_distance", _l("Levenshtein - Text change distance")),
    ]

@conditions_hookimpl
@@ -77,7 +78,7 @@ def ui_edit_stats_extras(watch):
    """Add Levenshtein stats to the UI using the global plugin system"""
    """Generate the HTML for Levenshtein stats - shared by both plugin systems"""
    if len(watch.history.keys()) < 2:
-        return "<p>Not enough history to calculate Levenshtein metrics</p>"
+        return f"<p>{_('Not enough history to calculate Levenshtein metrics')}</p>"


    # Protection against the algorithm getting stuck on huge documents
@@ -87,37 +88,37 @@ def ui_edit_stats_extras(watch):
            for idx in (-1, -2)
            if len(k) >= abs(idx)
    ):
-        return "<p>Snapshot too large for edit statistics, skipping.</p>"
+        return f"<p>{_('Snapshot too large for edit statistics, skipping.')}</p>"

    try:
        lev_data = levenshtein_ratio_recent_history(watch)
        if not lev_data or not isinstance(lev_data, dict):
-            return "<p>Unable to calculate Levenshtein metrics</p>"
-            
+            return f"<p>{_('Unable to calculate Levenshtein metrics')}</p>"
+
        html = f"""
        <div class="levenshtein-stats">
-            <h4>Levenshtein Text Similarity Details</h4>
+            <h4>{_('Levenshtein Text Similarity Details')}</h4>
            <table class="pure-table">
                <tbody>
                    <tr>
-                        <td>Raw distance (edits needed)</td>
+                        <td>{_('Raw distance (edits needed)')}</td>
                        <td>{lev_data['distance']}</td>
                    </tr>
                    <tr>
-                        <td>Similarity ratio</td>
+                        <td>{_('Similarity ratio')}</td>
                        <td>{lev_data['ratio']:.4f}</td>
                    </tr>
                    <tr>
-                        <td>Percent similar</td>
+                        <td>{_('Percent similar')}</td>
                        <td>{lev_data['percent_similar']}%</td>
                    </tr>
                </tbody>
            </table>
-            <p style="font-size: 80%;">Levenshtein metrics compare the last two snapshots, measuring how many character edits are needed to transform one into the other.</p>
+            <p style="font-size: 80%;">{_('Levenshtein metrics compare the last two snapshots, measuring how many character edits are needed to transform one into the other.')}</p>
        </div>
        """
        return html
    except Exception as e:
        logger.error(f"Error generating Levenshtein UI extras: {str(e)}")
-        return "<p>Error calculating Levenshtein metrics</p>"
+        return f"<p>{_('Error calculating Levenshtein metrics')}</p>"
        
@@ -4,6 +4,7 @@ Provides word count metrics for snapshot content.
 """
 import pluggy
 from loguru import logger
+from flask_babel import gettext as _, lazy_gettext as _l

 # Support both plugin systems
 conditions_hookimpl = pluggy.HookimplMarker("changedetectionio_conditions")
@@ -40,7 +41,7 @@ def register_operator_choices():
 def register_field_choices():
    # Add a field that will be available in conditions
    return [
-        ("word_count", "Word count of content"),
+        ("word_count", _l("Word count of content")),
    ]

@conditions_hookimpl
@@ -61,16 +62,16 @@ def _generate_stats_html(watch):
    
    html = f"""
    <div class="word-count-stats">
-        <h4>Content Analysis</h4>
+        <h4>{_('Content Analysis')}</h4>
        <table class="pure-table">
            <tbody>
                <tr>
-                    <td>Word count (latest snapshot)</td>
+                    <td>{_('Word count (latest snapshot)')}</td>
                    <td>{word_count}</td>
                </tr>
            </tbody>
        </table>
-        <p style="font-size: 80%;">Word count is a simple measure of content length, calculated by splitting text on whitespace.</p>
+        <p style="font-size: 80%;">{_('Word count is a simple measure of content length, calculated by splitting text on whitespace.')}</p>
    </div>
    """
    return html
@@ -49,6 +49,9 @@ async def capture_full_page_async(page, screenshot_format='JPEG', watch_uuid=Non
    if page_height > page.viewport_size['height']:
        if page_height < step_size:
            step_size = page_height # Incase page is bigger than default viewport but smaller than proposed step size
+        # Never set viewport taller than our max capture height - otherwise one screenshot chunk
+        # captures the whole (e.g. 8098px) page even when SCREENSHOT_MAX_HEIGHT=1000
+        step_size = min(step_size, SCREENSHOT_MAX_TOTAL_HEIGHT)
        viewport_start = time.time()
        logger.debug(f"{watch_info}Setting bigger viewport to step through large page width W{page.viewport_size['width']}xH{step_size} because page_height > viewport_size")
        # Set viewport to a larger size to capture more content at once
@@ -75,6 +75,9 @@ async def capture_full_page(page, screenshot_format='JPEG', watch_uuid=None, loc
    if page_height > page.viewport['height']:
        if page_height < step_size:
            step_size = page_height # Incase page is bigger than default viewport but smaller than proposed step size
+        # Never set viewport taller than our max capture height - otherwise one screenshot chunk
+        # captures the whole page even when SCREENSHOT_MAX_HEIGHT is set smaller
+        step_size = min(step_size, SCREENSHOT_MAX_TOTAL_HEIGHT)
        viewport_start = time.time()
        await page.setViewport({'width': page.viewport['width'], 'height': step_size})
        viewport_time = time.time() - viewport_start
@@ -86,8 +89,8 @@ async def capture_full_page(page, screenshot_format='JPEG', watch_uuid=None, loc
        # better than scrollTo incase they override it in the page
        await page.evaluate(
            """(y) => {
-                document.documentElement.scrollTop = y;
-                document.body.scrollTop = y;
+                const el = document.scrollingElement;
+                if (el) el.scrollTop = y;
            }""",
            y
        )
@@ -305,6 +308,8 @@ class fetcher(Fetcher):
                await asyncio.wait_for(self.browser.close(), timeout=3.0)
            except Exception as cleanup_error:
                logger.error(f"[{watch_uuid}] Failed to cleanup browser after page creation failure: {cleanup_error}")
+            finally:
+                self.browser = None
            raise
        
        # Add console handler to capture console.log from favicon fetcher
@@ -532,6 +537,14 @@ class fetcher(Fetcher):
            )
        except asyncio.TimeoutError:
            raise (BrowserFetchTimedOut(msg=f"Browser connected but was unable to process the page in {max_time} seconds."))
+        finally:
+            # Internal cleanup on any exception/timeout - call quit() immediately
+            # This prevents connection leaks during exception bursts
+            # Worker.py's quit() call becomes a redundant safety net (idempotent)
+            try:
+                await self.quit(watch={'uuid': watch_uuid} if watch_uuid else None)
+            except Exception as cleanup_error:
+                logger.error(f"[{watch_uuid}] Error during internal quit() cleanup: {cleanup_error}")


 # Plugin registration for built-in fetcher
@@ -1,4 +1,6 @@
+from flask_babel import lazy_gettext as _l
 from loguru import logger
+from urllib.parse import urljoin, urlparse
 import hashlib
 import os
 import re
@@ -7,11 +9,12 @@ import asyncio
 from changedetectionio import strtobool
 from changedetectionio.content_fetchers.exceptions import BrowserStepsInUnsupportedFetcher, EmptyReply, Non200ErrorCodeReceived
 from changedetectionio.content_fetchers.base import Fetcher
+from changedetectionio.validate_url import is_private_hostname, is_url_private_or_parser_confused


 # "html_requests" is listed as the default fetcher in store.py!
 class fetcher(Fetcher):
-    fetcher_description = "Basic fast Plaintext/HTTP Client"
+    fetcher_description = _l("Basic fast Plaintext/HTTP Client")

    def __init__(self, proxy_override=None, custom_browser_connection_url=None, **kwargs):
        super().__init__(**kwargs)
@@ -79,14 +82,50 @@ class fetcher(Fetcher):
        if strtobool(os.getenv('ALLOW_FILE_URI', 'false')) and url.startswith('file://'):
            from requests_file import FileAdapter
            session.mount('file://', FileAdapter())
+
+        allow_iana_restricted = strtobool(os.getenv('ALLOW_IANA_RESTRICTED_ADDRESSES', 'false'))
+
        try:
+            # Fresh DNS check at fetch time — catches DNS rebinding regardless of add-time cache.
+            # Validates every hostname both urlparse and urllib3 see, so parser-differential
+            # payloads (GHSA-rph4-96w6-q594) cannot smuggle an internal target past the gate.
+            if not allow_iana_restricted:
+                if is_url_private_or_parser_confused(url):
+                    raise Exception(f"Fetch blocked: '{url}' resolves to a private/reserved IP address "
+                                    f"or contains a parser-differential payload. "
+                                    f"Set ALLOW_IANA_RESTRICTED_ADDRESSES=true to allow.")
+
            r = session.request(method=request_method,
                                data=request_body.encode('utf-8') if type(request_body) is str else request_body,
                                url=url,
                                headers=request_headers,
                                timeout=timeout,
                                proxies=proxies,
-                                verify=False)
+                                verify=False,
+                                allow_redirects=False)
+
+            # Manually follow redirects so each hop's resolved IP can be validated,
+            # preventing SSRF via an open redirect on a public host.
+            current_url = url
+            for _ in range(10):
+                if not r.is_redirect:
+                    break
+                location = r.headers.get('Location', '')
+                redirect_url = urljoin(current_url, location)
+                if not allow_iana_restricted:
+                    if is_url_private_or_parser_confused(redirect_url):
+                        raise Exception(f"Redirect blocked: '{redirect_url}' resolves to a private/reserved IP address "
+                                        f"or contains a parser-differential payload.")
+                current_url = redirect_url
+                r = session.request('GET', redirect_url,
+                                    headers=request_headers,
+                                    timeout=timeout,
+                                    proxies=proxies,
+                                    verify=False,
+                                    allow_redirects=False)
+            else:
+                raise Exception("Too many redirects")
+
        except Exception as e:
            msg = str(e)
            if proxies and 'SOCKSHTTPSConnectionPool' in msg:
@@ -112,10 +151,32 @@ class fetcher(Fetcher):
                        # Default to UTF-8 for XML if no encoding found
                        r.encoding = 'utf-8'
                else:
-                    # For other content types, use chardet
-                    encoding = chardet.detect(r.content)['encoding']
-                    if encoding:
-                        r.encoding = encoding
+                    # No charset in HTTP header - sniff encoding in priority order matching browsers
+                    # (WHATWG encoding sniffing algorithm):
+                    # 1. BOM - highest confidence, check before anything else
+                    # 2. <meta charset> in first 2kb
+                    # 3. chardet statistical detection - last resort
+                    # See: https://github.com/dgtlmoon/changedetection.io/issues/3952
+                    boms = [
+                        (b'\xef\xbb\xbf', 'utf-8-sig'),
+                        (b'\xff\xfe', 'utf-16-le'),
+                        (b'\xfe\xff', 'utf-16-be'),
+                    ]
+                    bom_encoding = next((enc for bom, enc in boms if r.content.startswith(bom)), None)
+                    if bom_encoding:
+                        logger.info(f"URL: {url} Using encoding '{bom_encoding}' detected from BOM")
+                        r.encoding = bom_encoding
+                    else:
+                        meta_charset_match = re.search(rb'<meta[^>]+charset\s*=\s*["\']?\s*([^"\'\s;>]+)', r.content[:2000], re.IGNORECASE)
+                        if meta_charset_match:
+                            encoding = meta_charset_match.group(1).decode('ascii', errors='ignore')
+                            logger.info(f"URL: {url} No content-type encoding in HTTP headers - Using encoding '{encoding}' from HTML meta charset tag")
+                            r.encoding = encoding
+                        else:
+                            encoding = chardet.detect(r.content)['encoding']
+                            logger.warning(f"URL: {url} No charset in headers or meta tag, guessed encoding as '{encoding}' via chardet")
+                            if encoding:
+                                r.encoding = encoding

        self.headers = r.headers

@@ -38,26 +38,39 @@
      if (a.size !== b.size) {
        return b.size - a.size;
      }
-      
+
      // Second priority: apple-touch-icon over regular icon
      const isAppleA = /apple-touch-icon/.test(a.rel);
      const isAppleB = /apple-touch-icon/.test(b.rel);
      if (isAppleA && !isAppleB) return -1;
      if (!isAppleA && isAppleB) return 1;
-      
+
      // Third priority: icons with no size attribute (fallback icons) last
      const hasNoSizeA = !a.hasSizes;
      const hasNoSizeB = !b.hasSizes;
      if (hasNoSizeA && !hasNoSizeB) return 1;
      if (!hasNoSizeA && hasNoSizeB) return -1;
-      
+
      return 0;
    });

    const timeoutMs = 2000;
+    // 1 MB — matches the server-side limit in bump_favicon()
+    const MAX_BYTES = 1 * 1024 * 1024;

    for (const icon of icons) {
      try {
+        // Inline data URI — no network fetch needed, data is already here
+        if (icon.href.startsWith('data:')) {
+          const match = icon.href.match(/^data:([^;]+);base64,([A-Za-z0-9+/=]+)$/);
+          if (!match) continue;
+          const mime_type = match[1];
+          const base64 = match[2];
+          // Rough size check: base64 is ~4/3 the binary size
+          if (base64.length * 0.75 > MAX_BYTES) continue;
+          return { url: icon.href, mime_type, base64 };
+        }
+
        const controller = new AbortController();
        const timeout = setTimeout(() => controller.abort(), timeoutMs);

@@ -74,12 +87,15 @@

        const blob = await resp.blob();

+        if (blob.size > MAX_BYTES) continue;
+
        // Convert blob to base64
        const reader = new FileReader();
        return await new Promise(resolve => {
          reader.onloadend = () => {
            resolve({
              url: icon.href,
+              mime_type: blob.type,
              base64: reader.result.split(",")[1]
            });
          };
@@ -98,4 +114,3 @@
  // Auto-execute and return result for page.evaluate()
  return await window.getFaviconAsBlob();
 })();
-
@@ -56,6 +56,10 @@ def stitch_images_worker_raw_bytes(pipe_conn, original_page_height, capture_heig
            im.close()
        del images

+        # Clip stitched image to capture_height (chunks may overshoot by up to step_size-1 px)
+        if total_height > capture_height:
+            stitched = stitched.crop((0, 0, max_width, capture_height))
+
        # Draw caption only if page was trimmed
        if original_page_height > capture_height:
            draw = ImageDraw.Draw(stitched)
@@ -104,15 +104,17 @@ class fetcher(Fetcher):

            from selenium.webdriver.remote.remote_connection import RemoteConnection
            from selenium.webdriver.remote.webdriver import WebDriver as RemoteWebDriver
+            from selenium.webdriver.remote.client_config import ClientConfig
+            from urllib3.util import Timeout
            driver = None
            try:
-                # Create the RemoteConnection and set timeout (e.g., 30 seconds)
-                remote_connection = RemoteConnection(
-                    self.browser_connection_url,
+                connection_timeout = int(os.getenv("WEBDRIVER_CONNECTION_TIMEOUT", 90))
+                client_config = ClientConfig(
+                    remote_server_addr=self.browser_connection_url,
+                    timeout=Timeout(connect=connection_timeout, total=connection_timeout)
                )
-                remote_connection.set_timeout(30)  # seconds
+                remote_connection = RemoteConnection(client_config=client_config)

-                # Now create the driver with the RemoteConnection
                driver = RemoteWebDriver(
                    command_executor=remote_connection,
                    options=options
@@ -45,8 +45,38 @@ CHANGED_INTO_PLACEMARKER_CLOSED = '@changed_into_PLACEMARKER_CLOSED'
 # Compiled regex patterns for performance
 WHITESPACE_NORMALIZE_RE = re.compile(r'\s+')

+# Regexes built from the constants above — no brittle hardcoded strings
+_EXTRACT_REMOVED_RE = re.compile(
+    re.escape(REMOVED_PLACEMARKER_OPEN) + r'(.*?)' + re.escape(REMOVED_PLACEMARKER_CLOSED)
+    + r'|' +
+    re.escape(CHANGED_PLACEMARKER_OPEN) + r'(.*?)' + re.escape(CHANGED_PLACEMARKER_CLOSED)
+)
+_EXTRACT_ADDED_RE = re.compile(
+    re.escape(ADDED_PLACEMARKER_OPEN) + r'(.*?)' + re.escape(ADDED_PLACEMARKER_CLOSED)
+    + r'|' +
+    re.escape(CHANGED_INTO_PLACEMARKER_OPEN) + r'(.*?)' + re.escape(CHANGED_INTO_PLACEMARKER_CLOSED)
+)

-def render_inline_word_diff(before_line: str, after_line: str, ignore_junk: bool = False, markdown_style: str = None, tokenizer: str = 'words_and_html') -> tuple[str, bool]:
+
+def extract_changed_from(raw_diff: str) -> str:
+    """Extract only the removed/changed-from fragments from a raw diff string.
+
+    Useful for {{diff_changed_from}} — gives just the old value (e.g. old price),
+    not the full surrounding line. Multiple fragments joined with newlines.
+    """
+    return '\n'.join(next((g for g in m.groups() if g is not None), '') for m in _EXTRACT_REMOVED_RE.finditer(raw_diff))
+
+
+def extract_changed_to(raw_diff: str) -> str:
+    """Extract only the added/changed-into fragments from a raw diff string.
+
+    Useful for {{diff_changed_to}} — gives just the new value (e.g. new price),
+    not the full surrounding line. Multiple fragments joined with newlines.
+    """
+    return '\n'.join(next((g for g in m.groups() if g is not None), '') for m in _EXTRACT_ADDED_RE.finditer(raw_diff))
+
+
+def render_inline_word_diff(before_line: str, after_line: str, ignore_junk: bool = False, markdown_style: str = None, tokenizer: str = 'words_and_html', include_change_type_prefix: bool = True) -> tuple[str, bool]:
    """
    Render word-level differences between two lines inline using diff-match-patch library.

@@ -133,14 +163,20 @@ def render_inline_word_diff(before_line: str, after_line: str, ignore_junk: bool
        if removed_tokens:
            removed_full = ''.join(removed_tokens).rstrip()
            trailing_removed = ''.join(removed_tokens)[len(removed_full):] if len(''.join(removed_tokens)) > len(removed_full) else ''
-            result_parts.append(f'{CHANGED_PLACEMARKER_OPEN}{removed_full}{CHANGED_PLACEMARKER_CLOSED}{trailing_removed}')
+            if include_change_type_prefix:
+                result_parts.append(f'{CHANGED_PLACEMARKER_OPEN}{removed_full}{CHANGED_PLACEMARKER_CLOSED}{trailing_removed}')
+            else:
+                result_parts.append(f'{removed_full}{trailing_removed}')

        if added_tokens:
            if result_parts:  # Add newline between removed and added
                result_parts.append('\n')
            added_full = ''.join(added_tokens).rstrip()
            trailing_added = ''.join(added_tokens)[len(added_full):] if len(''.join(added_tokens)) > len(added_full) else ''
-            result_parts.append(f'{CHANGED_INTO_PLACEMARKER_OPEN}{added_full}{CHANGED_INTO_PLACEMARKER_CLOSED}{trailing_added}')
+            if include_change_type_prefix:
+                result_parts.append(f'{CHANGED_INTO_PLACEMARKER_OPEN}{added_full}{CHANGED_INTO_PLACEMARKER_CLOSED}{trailing_added}')
+            else:
+                result_parts.append(f'{added_full}{trailing_added}')

        return ''.join(result_parts), has_changes
    else:
@@ -150,21 +186,27 @@ def render_inline_word_diff(before_line: str, after_line: str, ignore_junk: bool
            if op == 0:  # Equal
                result_parts.append(text)
            elif op == 1:  # Insertion
-                # Don't wrap empty content (e.g., whitespace-only tokens after rstrip)
-                content = text.rstrip()
-                trailing = text[len(content):] if len(text) > len(content) else ''
-                if content:
-                    result_parts.append(f'{ADDED_PLACEMARKER_OPEN}{content}{ADDED_PLACEMARKER_CLOSED}{trailing}')
+                if not include_change_type_prefix:
+                    result_parts.append(text)
                else:
-                    result_parts.append(trailing)
+                    # Don't wrap empty content (e.g., whitespace-only tokens after rstrip)
+                    content = text.rstrip()
+                    trailing = text[len(content):] if len(text) > len(content) else ''
+                    if content:
+                        result_parts.append(f'{ADDED_PLACEMARKER_OPEN}{content}{ADDED_PLACEMARKER_CLOSED}{trailing}')
+                    else:
+                        result_parts.append(trailing)
            elif op == -1:  # Deletion
-                # Don't wrap empty content (e.g., whitespace-only tokens after rstrip)
-                content = text.rstrip()
-                trailing = text[len(content):] if len(text) > len(content) else ''
-                if content:
-                    result_parts.append(f'{REMOVED_PLACEMARKER_OPEN}{content}{REMOVED_PLACEMARKER_CLOSED}{trailing}')
+                if not include_change_type_prefix:
+                    result_parts.append(text)
                else:
-                    result_parts.append(trailing)
+                    # Don't wrap empty content (e.g., whitespace-only tokens after rstrip)
+                    content = text.rstrip()
+                    trailing = text[len(content):] if len(text) > len(content) else ''
+                    if content:
+                        result_parts.append(f'{REMOVED_PLACEMARKER_OPEN}{content}{REMOVED_PLACEMARKER_CLOSED}{trailing}')
+                    else:
+                        result_parts.append(trailing)

        return ''.join(result_parts), has_changes

@@ -360,7 +402,7 @@ def customSequenceMatcher(

            # Use inline word-level diff for single line replacements when word_diff is enabled
            if word_diff and len(before_lines) == 1 and len(after_lines) == 1:
-                inline_diff, has_changes = render_inline_word_diff(before_lines[0], after_lines[0], ignore_junk=ignore_junk, tokenizer=tokenizer)
+                inline_diff, has_changes = render_inline_word_diff(before_lines[0], after_lines[0], ignore_junk=ignore_junk, tokenizer=tokenizer, include_change_type_prefix=include_change_type_prefix)
                # Check if there are any actual changes (not just whitespace when ignore_junk is enabled)
                if ignore_junk and not has_changes:
                    # No real changes, skip this line
@@ -415,8 +457,8 @@ def render_diff(
    Returns:
        str: Rendered difference
    """
-    newest_lines = [line.rstrip() for line in newest_version_file_contents.splitlines()]
-    previous_lines = [line.rstrip() for line in previous_version_file_contents.splitlines()] if previous_version_file_contents else []
+    newest_lines = [line.rstrip() for line in (newest_version_file_contents or '').splitlines()]
+    previous_lines = [line.rstrip() for line in (previous_version_file_contents or '').splitlines()]
    now = time.time()
    logger.debug(
        f"diff options: "
@@ -4,6 +4,7 @@ import flask_login
 import locale
 import os
 import queue
+import re
 import sys
 import threading
 import time
@@ -27,7 +28,6 @@ from flask import (
    session,
    url_for,
 )
-from flask_compress import Compress as FlaskCompress
 from flask_restful import abort, Api
 from flask_cors import CORS

@@ -40,7 +40,7 @@ from loguru import logger

 from changedetectionio import __version__
 from changedetectionio import queuedWatchMetaData
-from changedetectionio.api import Watch, WatchHistory, WatchSingleHistory, WatchHistoryDiff, CreateWatch, Import, SystemInfo, Tag, Tags, Notifications, WatchFavicon
+from changedetectionio.api import Watch, WatchHistory, WatchSingleHistory, WatchHistoryDiff, CreateWatch, Import, SystemInfo, Tag, Tags, Notifications, WatchFavicon, Spec
 from changedetectionio.api.Search import Search
 from .time_handler import is_within_schedule
 from changedetectionio.languages import get_available_languages, get_language_codes, get_flag_for_locale, get_timeago_locale
@@ -69,19 +69,43 @@ socketio_server = None

 # Enable CORS, especially useful for the Chrome extension to operate from anywhere
 CORS(app)
+from werkzeug.routing import BaseConverter, ValidationError
+from uuid import UUID
+
+class StrictUUIDConverter(BaseConverter):
+    # Special sentinel values allowed in addition to strict UUIDs
+    _ALLOWED_SENTINELS = frozenset({'first'})
+
+    def to_python(self, value: str) -> str:
+        if value in self._ALLOWED_SENTINELS:
+            return value
+        try:
+            u = UUID(value)
+        except ValueError as e:
+            raise ValidationError() from e
+        # Reject non-standard formats (braces, URNs, no-hyphens)
+        if str(u) != value.lower():
+            raise ValidationError()
+        return str(u)
+
+    def to_url(self, value) -> str:
+        return str(value)
+
+# app setup (once)
+app.url_map.converters["uuid_str"] = StrictUUIDConverter

 # Flask-Compress handles HTTP compression, Socket.IO compression disabled to prevent memory leak.
 # There's also a bug between flask compress and socketio that causes some kind of slow memory leak
 # It's better to use compression on your reverse proxy (nginx etc) instead.
 if strtobool(os.getenv("FLASK_ENABLE_COMPRESSION")):
+    from flask_compress import Compress as FlaskCompress
    app.config['COMPRESS_MIN_SIZE'] = 2096
    app.config['COMPRESS_MIMETYPES'] = ['text/html', 'text/css', 'text/javascript', 'application/json', 'application/javascript', 'image/svg+xml']
    # Use gzip only - smaller memory footprint than zstd/brotli (4-8KB vs 200-500KB contexts)
    app.config['COMPRESS_ALGORITHM'] = ['gzip']
+    compress = FlaskCompress()
+    compress.init_app(app)

-compress = FlaskCompress()
-
-compress.init_app(app)
 app.config['TEMPLATES_AUTO_RELOAD'] = False


@@ -188,15 +212,24 @@ def _is_safe_valid_url(test_url):
    from .validate_url import is_safe_valid_url
    return is_safe_valid_url(test_url)

+@app.template_global('get_html_head_extras')
+def _get_html_head_extras():
+    from .pluggy_interface import collect_html_head_extras
+    return collect_html_head_extras()
+

@app.template_filter('format_number_locale')
 def _jinja2_filter_format_number_locale(value: float) -> str:
    "Formats for example 4000.10 to the local locale default of 4,000.10"
    # Format the number with two decimal places (locale format string will return 6 decimal)
    formatted_value = locale.format_string("%.2f", value, grouping=True)
-
    return formatted_value

+@app.template_filter('regex_search')
+def _jinja2_filter_regex_search(value, pattern):
+    import re
+    return re.search(pattern, str(value)) is not None
+
@app.template_global('is_checking_now')
 def _watch_is_checking_now(watch_obj, format="%Y-%m-%d %H:%M:%S"):
    return worker_pool.is_watch_running(watch_obj['uuid'])
@@ -360,6 +393,8 @@ def _jinja2_filter_fetcher_status_icons(fetcher_name):

    return ''

+_RE_SANITIZE_TAG = re.compile(r'[^a-zA-Z0-9]')
+
@app.template_filter('sanitize_tag_class')
 def _jinja2_filter_sanitize_tag_class(tag_title):
    """Sanitize a tag title to create a valid CSS class name.
@@ -371,16 +406,15 @@ def _jinja2_filter_sanitize_tag_class(tag_title):
    Returns:
        str: A sanitized string suitable for use as a CSS class name
    """
-    import re
    # Remove all non-alphanumeric characters and convert to lowercase
-    sanitized = re.sub(r'[^a-zA-Z0-9]', '', tag_title).lower()
+    sanitized = _RE_SANITIZE_TAG.sub('', tag_title).lower()
    # Ensure it starts with a letter (CSS requirement)
    if sanitized and not sanitized[0].isalpha():
        sanitized = 'tag' + sanitized
    return sanitized if sanitized else 'tag'

 # Import login_optionally_required from auth_decorator
-from changedetectionio.auth_decorator import login_optionally_required
+from changedetectionio.auth_decorator import SHARED_DIFF_READ_ONLY_ENDPOINTS, login_optionally_required

 # When nobody is logged in Flask-Login's current_user is set to an AnonymousUser object.
 class User(flask_login.UserMixin):
@@ -461,28 +495,21 @@ def changedetection_app(config=None, datastore_o=None):
    available_languages = get_available_languages()
    language_codes = get_language_codes()

-    def get_locale():
-        # Locale aliases: map browser language codes to translation directory names
-        # This handles cases where browsers send standard codes (e.g., zh-TW)
-        # but our translations use more specific codes (e.g., zh_Hant_TW)
-        locale_aliases = {
-            'zh-TW': 'zh_Hant_TW',  # Traditional Chinese: browser sends zh-TW, we use zh_Hant_TW
-            'zh_TW': 'zh_Hant_TW',  # Also handle underscore variant
-        }
+    _locale_aliases = {
+        'zh-TW': 'zh_Hant_TW',  # Traditional Chinese: browser sends zh-TW, we use zh_Hant_TW
+        'zh_TW': 'zh_Hant_TW',  # Also handle underscore variant
+    }
+    _locale_match_list = language_codes + list(_locale_aliases.keys())

+    def get_locale():
        # 1. Try to get locale from session (user explicitly selected)
        if 'locale' in session:
            return session['locale']

        # 2. Fall back to Accept-Language header
-        # Get the best match from browser's Accept-Language header
-        browser_locale = request.accept_languages.best_match(language_codes + list(locale_aliases.keys()))
-
-        # 3. Check if we need to map the browser locale to our internal locale
-        if browser_locale in locale_aliases:
-            return locale_aliases[browser_locale]
-
-        return browser_locale
+        browser_locale = request.accept_languages.best_match(_locale_match_list)
+        # 3. Map browser locale to our internal locale if needed
+        return _locale_aliases.get(browser_locale, browser_locale)

    # Initialize Babel with locale selector
    babel = Babel(app, locale_selector=get_locale)
@@ -495,6 +522,11 @@ def changedetection_app(config=None, datastore_o=None):
        available_languages=available_languages
    )

+    @app.context_processor
+    def inject_llm_features_disabled():
+        from changedetectionio.llm.evaluator import is_llm_features_disabled
+        return dict(llm_features_disabled=is_llm_features_disabled())
+
    # Set up a request hook to check authentication for all routes
    @app.before_request
    def check_authentication():
@@ -514,7 +546,7 @@ def changedetection_app(config=None, datastore_o=None):
            # Permitted
            elif request.endpoint and 'login' in request.endpoint:
                return None
-            elif request.endpoint and 'diff_history_page' in request.endpoint and datastore.data['settings']['application'].get('shared_diff_access'):
+            elif request.endpoint in SHARED_DIFF_READ_ONLY_ENDPOINTS and datastore.data['settings']['application'].get('shared_diff_access'):
                return None
            elif request.method in flask_login.config.EXEMPT_METHODS:
                return None
@@ -534,22 +566,22 @@ def changedetection_app(config=None, datastore_o=None):


    watch_api.add_resource(WatchHistoryDiff,
-                           '/api/v1/watch/<string:uuid>/difference/<string:from_timestamp>/<string:to_timestamp>',
+                           '/api/v1/watch/<uuid_str:uuid>/difference/<string:from_timestamp>/<string:to_timestamp>',
                           resource_class_kwargs={'datastore': datastore})
    watch_api.add_resource(WatchSingleHistory,
-                           '/api/v1/watch/<string:uuid>/history/<string:timestamp>',
+                           '/api/v1/watch/<uuid_str:uuid>/history/<string:timestamp>',
                           resource_class_kwargs={'datastore': datastore, 'update_q': update_q})
    watch_api.add_resource(WatchFavicon,
-                           '/api/v1/watch/<string:uuid>/favicon',
+                           '/api/v1/watch/<uuid_str:uuid>/favicon',
                           resource_class_kwargs={'datastore': datastore})
    watch_api.add_resource(WatchHistory,
-                           '/api/v1/watch/<string:uuid>/history',
+                           '/api/v1/watch/<uuid_str:uuid>/history',
                           resource_class_kwargs={'datastore': datastore})

    watch_api.add_resource(CreateWatch, '/api/v1/watch',
                           resource_class_kwargs={'datastore': datastore, 'update_q': update_q})

-    watch_api.add_resource(Watch, '/api/v1/watch/<string:uuid>',
+    watch_api.add_resource(Watch, '/api/v1/watch/<uuid_str:uuid>',
                           resource_class_kwargs={'datastore': datastore, 'update_q': update_q})

    watch_api.add_resource(SystemInfo, '/api/v1/systeminfo',
@@ -562,7 +594,7 @@ def changedetection_app(config=None, datastore_o=None):
    watch_api.add_resource(Tags, '/api/v1/tags',
                           resource_class_kwargs={'datastore': datastore})

-    watch_api.add_resource(Tag, '/api/v1/tag', '/api/v1/tag/<string:uuid>',
+    watch_api.add_resource(Tag, '/api/v1/tag', '/api/v1/tag/<uuid_str:uuid>',
                           resource_class_kwargs={'datastore': datastore, 'update_q': update_q})
                           
    watch_api.add_resource(Search, '/api/v1/search',
@@ -571,6 +603,8 @@ def changedetection_app(config=None, datastore_o=None):
    watch_api.add_resource(Notifications, '/api/v1/notifications',
                           resource_class_kwargs={'datastore': datastore})

+    watch_api.add_resource(Spec, '/api/v1/full-spec')
+
    @login_manager.user_loader
    def user_loader(email):
        user = User()
@@ -712,8 +746,14 @@ def changedetection_app(config=None, datastore_o=None):
    def static_content(group, filename):
        from flask import make_response
        import re
-        group = re.sub(r'[^\w.-]+', '', group.lower())
-        filename = re.sub(r'[^\w.-]+', '', filename.lower())
+
+        # Strict sanitization: only allow a-z, 0-9, and underscore (blocks .. and other traversal)
+        group = re.sub(r'[^a-z0-9_-]+', '', group.lower())
+        filename = filename
+
+        # Additional safety: reject if sanitization resulted in empty strings
+        if not group or not filename:
+            abort(404)

        if group == 'screenshot':
            # Could be sensitive, follow password requirements
@@ -946,6 +986,11 @@ def changedetection_app(config=None, datastore_o=None):
                    "queued_data": all_queued
                })

+    if strtobool(os.getenv('HISTORY_SNAPSHOT_FILE_ALLOW_OUTSIDE_WATCH_DATADIR', 'False')):
+        logger.warning("SECURITY WARNING: HISTORY_SNAPSHOT_FILE_ALLOW_OUTSIDE_WATCH_DATADIR is enabled — "
+                       "snapshot reads are NOT confined to the watch data directory. "
+                       "This disables protection against path traversal via restored backups (GHSA-8757-69j2-hx56).")
+
    # Start the async workers during app initialization
    # Can be overridden by ENV or use the default settings
    n_workers = int(os.getenv("FETCH_WORKERS", datastore.data['settings']['requests']['workers']))
@@ -987,15 +1032,16 @@ def check_for_new_version():
    import urllib3
    urllib3.disable_warnings(urllib3.exceptions.InsecureRequestWarning)

+    session = requests.Session()
+    session.verify = False
+
    while not app.config.exit.is_set():
        try:
-            r = requests.post("https://changedetection.io/check-ver.php",
+            r = session.post("https://changedetection.io/check-ver.php",
                              data={'version': __version__,
                                    'app_guid': datastore.data['app_guid'],
                                    'watch_count': len(datastore.data['watching'])
-                                    },
-
-                              verify=False)
+                                    })
        except:
            pass

@@ -5,6 +5,8 @@ from wtforms.widgets.core import TimeInput
 from flask_babel import lazy_gettext as _l, gettext

 from changedetectionio.blueprint.rss import RSS_FORMAT_TYPES, RSS_TEMPLATE_TYPE_OPTIONS, RSS_TEMPLATE_HTML_DEFAULT
+from changedetectionio.llm.ui_strings import LLM_INTENT_WATCH_PLACEHOLDER
+from changedetectionio.llm.evaluator import DEFAULT_CHANGE_SUMMARY_PROMPT, LLM_DEFAULT_MAX_SUMMARY_TOKENS, LLM_DEFAULT_THINKING_BUDGET
 from changedetectionio.conditions.form import ConditionFormRow
 from changedetectionio.notification_service import NotificationContextData
 from changedetectionio.strtobool import strtobool
@@ -15,7 +17,9 @@ from wtforms import (
    Form,
    Field,
    FloatField,
+    HiddenField,
    IntegerField,
+    PasswordField,
    RadioField,
    SelectField,
    StringField,
@@ -276,12 +280,44 @@ class TimeBetweenCheckForm(Form):
        return True


+class LabelAfterInputTableWidget(widgets.TableWidget):
+    """
+    Variant of WTForms' TableWidget that renders the input cell before the label cell,
+    so each row is <td>input</td><th>label</th> instead of the default <th>label</th><td>input</td>.
+    """
+
+    def __call__(self, field, **kwargs):
+        from markupsafe import Markup
+        from wtforms.widgets import html_params
+
+        html = []
+        if self.with_table_tag:
+            kwargs.setdefault("id", field.id)
+            html.append(f"<table {html_params(**kwargs)}>")
+        hidden = ""
+        for subfield in field:
+            if subfield.type in ("HiddenField", "CSRFTokenField"):
+                hidden += str(subfield)
+            else:
+                html.append(
+                    f"<tr><td>{hidden}{subfield}</td><th>{subfield.label}</th></tr>"
+                )
+                hidden = ""
+        if self.with_table_tag:
+            html.append("</table>")
+        if hidden:
+            html.append(hidden)
+        return Markup("".join(html))
+
+
 class EnhancedFormField(FormField):
    """
    An enhanced FormField that supports conditional validation with top-level error messages.
    Adds a 'top_errors' property for validation errors at the FormField level.
    """

+    widget = LabelAfterInputTableWidget()
+
    def __init__(self, form_class, label=None, validators=None, separator="-",
                 conditional_field=None, conditional_message=None, conditional_test_function=None, **kwargs):
        """
@@ -548,6 +584,17 @@ def validate_url(test_url):
        raise ValidationError('Watch protocol is not permitted or invalid URL format')


+class validateLLMApiBaseSafe(object):
+    """Block private/loopback/reserved api_base values (SSRF) unless the operator
+    has opted in via ALLOW_IANA_RESTRICTED_ADDRESSES=true."""
+
+    def __call__(self, form, field):
+        from changedetectionio.validate_url import is_llm_api_base_safe
+        ok, reason = is_llm_api_base_safe(field.data)
+        if not ok:
+            raise ValidationError(reason)
+
+
 class ValidateSinglePythonRegexString(object):
    def __init__(self, message=None):
        self.message = message
@@ -608,16 +655,15 @@ class ValidateCSSJSONXPATHInput(object):
                    raise ValidationError("XPath not permitted in this field!")
                from lxml import etree, html
                import elementpath
-                # xpath 2.0-3.1
-                from elementpath.xpath3 import XPath3Parser
+                from changedetectionio.html_tools import SafeXPath3Parser
                tree = html.fromstring("<html></html>")
                line = line.replace('xpath:', '')

                try:
-                    elementpath.select(tree, line.strip(), parser=XPath3Parser)
+                    elementpath.select(tree, line.strip(), parser=SafeXPath3Parser)
                except elementpath.ElementPathError as e:
-                    message = field.gettext('\'%s\' is not a valid XPath expression. (%s)')
-                    raise ValidationError(message % (line, str(e)))
+                    message = field.gettext('\'%(expression)s\' is not a valid XPath expression. (%(error)s)')
+                    raise ValidationError(message % {'expression': line, 'error': str(e)})
                except:
                    raise ValidationError("A system-error occurred when validating your XPath expression")

@@ -631,8 +677,8 @@ class ValidateCSSJSONXPATHInput(object):
                try:
                    tree.xpath(line.strip())
                except etree.XPathEvalError as e:
-                    message = field.gettext('\'%s\' is not a valid XPath expression. (%s)')
-                    raise ValidationError(message % (line, str(e)))
+                    message = field.gettext('\'%(expression)s\' is not a valid XPath expression. (%(error)s)')
+                    raise ValidationError(message % {'expression': line, 'error': str(e)})
                except:
                    raise ValidationError("A system-error occurred when validating your XPath expression")

@@ -651,8 +697,8 @@ class ValidateCSSJSONXPATHInput(object):
                try:
                    parse(input)
                except (JsonPathParserError, JsonPathLexerError) as e:
-                    message = field.gettext('\'%s\' is not a valid JSONPath expression. (%s)')
-                    raise ValidationError(message % (input, str(e)))
+                    message = field.gettext('\'%(expression)s\' is not a valid JSONPath expression. (%(error)s)')
+                    raise ValidationError(message % {'expression': input, 'error': str(e)})
                except:
                    raise ValidationError("A system-error occurred when validating your JSONPath expression")

@@ -668,13 +714,15 @@ class ValidateCSSJSONXPATHInput(object):
                    # `jq` requires full compilation in windows and so isn't generally available
                    raise ValidationError("jq not support not found")

+                from changedetectionio.html_tools import validate_jq_expression
                input = line.replace('jq:', '')

                try:
+                    validate_jq_expression(input)
                    jq.compile(input)
                except (ValueError) as e:
-                    message = field.gettext('\'%s\' is not a valid jq expression. (%s)')
-                    raise ValidationError(message % (input, str(e)))
+                    message = field.gettext('\'%(expression)s\' is not a valid jq expression. (%(error)s)')
+                    raise ValidationError(message % {'expression': input, 'error': str(e)})
                except:
                    raise ValidationError("A system-error occurred when validating your jq expression")

@@ -724,7 +772,7 @@ class ValidateStartsWithRegex(object):
                raise ValidationError(self.message or _l("Invalid value."))

 class quickWatchForm(Form):
-    url = fields.URLField(_l('URL'), validators=[validateURL()])
+    url = StringField('URL', validators=[validateURL()])
    tags = StringTagUUID(_l('Group tag'), validators=[validators.Optional()])
    watch_submit_button = SubmitField(_l('Watch'), render_kw={"class": "pure-button pure-button-primary"})
    processor = RadioField(_l('Processor'), choices=lambda: processors.available_processors(), default=processors.get_default_processor)
@@ -770,16 +818,16 @@ class SingleBrowserStep(Form):
    operation = SelectField(_l('Operation'), [validators.Optional()], choices=browser_step_ui_config.keys())

    # maybe better to set some <script>var..
-    selector = StringField(_l('Selector'), [validators.Optional()], render_kw={"placeholder": "CSS or xPath selector"})
-    optional_value = StringField(_l('value'), [validators.Optional()], render_kw={"placeholder": "Value"})
+    selector = StringField(_l('Selector'), [validators.Optional()], render_kw={"placeholder": _l("CSS or xPath selector")})
+    optional_value = StringField(_l('value'), [validators.Optional()], render_kw={"placeholder": _l("Value")})
 #   @todo move to JS? ajax fetch new field?
 #    remove_button = SubmitField(_l('-'), render_kw={"type": "button", "class": "pure-button pure-button-primary", 'title': 'Remove'})
 #    add_button = SubmitField(_l('+'), render_kw={"type": "button", "class": "pure-button pure-button-primary", 'title': 'Add new step after'})

 class processor_text_json_diff_form(commonSettingsForm):

-    url = fields.URLField('Web Page URL', validators=[validateURL()])
-    tags = StringTagUUID('Group Tag', [validators.Optional()], default='')
+    url = StringField(_l('Web Page URL'), validators=[validateURL()])
+    tags = StringTagUUID(_l('Group Tag'), [validators.Optional()], default='')

    time_between_check = EnhancedFormField(
        TimeBetweenCheckForm,
@@ -793,10 +841,18 @@ class processor_text_json_diff_form(commonSettingsForm):

    time_between_check_use_default = BooleanField(_l('Use global settings for time between check and scheduler.'), default=False)

+    llm_intent = TextAreaField(_l('AI Change Intent'), validators=[validators.Optional(), validators.Length(max=2000)],
+                               render_kw={"rows": "5", "placeholder": LLM_INTENT_WATCH_PLACEHOLDER})
+
+    llm_change_summary = TextAreaField(_l('AI Change Summary'), validators=[validators.Optional(), validators.Length(max=2000)],
+                               render_kw={"rows": "5", "placeholder": DEFAULT_CHANGE_SUMMARY_PROMPT},
+                               default='')
+
    include_filters = StringListField(_l('CSS/JSONPath/JQ/XPath Filters'), [ValidateCSSJSONXPATHInput()], default='')

    subtractive_selectors = StringListField(_l('Remove elements'), [ValidateCSSJSONXPATHInput(allow_json=False)])

+    extract_lines_containing = StringListField(_l('Extract lines containing'), [validators.Optional()])
    extract_text = StringListField(_l('Extract text'), [ValidateListRegex()])

    title = StringField(_l('Title'), default='')
@@ -916,7 +972,7 @@ class processor_text_json_diff_form(commonSettingsForm):

 class SingleExtraProxy(Form):
    # maybe better to set some <script>var..
-    proxy_name = StringField(_l('Name'), [validators.Optional()], render_kw={"placeholder": "Name"})
+    proxy_name = StringField(_l('Name'), [validators.Optional()], render_kw={"placeholder": _l("Name")})
    proxy_url = StringField(_l('Proxy URL'), [
        validators.Optional(),
        ValidateStartsWithRegex(
@@ -928,7 +984,7 @@ class SingleExtraProxy(Form):
    ], render_kw={"placeholder": "socks5:// or regular proxy http://user:pass@...:3128", "size":50})

 class SingleExtraBrowser(Form):
-    browser_name = StringField(_l('Name'), [validators.Optional()], render_kw={"placeholder": "Name"})
+    browser_name = StringField(_l('Name'), [validators.Optional()], render_kw={"placeholder": _l("Name")})
    browser_connection_url = StringField(_l('Browser connection URL'), [
        validators.Optional(),
        ValidateStartsWithRegex(
@@ -987,7 +1043,7 @@ class globalSettingsApplicationForm(commonSettingsForm):
    api_access_token_enabled = BooleanField(_l('API access token security check enabled'), default=True, validators=[validators.Optional()])
    base_url = StringField(_l('Notification base URL override'),
                           validators=[validators.Optional()],
-                           render_kw={"placeholder": os.getenv('BASE_URL', 'Not set')}
+                           render_kw={"placeholder": os.getenv('BASE_URL', _l('Not set'))}
                           )
    empty_pages_are_a_change =  BooleanField(_l('Treat empty pages as a change?'), default=False)
    fetch_backend = RadioField(_l('Fetch Method'), default="html_requests", choices=content_fetchers.available_fetchers(), validators=[ValidateContentFetcherIsReady()])
@@ -997,7 +1053,7 @@ class globalSettingsApplicationForm(commonSettingsForm):

    # Screenshot comparison settings
    min_change_percentage = FloatField(
-        'Screenshot: Minimum Change Percentage',
+        _l('Screenshot: Minimum Change Percentage'),
        validators=[
            validators.Optional(),
            validators.NumberRange(min=0.0, max=100.0, message=_l('Must be between 0 and 100'))
@@ -1006,7 +1062,7 @@ class globalSettingsApplicationForm(commonSettingsForm):
        render_kw={"placeholder": "0.1", "style": "width: 8em;"}
    )

-    password = SaltyPasswordField(_l('Password'))
+    password = SaltyPasswordField(_l('Password'), render_kw={"autocomplete": "new-password"})
    pager_size = IntegerField(_l('Pager size'),
                              render_kw={"style": "width: 5em;"},
                              validators=[validators.NumberRange(min=0,
@@ -1037,6 +1093,149 @@ class globalSettingsApplicationForm(commonSettingsForm):
    ui = FormField(globalSettingsApplicationUIForm)


+class globalSettingsLLMForm(Form):
+    """
+    LLM / AI provider settings — stored under datastore['settings']['application']['llm'].
+
+    Uses litellm under the hood, so the model string encodes both the provider and model.
+    No separate provider dropdown needed — litellm routes automatically:
+      gpt-4o-mini                           → OpenAI
+      claude-3-5-haiku-20251001             → Anthropic
+      ollama/llama3.2                       → Ollama
+      openrouter/google/gemma-3-12b-it:free → OpenRouter (free tier)
+      gemini/gemini-2.0-flash               → Google Gemini
+      azure/gpt-4o                          → Azure OpenAI
+    """
+    model = StringField(
+        _l('Model'),
+        validators=[validators.Optional()],
+        render_kw={"placeholder": "gpt-4o-mini", "style": "width: 24em;"},
+    )
+    api_key = PasswordField(
+        _l('API Key'),
+        validators=[validators.Optional()],
+        render_kw={
+            "autocomplete": "off",
+            "style": "width: 24em;",
+        },
+    )
+    api_base = StringField(
+        _l('API Base URL'),
+        validators=[validators.Optional(), validateLLMApiBaseSafe()],
+        render_kw={
+            "placeholder": "http://localhost:11434  (Ollama / custom endpoints only)",
+            "style": "width: 24em;",
+        },
+    )
+    # Persisted by the Provider dropdown JS — lets the backend distinguish a self-hosted
+    # OpenAI-compatible endpoint (vLLM, LM Studio, llama.cpp) from cloud OpenAI, so we can
+    # apply reasoning-friendly token caps only when the user opted in.
+    provider_kind = HiddenField(
+        validators=[validators.Optional()],
+        default='',
+    )
+    # Multiplier applied to LLM max_tokens caps when provider_kind is 'ollama' or
+    # 'openai_compatible' — endpoints that commonly serve reasoning models (Qwen3,
+    # DeepSeek-R1, Gemma 3, etc.) which emit chain-of-thought into
+    # message.reasoning_content before the final answer lands in message.content.
+    # Cloud providers with non-reasoning defaults (OpenAI, Anthropic, Gemini,
+    # OpenRouter) stay on the original tight caps so existing users see no
+    # behavior or cost change. Users on paid Ollama / openai_compatible endpoints
+    # who care about cost can dial this down to 1x.
+    local_token_multiplier = IntegerField(
+        _l('Token multiplier for local reasoning models'),
+        validators=[validators.Optional(), validators.NumberRange(min=1, max=20)],
+        default=5,
+        render_kw={"placeholder": "5", "style": "width: 6em;"},
+    )
+    change_summary_default = TextAreaField(
+        _l('Default AI Change Summary prompt'),
+        validators=[validators.Optional(), validators.Length(max=2000)],
+        render_kw={
+            "rows": "5",
+            "placeholder": DEFAULT_CHANGE_SUMMARY_PROMPT,
+            "style": "width: 100%; ",
+        },
+        default='',
+    )
+    max_tokens_per_count_period = IntegerField(
+        _l('Max tokens per watch per period'),
+        validators=[validators.Optional(), validators.NumberRange(min=0)],
+        default=0,
+        render_kw={
+            "placeholder": "0 = unlimited",
+            "style": "width: 8em;",
+        },
+    )
+    token_budget_month = IntegerField(
+        _l('Monthly token budget'),
+        validators=[validators.Optional(), validators.NumberRange(min=0)],
+        default=0,
+        render_kw={"style": "width: 10em;"},
+    )
+    max_input_chars = IntegerField(
+        _l('Max input characters'),
+        validators=[validators.Optional(), validators.NumberRange(min=1)],
+        default=100000,
+        render_kw={
+            "placeholder": "100000",
+            "style": "width: 10em;",
+        },
+    )
+    # Master on/off switch for ALL LLM lookups at runtime. When False, every entry point
+    # in evaluator.py (and the restock fallback) short-circuits with a logger.debug
+    # message — even if a provider+model is still configured. Saved config and the
+    # "configured" badge remain visible so the user can toggle back on without re-entering.
+    enabled = BooleanField(
+        _l('Enable AI / LLM features'),
+        default=True,
+    )
+    override_diff_with_summary = BooleanField(
+        _l('Replace {{diff}} notification token with AI summary'),
+        default=True,
+    )
+    restock_use_fallback_extract = BooleanField(
+        _l('Use LLM as a fallback for extracting price and restock info'),
+        default=True,
+    )
+    debug = BooleanField(
+        _l('Enable LLM debug logging'),
+        default=False,
+    )
+    thinking_budget = SelectField(
+        _l('AI thinking budget (tokens)'),
+        choices=[
+            ('0',    _l('Off (no thinking)')),
+            ('100',  '100'),
+            ('500',  '500'),
+            ('2000', '2000'),
+        ],
+        default=str(LLM_DEFAULT_THINKING_BUDGET),
+        validators=[validators.Optional()],
+    )
+    max_summary_tokens = SelectField(
+        _l('Max AI summary length (tokens)'),
+        choices=[
+            ('500',   '500'),
+            ('1000',  '1000'),
+            ('3000',  '3000'),
+            ('5000',  '5000'),
+            ('10000', '10000'),
+            ('15000', '15000'),
+        ],
+        default=str(LLM_DEFAULT_MAX_SUMMARY_TOKENS),
+        validators=[validators.Optional()],
+    )
+    budget_action = RadioField(
+        _l('When monthly token budget is reached'),
+        choices=[
+            ('skip_llm',   _l('Skip AI summarisation only (watch still checks)')),
+            ('skip_check', _l('Skip the watch check entirely')),
+        ],
+        default='skip_llm',
+    )
+
+
 class globalSettingsForm(Form):
    # Define these as FormFields/"sub forms", this way it matches the JSON storage
    # datastore.data['settings']['application']..
@@ -1049,6 +1248,7 @@ class globalSettingsForm(Form):

    requests = FormField(globalSettingsRequestForm)
    application = FormField(globalSettingsApplicationForm)
+    llm = FormField(globalSettingsLLMForm)
    save_button = SubmitField(_l('Save'), render_kw={"class": "pure-button pure-button-primary"})


@@ -4,6 +4,7 @@ from loguru import logger
 from typing import List
 import html
 import json
+import os
 import re

 # HTML added to be sure each result matching a filter (.example) gets converted to a new line by Inscriptis
@@ -13,6 +14,45 @@ PERL_STYLE_REGEX = r'^/(.*?)/([a-z]*)?$'

 TITLE_RE = re.compile(r"<title[^>]*>(.*?)</title>", re.I | re.S)
 META_CS  = re.compile(r'<meta[^>]+charset=["\']?\s*([a-z0-9_\-:+.]+)', re.I)
+
+# jq builtins that can leak sensitive data or cause harm when user-supplied expressions are executed.
+# env/$ENV reads all process environment variables (passwords, API keys, etc.)
+# include/import can read arbitrary files from disk
+# input/inputs reads beyond the supplied JSON data
+# debug/stderr leaks data to stderr
+# halt/halt_error terminates the process (DoS)
+_JQ_BLOCKED_PATTERNS = [
+    (re.compile(r'\benv\b'),                    'env (reads environment variables)'),
+    (re.compile(r'\$ENV\b'),                    '$ENV (reads environment variables)'),
+    (re.compile(r'\binclude\b'),                'include (reads files from disk)'),
+    (re.compile(r'\bimport\b'),                 'import (reads files from disk)'),
+    (re.compile(r'\binputs?\b'),                'input/inputs (reads beyond provided data)'),
+    (re.compile(r'\bdebug\b'),                  'debug (leaks data to stderr)'),
+    (re.compile(r'\bstderr\b'),                 'stderr (leaks data to stderr)'),
+    (re.compile(r'\bhalt(?:_error)?\b'),        'halt/halt_error (terminates the process)'),
+    (re.compile(r'\$__loc__\b'),                '$__loc__ (leaks file path information)'),
+    (re.compile(r'\bbuiltins\b'),               'builtins (enumerates available functions)'),
+    (re.compile(r'\bmodulemeta\b'),             'modulemeta (leaks module information)'),
+    (re.compile(r'\$JQ_BUILD_CONFIGURATION\b'), '$JQ_BUILD_CONFIGURATION (leaks build information)'),
+]
+
+def validate_jq_expression(expression: str) -> None:
+    """Raise ValueError if the jq expression uses any dangerous builtin.
+
+    User-supplied jq expressions are executed server-side. Without this check,
+    builtins like `env` expose every process environment variable (SALTED_PASS,
+    proxy credentials, API keys, etc.) as watch output.
+    """
+    from changedetectionio.strtobool import strtobool
+    if strtobool(os.getenv('JQ_ALLOW_RISKY_EXPRESSIONS', 'false')):
+        return
+
+    for pattern, description in _JQ_BLOCKED_PATTERNS:
+        if pattern.search(expression):
+            msg = f"jq expression uses disallowed builtin: {description}"
+            logger.critical(f"Security: blocked jq expression containing '{description}' - expression: {expression!r}")
+            raise ValueError(msg)
+
 META_CT  = re.compile(r'<meta[^>]+http-equiv=["\']?content-type["\']?[^>]*content=["\'][^>]*charset=([a-z0-9_\-:+.]+)', re.I)

 # 'price' , 'lowPrice', 'highPrice' are usually under here
@@ -23,6 +63,59 @@ class JSONNotFound(ValueError):
    def __init__(self, msg):
        ValueError.__init__(self, msg)

+
+_DEFAULT_UNSAFE_XPATH3_FUNCTIONS = [
+    'unparsed-text',
+    'unparsed-text-lines',
+    'unparsed-text-available',
+    'doc',
+    'doc-available',
+    'json-doc',
+    'json-doc-available',
+    'collection',           # XPath 2.0+: loads XML node collections from arbitrary URIs
+    'uri-collection',       # XPath 3.0+: enumerates URIs from resource collections
+    'transform',            # XPath 3.1: XSLT transformation (currently raises, block proactively)
+    'load-xquery-module',   # XPath 3.1: loads XQuery modules (currently raises, block proactively)
+    'environment-variable',
+    'available-environment-variables',
+]
+
+
+def _build_safe_xpath3_parser():
+    """Return an XPath3Parser subclass with filesystem/environment access functions removed.
+
+    XPath 3.0 includes functions that can read arbitrary files or environment variables:
+      - unparsed-text / unparsed-text-lines / unparsed-text-available  (file read)
+      - doc / doc-available                                             (XML fetch from URI)
+      - environment-variable / available-environment-variables         (env var leakage)
+
+    Subclassing gives us an independent symbol_table copy (not shared with the parent class),
+    so removing entries here does not affect XPath3Parser itself.
+
+    Override the blocked list via the XPATH_BLOCKED_FUNCTIONS environment variable
+    (comma-separated, e.g. "unparsed-text,doc,environment-variable").
+    """
+    import os
+    from elementpath.xpath3 import XPath3Parser
+
+    class SafeXPath3Parser(XPath3Parser):
+        pass
+
+    env_override = os.getenv('XPATH_BLOCKED_FUNCTIONS')
+    if env_override is not None:
+        blocked = [f.strip() for f in env_override.split(',') if f.strip()]
+    else:
+        blocked = _DEFAULT_UNSAFE_XPATH3_FUNCTIONS
+
+    for _fn in blocked:
+        SafeXPath3Parser.symbol_table.pop(_fn, None)
+
+    return SafeXPath3Parser
+
+
+# Module-level singleton — built once, reused everywhere.
+SafeXPath3Parser = _build_safe_xpath3_parser()
+
 # Doesn't look like python supports forward slash auto enclosure in re.findall
 # So convert it to inline flag "(?i)foobar" type configuration
@lru_cache(maxsize=100)
@@ -183,15 +276,13 @@ def xpath_filter(xpath_filter, html_content, append_pretty_line_formatting=False
    """
    from lxml import etree, html
    import elementpath
-    # xpath 2.0-3.1
-    from elementpath.xpath3 import XPath3Parser

    parser = etree.HTMLParser()
    tree = None
    try:
        if is_xml:
            # So that we can keep CDATA for cdata_in_document_to_text() to process
-            parser = etree.XMLParser(strip_cdata=False)
+            parser = etree.XMLParser(strip_cdata=False, resolve_entities=False, no_network=True)
            # For XML/RSS content, use etree.fromstring to properly handle XML declarations
            tree = etree.fromstring(html_content.encode('utf-8') if isinstance(html_content, str) else html_content, parser=parser)
        else:
@@ -210,7 +301,7 @@ def xpath_filter(xpath_filter, html_content, append_pretty_line_formatting=False
            # This allows //title to match elements in the default namespace
            namespaces[''] = tree.nsmap[None]

-        r = elementpath.select(tree, xpath_filter.strip(), namespaces=namespaces, parser=XPath3Parser)
+        r = elementpath.select(tree, xpath_filter.strip(), namespaces=namespaces, parser=SafeXPath3Parser)
        #@note: //title/text() now works with default namespaces (fixed by registering '' prefix)
        #@note: //title/text() wont work where <title>CDATA.. (use cdata_in_document_to_text first)

@@ -235,6 +326,9 @@ def xpath_filter(xpath_filter, html_content, append_pretty_line_formatting=False
            else:
                html_block += elementpath_tostring(element)

+        # Drop element references before the finally block so tree.clear() can release
+        # the libxml2 document immediately (elements pin the C-level doc via refcount).
+        del r
        return html_block
    finally:
        # Explicitly clear the tree to free memory
@@ -252,7 +346,7 @@ def xpath1_filter(xpath_filter, html_content, append_pretty_line_formatting=Fals
    try:
        if is_xml:
            # So that we can keep CDATA for cdata_in_document_to_text() to process
-            parser = etree.XMLParser(strip_cdata=False)
+            parser = etree.XMLParser(strip_cdata=False, resolve_entities=False, no_network=True)
            # For XML/RSS content, use etree.fromstring to properly handle XML declarations
            tree = etree.fromstring(html_content.encode('utf-8') if isinstance(html_content, str) else html_content, parser=parser)
        else:
@@ -330,12 +424,16 @@ def _parse_json(json_data, json_filter):
            raise Exception("jq not support not found")

        if json_filter.startswith("jq:"):
-            jq_expression = jq.compile(json_filter.removeprefix("jq:"))
+            expr = json_filter.removeprefix("jq:")
+            validate_jq_expression(expr)
+            jq_expression = jq.compile(expr)
            match = jq_expression.input(json_data).all()
            return _get_stripped_text_from_json_match(match)

        if json_filter.startswith("jqraw:"):
-            jq_expression = jq.compile(json_filter.removeprefix("jqraw:"))
+            expr = json_filter.removeprefix("jqraw:")
+            validate_jq_expression(expr)
+            jq_expression = jq.compile(expr)
            match = jq_expression.input(json_data).all()
            return '\n'.join(str(item) for item in match)

@@ -439,13 +537,25 @@ def extract_json_as_string(content, json_filter, ensure_is_ldjson_info_type=None
        except json.JSONDecodeError as e:
            logger.warning(f"Error processing JSON {content[:20]}...{str(e)})")
    else:
-        # Probably something else, go fish inside for it
-        try:
-            stripped_text_from_html = extract_json_blob_from_html(content=content,
-                                                                  ensure_is_ldjson_info_type=ensure_is_ldjson_info_type,
-                                                                  json_filter=json_filter                                                                  )
-        except json.JSONDecodeError as e:
-            logger.warning(f"Error processing JSON while extracting JSON from HTML blob {content[:20]}...{str(e)})")
+        # Check for JSONP wrapper: someCallback({...}) or some.namespace({...})
+        # Server may claim application/json but actually return JSONP
+        jsonp_match = re.match(r'^\w[\w.]*\s*\((.+)\)\s*;?\s*$', content.lstrip("\ufeff").strip(), re.DOTALL)
+        if jsonp_match:
+            try:
+                inner = jsonp_match.group(1).strip()
+                logger.warning(f"Content looks like JSONP, attempting to extract inner JSON for filter '{json_filter}'")
+                stripped_text_from_html = _parse_json(json.loads(inner), json_filter)
+            except json.JSONDecodeError as e:
+                logger.warning(f"Error processing JSONP inner content {content[:20]}...{str(e)})")
+
+        if not stripped_text_from_html:
+            # Probably something else, go fish inside for it
+            try:
+                stripped_text_from_html = extract_json_blob_from_html(content=content,
+                                                                      ensure_is_ldjson_info_type=ensure_is_ldjson_info_type,
+                                                                      json_filter=json_filter)
+            except json.JSONDecodeError as e:
+                logger.warning(f"Error processing JSON while extracting JSON from HTML blob {content[:20]}...{str(e)})")

    if not stripped_text_from_html:
        # Re 265 - Just return an empty string when filter not found
@@ -561,10 +671,33 @@ def html_to_text(html_content: str, render_anchor_tag_content=False, is_rss=Fals
        )
    else:
        parser_config = None
-
    if is_rss:
        html_content = re.sub(r'<title([\s>])', r'<h1\1', html_content)
        html_content = re.sub(r'</title>', r'</h1>', html_content)
+    else:
+        # Use BS4 html.parser to strip bloat — SPA's often dump 10MB+ of CSS/JS into <head>,
+        # causing inscriptis to silently give up. Regex-based stripping is unsafe because tags
+        # can appear inside JSON data attributes with JS-escaped closing tags (e.g. <\/script>),
+        # causing the regex to scan past the intended close and eat real page content.
+        from bs4 import BeautifulSoup
+        soup = BeautifulSoup(html_content, 'html.parser')
+        # Strip tags that inscriptis cannot render as meaningful text and which can be very large.
+        # svg/math: produce path-data/MathML garbage; canvas/iframe/template: no inscriptis handlers.
+        # video/audio/picture are kept — they may contain meaningful fallback text or captions.
+        for tag in soup.find_all(['head', 'script', 'style', 'noscript', 'svg',
+                                  'math', 'canvas', 'iframe', 'template']):
+            tag.decompose()
+
+        # SPAs often use <body style="display:none"> to hide content until JS loads.
+        # inscriptis respects CSS display rules, so strip hiding styles from the body tag.
+        body_tag = soup.find('body')
+        if body_tag and body_tag.get('style'):
+            style = body_tag['style']
+            if re.search(r'\b(?:display\s*:\s*none|visibility\s*:\s*hidden)\b', style, re.IGNORECASE):
+                logger.debug(f"html_to_text: Removing hiding styles from body tag (found: '{style}')")
+                del body_tag['style']
+
+        html_content = str(soup)

    text_content = get_text(html_content, config=parser_config)
    return text_content
@@ -28,17 +28,20 @@ def get_timeago_locale(flask_locale):
        str: timeago library locale code (e.g., 'en', 'zh_CN', 'pt_PT')
    """
    locale_map = {
-        'zh': 'zh_CN',      # Chinese Simplified
+        'zh': 'zh_CN',          # Chinese Simplified
        # timeago library just hasn't been updated to use the more modern locale naming convention, before BCP 47 / RFC 5646.
-        'zh_TW': 'zh_TW',   # Chinese Traditional (timeago uses zh_TW)
+        'zh_TW': 'zh_TW',       # Chinese Traditional (timeago uses zh_TW)
        'zh_Hant_TW': 'zh_TW',  # Flask-Babel normalizes zh_TW to zh_Hant_TW, map back to timeago's zh_TW
-        'pt': 'pt_PT',      # Portuguese (Portugal)
-        'sv': 'sv_SE',      # Swedish
-        'no': 'nb_NO',      # Norwegian Bokmål
-        'hi': 'in_HI',      # Hindi
-        'cs': 'en',         # Czech not supported by timeago, fallback to English
-        'en_GB': 'en',      # British English - timeago uses 'en'
-        'en_US': 'en',      # American English - timeago uses 'en'
+        'pt': 'pt_PT',          # Portuguese (Portugal)
+        'pt_BR': 'pt_BR',       # Portuguese (Brasil)
+        'sv': 'sv_SE',          # Swedish
+        'no': 'nb_NO',          # Norwegian Bokmål
+        'hi': 'in_HI',          # Hindi
+        'cs': 'en',             # Czech not supported by timeago, fallback to English
+        'ja': 'ja',             # Japanese
+        'uk': 'uk',             # Ukrainian
+        'en_GB': 'en',          # British English - timeago uses 'en'
+        'en_US': 'en',          # American English - timeago uses 'en'
    }
    return locale_map.get(flask_locale, flask_locale)

@@ -52,7 +55,8 @@ LANGUAGE_DATA = {
    'ko': {'flag': 'fi fi-kr fis', 'name': '한국어'},
    'cs': {'flag': 'fi fi-cz fis', 'name': 'Čeština'},
    'es': {'flag': 'fi fi-es fis', 'name': 'Español'},
-    'pt': {'flag': 'fi fi-pt fis', 'name': 'Português'},
+    'pt': {'flag': 'fi fi-pt fis', 'name': 'Português (Portugal)'},
+    'pt_BR': {'flag': 'fi fi-br fis', 'name': 'Português (Brasil)'},
    'it': {'flag': 'fi fi-it fis', 'name': 'Italiano'},
    'ja': {'flag': 'fi fi-jp fis', 'name': '日本語'},
    'zh': {'flag': 'fi fi-cn fis', 'name': '中文 (简体)'},
@@ -67,6 +71,7 @@ LANGUAGE_DATA = {
    'tr': {'flag': 'fi fi-tr fis', 'name': 'Türkçe'},
    'ar': {'flag': 'fi fi-sa fis', 'name': 'العربية'},
    'hi': {'flag': 'fi fi-in fis', 'name': 'हिन्दी'},
+    'uk': {'flag': 'fi fi-ua fis', 'name': 'Українська'},
 }


@@ -0,0 +1 @@
+# LLM intent-based change evaluation
@@ -0,0 +1,52 @@
+"""
+BM25-based relevance trimming for large snapshot text.
+
+When a snapshot is large and no CSS pre-filter has narrowed it down,
+we use BM25 to select the lines most relevant to the user's intent
+before sending to the LLM. This keeps the context focused without
+an arbitrary char truncation.
+
+Pure functions — no side effects, fully testable.
+"""
+
+MAX_CONTEXT_CHARS = 15_000
+
+
+def trim_to_relevant(text: str, query: str, max_chars: int = MAX_CONTEXT_CHARS) -> str:
+    """
+    Return the lines from `text` most relevant to `query` up to `max_chars`.
+    If text fits within budget, return it unchanged.
+    Falls back to head-truncation if rank_bm25 is unavailable.
+    """
+    if not text or not query:
+        return text or ''
+
+    if len(text) <= max_chars:
+        return text
+
+    lines = [l for l in text.splitlines() if l.strip()]
+    if not lines:
+        return text[:max_chars]
+
+    try:
+        from rank_bm25 import BM25Okapi
+    except ImportError:
+        # rank-bm25 not installed — fall back to simple head truncation
+        return text[:max_chars]
+
+    tokenized = [line.lower().split() for line in lines]
+    bm25 = BM25Okapi(tokenized)
+    scores = bm25.get_scores(query.lower().split())
+
+    ranked = sorted(enumerate(zip(scores, lines)), key=lambda x: x[1][0], reverse=True)
+
+    selected_indices, total = [], 0
+    for idx, (_score, line) in ranked:
+        if total + len(line) + 1 > max_chars:
+            break
+        selected_indices.append(idx)
+        total += len(line) + 1
+
+    # Re-order selected lines to preserve original document order
+    ordered = [lines[i] for i in sorted(selected_indices)]
+    return '\n'.join(ordered)
@@ -0,0 +1,162 @@
+"""
+Thin wrapper around litellm.completion.
+Keeps litellm import isolated so the rest of the codebase doesn't depend on it directly,
+and makes the call easy to mock in tests.
+"""
+
+import logging
+import os
+from loguru import logger
+
+# Default output token cap for JSON-returning calls (intent eval, preview, setup).
+# These return small JSON objects — 400 is enough for a verbose explanation while
+# still preventing runaway cost. Change summaries pass their own max_tokens via
+# _summary_max_tokens() and are NOT subject to this cap.
+_MAX_COMPLETION_TOKENS = 400
+
+DEFAULT_TIMEOUT = int(os.getenv('LLM_TIMEOUT', 60))
+DEFAULT_RETRIES = 3
+
+
+class _LoguruInterceptHandler(logging.Handler):
+    # Routes litellm's stdlib log records through loguru so debug output
+    # uses the same format/sink as the rest of the app.
+    def emit(self, record):
+        try:
+            level = logger.level(record.levelname).name
+        except (ValueError, AttributeError):
+            level = record.levelno
+        logger.opt(exception=record.exc_info).log(level, record.getMessage())
+
+
+_debug_installed = False
+
+
+def _install_litellm_debug():
+    # Attach our loguru intercept and clear any pre-existing handlers so litellm's
+    # own stdout StreamHandler (installed by _turn_on_debug / set_verbose) doesn't
+    # double-emit. Setting the logger level to DEBUG is enough to make litellm
+    # produce debug records — we don't call _turn_on_debug() for that reason.
+    global _debug_installed
+    if _debug_installed:
+        return
+
+    handler = _LoguruInterceptHandler()
+    handler.setLevel(logging.DEBUG)
+    for _name in ('LiteLLM', 'litellm', 'litellm.utils', 'litellm.router'):
+        _lg = logging.getLogger(_name)
+        _lg.handlers = []
+        _lg.setLevel(logging.DEBUG)
+        _lg.addHandler(handler)
+        _lg.propagate = False
+
+    _debug_installed = True
+    logger.info("LLM client: litellm debug logging routed through loguru")
+
+
+def completion(model: str, messages: list, api_key: str = None,
+               api_base: str = None, timeout: int = DEFAULT_TIMEOUT,
+               max_tokens: int = None, extra_body: dict = None,
+               debug: bool = False) -> tuple[str, int, int, int]:
+    """
+    Call the LLM and return (response_text, total_tokens, input_tokens, output_tokens).
+    Retries up to DEFAULT_RETRIES times on timeout or connection errors.
+    Token counts are 0 if the provider doesn't return usage data.
+    Raises on network/auth errors — callers handle gracefully.
+    """
+    try:
+        import litellm
+    except ImportError:
+        raise RuntimeError("litellm is not installed. Add it to requirements.txt.")
+
+    if debug:
+        _install_litellm_debug()
+
+    _timeout = timeout if timeout is not None else DEFAULT_TIMEOUT
+
+    kwargs = {
+        'model': model,
+        'messages': messages,
+        'timeout': _timeout,
+        'temperature': 0,
+        'max_tokens': max_tokens if max_tokens is not None else _MAX_COMPLETION_TOKENS,
+    }
+    if api_key:
+        kwargs['api_key'] = api_key
+    if api_base:
+        kwargs['api_base'] = api_base
+    if extra_body:
+        kwargs['extra_body'] = extra_body
+
+    _retryable = (litellm.Timeout, litellm.APIConnectionError)
+
+    logger.debug(
+        f"LLM client: calling model={model!r} api_base={api_base!r} "
+        f"timeout={_timeout}s max_tokens={kwargs['max_tokens']}"
+    )
+    logger.trace(messages)
+
+    for attempt in range(1, DEFAULT_RETRIES + 1):
+        try:
+            response = litellm.completion(**kwargs)
+            choice   = response.choices[0]
+            message  = choice.message
+            finish   = getattr(choice, 'finish_reason', None)
+
+            text = message.content or ''
+
+            if not text:
+                # Some providers (e.g. Gemini) put text in message.parts instead of .content
+                parts = getattr(message, 'parts', None)
+                if parts:
+                    text = ''.join(getattr(p, 'text', '') or '' for p in parts).strip()
+                    logger.debug(f"LLM client: extracted text from message.parts ({len(parts)} parts) model={model!r}")
+
+            if finish == 'length':
+                logger.warning(
+                    f"LLM client: response truncated (finish_reason='length') model={model!r} "
+                    f"— increase max_tokens; got {len(text)} chars so far"
+                )
+
+            if not text:
+                logger.warning(
+                    f"LLM client: empty content from model={model!r} "
+                    f"finish_reason={finish!r} "
+                    f"message={message!r}"
+                )
+
+            usage = getattr(response, 'usage', None)
+            input_tokens  = int(getattr(usage, 'prompt_tokens',     0) or 0) if usage else 0
+            output_tokens = int(getattr(usage, 'completion_tokens', 0) or 0) if usage else 0
+            total_tokens  = int(getattr(usage, 'total_tokens',      0) or 0) if usage else (input_tokens + output_tokens)
+            logger.debug(
+                f"LLM client: model={model!r} finish={finish!r} "
+                f"tokens={total_tokens} (in={input_tokens} out={output_tokens}) "
+                f"text_len={len(text)}"
+            )
+            return text, total_tokens, input_tokens, output_tokens
+
+        except _retryable as e:
+            # litellm formats its Timeout message with None when the provider doesn't
+            # propagate the timeout value — patch the exception args in-place so every
+            # caller that logs str(e) sees the real number.
+            _fix = f'after {_timeout} seconds'
+            try:
+                e.args = tuple(str(a).replace('after None seconds', _fix) for a in e.args)
+            except Exception:
+                pass
+            if attempt < DEFAULT_RETRIES:
+                logger.warning(
+                    f"LLM call timed out/connection error (attempt {attempt}/{DEFAULT_RETRIES}), "
+                    f"retrying — model={model!r} timeout={_timeout}s error={e}"
+                )
+                continue
+            logger.warning(
+                f"LLM call failed after {DEFAULT_RETRIES} attempts ({_timeout}s timeout) "
+                f"model={model!r} error={e}"
+            )
+            raise
+
+        except Exception as e:
+            logger.warning(f"LLM call failed: model={model!r} error={e}")
+            raise
@@ -0,0 +1,804 @@
+"""
+LLM evaluation orchestration.
+
+Two public entry points:
+  - run_setup(watch, datastore)        — one-time: decide if pre-filter needed
+  - evaluate_change(watch, datastore, diff, current_snapshot) — per-change evaluation
+
+Intent resolution: watch.llm_intent → first tag with llm_intent → None (no evaluation)
+Cache: each (intent, diff) pair is evaluated exactly once, result stored in watch.
+
+Environment variable overrides (take priority over datastore settings):
+  LLM_MODEL    — model string (e.g. "gpt-4o-mini", "ollama/llama3.2")
+  LLM_API_KEY  — API key for cloud providers
+  LLM_API_BASE — base URL for local/custom endpoints (e.g. http://localhost:11434)
+"""
+
+import hashlib
+import os
+from dataclasses import dataclass
+from datetime import datetime, timezone
+from loguru import logger
+
+from changedetectionio.strtobool import strtobool
+
+from . import client as llm_client
+from .prompt_builder import (
+    build_change_summary_prompt, build_change_summary_system_prompt,
+    build_eval_prompt, build_eval_system_prompt,
+    build_preview_prompt, build_preview_system_prompt,
+    build_setup_prompt, build_setup_system_prompt,
+)
+from .response_parser import parse_eval_response, parse_preview_response, parse_setup_response
+
+from changedetectionio.model.LLMSettings import (
+    LLMSettings,
+    LLM_DEFAULT_MAX_INPUT_CHARS as _DEFAULT_MAX_INPUT_CHARS,
+    LLM_DEFAULT_MAX_SUMMARY_TOKENS,
+    LLM_DEFAULT_THINKING_BUDGET,
+)
+
+
+def is_llm_features_disabled() -> bool:
+    """True when the LLM_FEATURES_DISABLED env var is set to a truthy value."""
+    return bool(strtobool(os.getenv('LLM_FEATURES_DISABLED', '')))
+
+
+def get_llm_settings(datastore) -> LLMSettings:
+    """Hydrate the LLM config dict at settings.application.llm into a validated model.
+
+    Returns a default-constructed LLMSettings when the dict is missing or empty —
+    callers never have to None-check the result. The storage layer remains a plain
+    dict; this is only the validation/typing layer for reads.
+    """
+    cfg = datastore.data.get('settings', {}).get('application', {}).get('llm') or {}
+    return LLMSettings.model_validate(cfg)
+
+
+def _get_max_input_chars(datastore) -> int:
+    """Max input characters to send to the LLM. Resolution: env var → datastore → 100,000.
+    Always returns at least 1 — unlimited is not permitted.
+    """
+    env_val = os.getenv('LLM_MAX_INPUT_CHARS', '').strip()
+    if env_val.isdigit() and int(env_val) > 0:
+        return int(env_val)
+    stored = get_llm_settings(datastore).max_input_chars
+    if stored and stored > 0:
+        return stored
+    return _DEFAULT_MAX_INPUT_CHARS
+
+
+class LLMInputTooLargeError(Exception):
+    pass
+
+
+def _check_input_size(text: str, max_chars: int) -> None:
+    """Raise LLMInputTooLargeError if text exceeds max_chars."""
+    if len(text) > max_chars:
+        raise LLMInputTooLargeError(
+            f"Change too large for AI summary ({len(text):,} chars, limit {max_chars:,})"
+        )
+
+
+def _thinking_extra_body(model: str, budget: int) -> dict | None:
+    """Return litellm extra_body to control thinking for models that support it.
+
+    The `thinkingConfig.thinkingBudget` payload is Gemini-specific (Anthropic and
+    OpenAI reasoning models use different parameters), so we gate on the gemini/
+    provider prefix first, then defer to litellm's model registry for the actual
+    "does this model think?" decision. That picks up new Gemini variants and
+    rolling aliases (`gemini-flash-latest`, etc.) as litellm's registry tracks
+    them, without us hardcoding model names here.
+    """
+    if not model.startswith('gemini/'):
+        return None
+    try:
+        import litellm
+        if not litellm.get_model_info(model).get('supports_reasoning'):
+            return None
+    except Exception:
+        # Unknown model or registry lookup failed — skip the thinking config
+        # rather than guess. Worst case: thinking stays at the provider default.
+        return None
+    return {'generationConfig': {'thinkingConfig': {'thinkingBudget': budget}}}
+
+
+def _cached_system(text: str, model: str = '') -> dict:
+    """Wrap a system prompt, adding Anthropic prompt-caching headers only for Anthropic models.
+    Gemini and other providers have their own caching APIs that break when they receive
+    cache_control, so we only apply it where it's supported.
+    """
+    is_anthropic = model.startswith('claude') or model.startswith('anthropic/')
+    if is_anthropic:
+        return {'role': 'system', 'content': [{'type': 'text', 'text': text, 'cache_control': {'type': 'ephemeral'}}]}
+    return {'role': 'system', 'content': text}
+
+
+# Output-token cap for the JSON-returning calls (intent eval, preview, setup/prefilter).
+# Mirrors client.py's _MAX_COMPLETION_TOKENS so the multiplier helper has a base value
+# to scale; cloud-LLM users hit this default unmodified, preserving prior cost defaults.
+JSON_RESPONSE_MAX_TOKENS = 400
+
+# Default prompt used when the user hasn't configured llm_change_summary.
+# This owns the OUTPUT FORMAT (structure, sections, style, language). The system prompt
+# in prompt_builder.build_change_summary_system_prompt() only covers how to READ the diff.
+# Users can replace this entirely (e.g. "Just tell me the new timestamp.") without
+# fighting hard-coded structure rules from the system prompt.
+DEFAULT_CHANGE_SUMMARY_PROMPT = (
+    "Describe what changed in plain English using these sections, in this fixed order — "
+    "omit a section entirely if there is nothing to report for it:\n"
+    "  Added: ...\n"
+    "  Changed: ...\n"
+    "  Removed: ...\n"
+    "The Removed section MUST always be last. Never place removals before additions or changes.\n\n"
+    "List items as bullet points with key details for each one. Be considerate of the style "
+    "of content you are summarising and adjust your report accordingly.\n"
+    "Do not list standalone timestamps like '3 hours ago', 'Yesterday', '2 minutes ago' as added "
+    "or removed items — they are not meaningful content changes.\n"
+    "For content-heavy pages (news, listings, feeds): quote or paraphrase the specific new "
+    "headlines, items, or entries that were added — do not collapse them into vague phrases "
+    "like 'new articles were added' or 'section was expanded'.\n"
+    "For large blocks of new text (full articles, documents, long paragraphs): briefly summarise "
+    "the substance in 1-2 sentences capturing the key point — do not just repeat the title.\n\n"
+    "Do not quote non-English text verbatim; translate and summarise all content into English. "
+    "Your entire response must be in English."
+)
+
+
+def _summary_max_tokens(diff: str, max_cap: int = LLM_DEFAULT_MAX_SUMMARY_TOKENS) -> int:
+    """Scale completion tokens to diff size: floor 400, ~1 token per 4 chars, ceiling max_cap."""
+    return max(400, min(len(diff) // 4, max_cap))
+
+
+def apply_local_token_multiplier(base_max_tokens: int, llm_cfg: dict) -> int:
+    """
+    Scale max_tokens for endpoints that commonly serve reasoning models
+    (Ollama — self-hosted or ollama.com cloud — and OpenAI-compatible servers like
+    vLLM, LM Studio, llama.cpp).
+
+    Reasoning models (Qwen3, DeepSeek-R1, Gemma 3, etc.) emit chain-of-thought into
+    `message.reasoning_content` BEFORE the final answer lands in `message.content`.
+    Without enough headroom the request truncates mid-thought (`finish_reason='length'`
+    or `'stop'` with empty content) and the answer never lands — callers see an empty
+    string and silently fall through to safe defaults, hiding the problem.
+
+    Cloud providers with stable, non-reasoning defaults (OpenAI, Anthropic, Gemini,
+    OpenRouter) keep their original tight caps so existing users see no behavior or
+    cost change. Ollama / OpenAI-compatible users can dial the multiplier down to 1x
+    in Settings → AI → Provider if they want to keep costs tight on a paid endpoint.
+
+    Activated when `llm_cfg['provider_kind']` is `'ollama'` or `'openai_compatible'`.
+    Multiplier defaults to 5x and is user-configurable in Settings → AI → Provider.
+    """
+    if (llm_cfg or {}).get('provider_kind') not in ('ollama', 'openai_compatible'):
+        return base_max_tokens
+    try:
+        multiplier = int(llm_cfg.get('local_token_multiplier') or 5)
+    except (TypeError, ValueError):
+        multiplier = 5
+    # Clamp to the same 1-20 range the form enforces. Defense-in-depth against
+    # corrupted datastore values that bypassed form validation (manual JSON edits,
+    # future migrations, plugins): a runaway multiplier could otherwise produce
+    # absurdly large max_tokens caps and exhaust local-endpoint memory.
+    multiplier = max(1, min(multiplier, 20))
+    return base_max_tokens * multiplier
+
+
+# ---------------------------------------------------------------------------
+# Intent resolution
+# ---------------------------------------------------------------------------
+
+def resolve_llm_field(watch, datastore, field: str) -> tuple[str, str]:
+    """
+    Generic cascade resolver for any LLM per-watch field.
+    Returns (value, source) where source is 'watch' or tag title.
+    Returns ('', '') if not set anywhere.
+    """
+    value = (watch.get(field) or '').strip()
+    if value:
+        return value, 'watch'
+
+    for tag_uuid in watch.get('tags', []):
+        tag = datastore.data['settings']['application'].get('tags', {}).get(tag_uuid)
+        if tag:
+            tag_value = (tag.get(field) or '').strip()
+            if tag_value:
+                return tag_value, tag.get('title', 'tag')
+
+    return '', ''
+
+
+def resolve_intent(watch, datastore) -> tuple[str, str]:
+    """
+    Return (intent, source) where source is 'watch' or tag title.
+    Returns ('', '') if no intent is configured anywhere.
+    """
+    intent = (watch.get('llm_intent') or '').strip()
+    if intent:
+        return intent, 'watch'
+
+    for tag_uuid in watch.get('tags', []):
+        tag = datastore.data['settings']['application'].get('tags', {}).get(tag_uuid)
+        if tag:
+            tag_intent = (tag.get('llm_intent') or '').strip()
+            if tag_intent:
+                return tag_intent, tag.get('title', 'tag')
+
+    return '', ''
+
+
+# ---------------------------------------------------------------------------
+# LLM config helper
+# ---------------------------------------------------------------------------
+
+def get_llm_config(datastore) -> dict | None:
+    """
+    Return LLM config dict or None if not configured.
+
+    Resolution order (first non-empty model wins):
+      1. Environment variables: LLM_MODEL, LLM_API_KEY, LLM_API_BASE
+      2. Datastore settings (set via UI)
+    """
+    if is_llm_features_disabled():
+        return None
+    # 1. Environment variable override
+    env_model = os.getenv('LLM_MODEL', '').strip()
+    if env_model:
+        return {
+            'model': env_model,
+            'api_key': os.getenv('LLM_API_KEY', '').strip(),
+            'api_base': os.getenv('LLM_API_BASE', '').strip(),
+        }
+
+    # 2. Datastore settings
+    cfg = datastore.data['settings']['application'].get('llm') or {}
+    if not cfg.get('model'):
+        return None
+    return cfg
+
+
+def llm_configured_via_env() -> bool:
+    """True when LLM config comes from environment variables, not the UI."""
+    if is_llm_features_disabled():
+        return False
+    return bool(os.getenv('LLM_MODEL', '').strip())
+
+
+def _runtime_llm_config(datastore) -> dict | None:
+    """
+    Runtime gate used by every LLM entry point in this module (and the restock
+    fallback). Returns the resolved config dict only when both:
+      - the master 'llm_enabled' toggle is on (default True)
+      - a provider+model is actually configured
+
+    When the toggle is off but a config exists, logs a debug message and returns
+    None so callers fall through their existing "not configured" early-return path.
+
+    The settings UI deliberately still calls get_llm_config() directly so the
+    "AI / LLM configured: ..." badge keeps showing the saved provider even while
+    the toggle is off.
+    """
+    cfg = get_llm_config(datastore)
+    if not get_llm_settings(datastore).enabled:
+        if cfg:
+            logger.debug("LLM features disabled via settings (enabled=False) — skipping LLM lookup")
+        return None
+    return cfg
+
+
+# ---------------------------------------------------------------------------
+# Global monthly token budget
+# ---------------------------------------------------------------------------
+
+def _get_month_key() -> str:
+    """Returns 'YYYY-MM' for the current UTC month."""
+    return datetime.now(timezone.utc).strftime("%Y-%m")
+
+
+def get_global_token_budget_month(datastore=None) -> int:
+    """
+    Monthly token budget ceiling. Resolution order:
+      1. LLM_TOKEN_BUDGET_MONTH env var (takes priority, makes field read-only in UI)
+      2. datastore settings (set via UI)
+    Returns 0 (no limit) if not set anywhere.
+    """
+    try:
+        env_val = int(os.getenv('LLM_TOKEN_BUDGET_MONTH', '0'))
+        if env_val > 0:
+            return env_val
+    except (ValueError, TypeError):
+        pass
+    if datastore is not None:
+        try:
+            stored = datastore.data['settings']['application'].get('llm') or {}
+            val = int(stored.get('token_budget_month') or 0)
+            return max(0, val)
+        except (ValueError, TypeError):
+            pass
+    return 0
+
+
+def _estimate_cost_usd(model: str, input_tokens: int, output_tokens: int) -> float:
+    """
+    Return estimated cost in USD using litellm's pricing database.
+    Returns 0.0 for unknown models (local/Ollama/custom endpoints).
+    Never raises — cost estimation is best-effort.
+    """
+    if not model or (not input_tokens and not output_tokens):
+        return 0.0
+    try:
+        from litellm.cost_calculator import cost_per_token
+        prompt_cost, completion_cost = cost_per_token(
+            model=model,
+            prompt_tokens=input_tokens,
+            completion_tokens=output_tokens,
+        )
+        return float(prompt_cost + completion_cost)
+    except Exception:
+        return 0.0
+
+
+def accumulate_global_tokens(datastore, tokens: int,
+                              input_tokens: int = 0, output_tokens: int = 0,
+                              model: str = '') -> None:
+    """
+    Add *tokens* to both the all-time and this-month global counters.
+    When input_tokens / output_tokens / model are supplied the estimated
+    USD cost is accumulated alongside the token counts.
+    Resets monthly counters automatically on month rollover.
+
+    These counters live at datastore.data['settings']['application']['llm']
+    and are intentionally read-only from the API/form side — they are only
+    ever written here, in a controlled way.
+    """
+    if tokens <= 0:
+        return
+
+    current_month = _get_month_key()
+    cost = _estimate_cost_usd(model, input_tokens, output_tokens)
+    settings = get_llm_settings(datastore)
+
+    # Month rollover: reset monthly counters
+    if settings.tokens_month_key != current_month:
+        settings.tokens_this_month = 0
+        settings.cost_usd_this_month = 0.0
+        settings.tokens_month_key = current_month
+
+    settings.tokens_total_cumulative += tokens
+    settings.tokens_this_month       += tokens
+    settings.cost_usd_total_cumulative += cost
+    settings.cost_usd_this_month       += cost
+
+    # Round-trip through model_dump so storage stays a plain dict and the schema
+    # contract (extra='forbid', type coercion) is re-enforced on every write.
+    datastore.data['settings']['application']['llm'] = settings.model_dump()
+    datastore.commit()
+
+
+def is_global_token_budget_exceeded(datastore) -> bool:
+    """
+    Returns True when a monthly token budget is configured (via
+    LLM_TOKEN_BUDGET_MONTH) and the current month's usage has reached
+    or exceeded that budget.
+    """
+    budget = get_global_token_budget_month(datastore)
+    if not budget:
+        return False
+
+    llm_cfg = datastore.data['settings']['application'].get('llm') or {}
+    if llm_cfg.get('tokens_month_key') != _get_month_key():
+        # Counter hasn't been updated yet this month → zero usage
+        return False
+
+    return (llm_cfg.get('tokens_this_month') or 0) >= budget
+
+
+# ---------------------------------------------------------------------------
+# One-time setup: derive pre-filter
+# ---------------------------------------------------------------------------
+
+def _check_token_budget(watch, cfg, tokens_this_call: int = 0) -> bool:
+    """
+    Per-watch per-period token cap.
+
+    Period is currently month (matches the global counter rollover); the field
+    name `max_tokens_per_count_period` is period-agnostic so a configurable
+    day/week/month can land later without renaming storage.
+
+    On non-zero tokens_this_call:
+      - rolls over watch['llm_tokens_this_period'] if a new period started
+      - increments the per-period counter
+      - also increments the existing lifetime counter (UI stat, unchanged)
+    Returns False once the per-period counter exceeds max_tokens_per_count_period
+    so subsequent evaluate_change calls bail out for this watch until rollover.
+
+    Note: only evaluate_change actually gates on the return value (the other
+    callers invoke this for the side-effect of accumulating tokens).
+    """
+    if tokens_this_call > 0:
+        current_period = _get_month_key()
+        # Rollover: new period zeroes the per-period counter
+        if watch.get('llm_tokens_period_key') != current_period:
+            watch['llm_tokens_this_period'] = 0
+            watch['llm_tokens_period_key'] = current_period
+        watch['llm_tokens_this_period'] = (watch.get('llm_tokens_this_period') or 0) + tokens_this_call
+        # Informational lifetime counter (UI shows this; not used for the cap)
+        watch['llm_tokens_used_cumulative'] = (watch.get('llm_tokens_used_cumulative') or 0) + tokens_this_call
+
+    max_per_period = int(cfg.get('max_tokens_per_count_period') or 0)
+    if max_per_period:
+        # Pre-flight (tokens_this_call=0) and post-call paths both read the
+        # same counter — but a stale period key means "no usage yet this period".
+        if watch.get('llm_tokens_period_key') == _get_month_key():
+            total = watch.get('llm_tokens_this_period') or 0
+            if total > max_per_period:
+                logger.warning(
+                    f"LLM per-period token budget exceeded for {watch.get('uuid')}: "
+                    f"{total} tokens > limit {max_per_period}"
+                )
+                return False
+
+    return True
+
+
+def run_setup(watch, datastore, snapshot_text: str) -> None:
+    """
+    Ask the LLM whether a CSS pre-filter would improve precision for this intent.
+    Stores result in watch['llm_prefilter'] (str selector or None).
+    Called once when intent is first set, and again if pre-filter returns zero matches.
+    """
+    cfg = _runtime_llm_config(datastore)
+    if not cfg:
+        return
+
+    intent, _ = resolve_intent(watch, datastore)
+    if not intent:
+        return
+
+    url = watch.get('url', '')
+    system_prompt = build_setup_system_prompt()
+    user_prompt = build_setup_prompt(intent, snapshot_text, url=url)
+    settings = get_llm_settings(datastore)
+
+    try:
+        raw, tokens, *_ = llm_client.completion(
+            model=cfg['model'],
+            messages=[
+                _cached_system(system_prompt, model=cfg['model']),
+                {'role': 'user', 'content': user_prompt},
+            ],
+            api_key=cfg.get('api_key'),
+            api_base=cfg.get('api_base'),
+            max_tokens=apply_local_token_multiplier(JSON_RESPONSE_MAX_TOKENS, cfg),
+            extra_body=_thinking_extra_body(cfg['model'], settings.thinking_budget),
+            debug=settings.debug,
+        )
+        _check_token_budget(watch, cfg, tokens)
+        accumulate_global_tokens(datastore, tokens, model=cfg['model'])
+        result = parse_setup_response(raw)
+        watch['llm_prefilter'] = result['selector']
+        logger.debug(f"LLM setup for {watch.get('uuid')}: prefilter={result['selector']} reason={result['reason']}")
+    except Exception as e:
+        logger.warning(f"LLM setup call failed for {watch.get('uuid')}: {e}")
+        watch['llm_prefilter'] = None
+
+
+# ---------------------------------------------------------------------------
+# AI Change Summary — human-readable description of what changed
+# ---------------------------------------------------------------------------
+
+def get_effective_summary_prompt(watch, datastore) -> str:
+    """Return the prompt that summarise_change will use.
+
+    Cascade: watch → tag → global settings default → hardcoded fallback.
+    """
+    prompt, _ = resolve_llm_field(watch, datastore, 'llm_change_summary')
+    if prompt:
+        return prompt
+    global_default = get_llm_settings(datastore).change_summary_default.strip()
+    return global_default or DEFAULT_CHANGE_SUMMARY_PROMPT
+
+
+def compute_summary_cache_key(diff_text: str, prompt: str) -> str:
+    """Stable 16-char hex key for a (diff, prompt) pair.  Stored alongside the summary file."""
+    h = hashlib.md5()
+    h.update(diff_text.encode('utf-8', errors='replace'))
+    h.update(b'\x00')
+    h.update(prompt.encode('utf-8', errors='replace'))
+    return h.hexdigest()[:16]
+
+
+@dataclass(frozen=True)
+class DiffPrefs:
+    """
+    User-facing diff display preferences. Part of the LLM summary cache key so
+    that toggling a preference produces a fresh summary.
+
+    Field defaults are the single source of truth — the UI query-arg defaults in
+    diff.py's from_request_args() and the worker pre-cache's bare DiffPrefs()
+    both rely on these.
+    """
+    all_changes:       bool = False
+    ignore_whitespace: bool = False
+    show_removed:      bool = True
+    show_added:        bool = True
+
+    @classmethod
+    def from_request_args(cls, args) -> 'DiffPrefs':
+        """Parse from a Flask request.args (or any .get(key, default)-shaped mapping)."""
+        return cls(
+            all_changes       = args.get('all_changes', '0') == '1',
+            ignore_whitespace = args.get('ignore_whitespace', '0') == '1',
+            show_removed      = args.get('removed', '1') == '1',
+            show_added        = args.get('added', '1') == '1',
+        )
+
+    def cache_key_suffix(self) -> str:
+        return (
+            f'\x00prefs:all={int(self.all_changes)},ws={int(self.ignore_whitespace)}'
+            f',rm={int(self.show_removed)},add={int(self.show_added)}'
+        )
+
+
+def build_summary_cache_prompt(effective_prompt: str, max_summary_tokens: int,
+                                prefs: DiffPrefs = None, model: str = '') -> str:
+    """
+    Compose the full cache-key string passed to save/get_llm_diff_summary.
+
+    Default prefs are DiffPrefs() — must match the UI's query-arg defaults so a
+    worker-side pre-cache is hit by an unmodified UI request. Same helper must
+    be used by both the worker pre-cache write and the UI diff route read,
+    otherwise the prompt hashes diverge and the cache file isn't found.
+
+    The active model name is folded into the key so switching models
+    (e.g. qwen3 → gpt-4o) invalidates stale summaries that were generated
+    by a different model with potentially different phrasing/quality.
+    """
+    if prefs is None:
+        prefs = DiffPrefs()
+    return (
+        effective_prompt
+        + prefs.cache_key_suffix()
+        + f'\x00sys:{build_change_summary_system_prompt()}'
+        + f'\x00max_tokens:{max_summary_tokens}'
+        + f'\x00model:{model}'
+    )
+
+
+def summarise_change(watch, datastore, diff: str, current_snapshot: str = '') -> str:
+    """
+    Generate a plain-language summary of the change using the watch's
+    llm_change_summary prompt (cascades from tag if not set on watch).
+
+    Returns the summary string, or '' on failure.
+    The result replaces {{ diff }} in notifications so the user gets a
+    readable description instead of raw +/- diff lines.
+    """
+    cfg = _runtime_llm_config(datastore)
+    if not cfg:
+        return ''
+
+    if is_global_token_budget_exceeded(datastore):
+        budget = get_global_token_budget_month(datastore)
+        llm_cfg = datastore.data['settings']['application'].get('llm') or {}
+        used = llm_cfg.get('tokens_this_month', 0)
+        logger.warning(
+            f"LLM summarise_change skipped: monthly budget {budget:,} reached "
+            f"({used:,} used this month)"
+        )
+        return ''
+
+    custom_prompt = get_effective_summary_prompt(watch, datastore)
+    if not diff.strip():
+        return ''
+
+    _check_input_size(diff, _get_max_input_chars(datastore))
+    url = watch.get('url', '')
+    title = watch.get('page_title') or watch.get('title') or ''
+
+    system_prompt = build_change_summary_system_prompt()
+    user_prompt = build_change_summary_prompt(
+        diff=diff,
+        custom_prompt=custom_prompt,
+        current_snapshot=current_snapshot,
+        url=url,
+        title=title,
+    )
+
+    settings = get_llm_settings(datastore)
+    _extra_body = _thinking_extra_body(cfg['model'], settings.thinking_budget)
+
+    try:
+        _resp = llm_client.completion(
+            model=cfg['model'],
+            messages=[
+                _cached_system(system_prompt, model=cfg['model']),
+                {'role': 'user', 'content': user_prompt},
+            ],
+            api_key=cfg.get('api_key'),
+            api_base=cfg.get('api_base'),
+            max_tokens=apply_local_token_multiplier(
+                _summary_max_tokens(diff, max_cap=settings.max_summary_tokens),
+                cfg,
+            ),
+            extra_body=_extra_body,
+            debug=settings.debug,
+        )
+        raw, tokens = _resp[0], _resp[1]
+        input_tokens  = _resp[2] if len(_resp) > 2 else 0
+        output_tokens = _resp[3] if len(_resp) > 3 else 0
+        summary = raw.strip()
+        _check_token_budget(watch, cfg, tokens)
+        watch['llm_last_tokens_used'] = tokens
+        watch['llm_tokens_used_cumulative'] = (watch.get('llm_tokens_used_cumulative') or 0) + tokens
+        accumulate_global_tokens(datastore, tokens,
+                                 input_tokens=input_tokens,
+                                 output_tokens=output_tokens,
+                                 model=cfg['model'])
+        logger.debug(
+            f"LLM change summary {watch.get('uuid')}: tokens={tokens} "
+            f"summary={summary[:80]}"
+        )
+        return summary
+    except Exception as e:
+        raise
+
+
+# ---------------------------------------------------------------------------
+# Live-preview extraction (current content, no diff)
+# ---------------------------------------------------------------------------
+
+def preview_extract(watch, datastore, content: str) -> dict | None:
+    """
+    For the live-preview endpoint: extract relevant information from the
+    *current* page content according to the watch's intent.
+
+    Unlike evaluate_change (which compares a diff), this asks the LLM to
+    directly answer the intent against the current snapshot — giving the user
+    immediate feedback like "30 articles listed" or "Price: $149, 25% off".
+
+    Returns {'found': bool, 'answer': str} or None if LLM not configured / no intent.
+    """
+    cfg = _runtime_llm_config(datastore)
+    if not cfg:
+        return None
+
+    intent, _ = resolve_intent(watch, datastore)
+    if not intent or not content.strip():
+        return None
+
+    _check_input_size(content, _get_max_input_chars(datastore))
+    url = watch.get('url', '')
+    title = watch.get('page_title') or watch.get('title') or ''
+
+    system_prompt = build_preview_system_prompt()
+    user_prompt = build_preview_prompt(intent, content, url=url, title=title)
+    settings = get_llm_settings(datastore)
+
+    try:
+        raw, tokens, *_ = llm_client.completion(
+            model=cfg['model'],
+            messages=[
+                _cached_system(system_prompt, model=cfg['model']),
+                {'role': 'user', 'content': user_prompt},
+            ],
+            api_key=cfg.get('api_key'),
+            api_base=cfg.get('api_base'),
+            max_tokens=apply_local_token_multiplier(JSON_RESPONSE_MAX_TOKENS, cfg),
+            extra_body=_thinking_extra_body(cfg['model'], settings.thinking_budget),
+            debug=settings.debug,
+        )
+        accumulate_global_tokens(datastore, tokens, model=cfg['model'])
+        result = parse_preview_response(raw)
+        logger.debug(
+            f"LLM preview {watch.get('uuid')}: found={result['found']} "
+            f"tokens={tokens} answer={result['answer'][:80]}"
+        )
+        return result
+    except Exception as e:
+        logger.warning(f"LLM preview extraction failed for {watch.get('uuid')}: {e}")
+        return None
+
+
+# ---------------------------------------------------------------------------
+# Per-change evaluation
+# ---------------------------------------------------------------------------
+
+def evaluate_change(watch, datastore, diff: str, current_snapshot: str = '') -> dict | None:
+    """
+    Evaluate whether `diff` matches the watch's intent.
+    Returns {'important': bool, 'summary': str} or None if LLM not configured / no intent.
+
+    Results are cached by (intent, diff) hash — each unique diff is evaluated exactly once.
+    """
+    cfg = _runtime_llm_config(datastore)
+    if not cfg:
+        return None
+
+    intent, source = resolve_intent(watch, datastore)
+    if not intent:
+        return None
+
+    if not diff or not diff.strip():
+        return {'important': False, 'summary': ''}
+
+    _check_input_size(diff, _get_max_input_chars(datastore))
+
+    # Cache lookup — evaluations are deterministic once cached
+    cache_key = hashlib.sha256(f"{intent}||{diff}".encode()).hexdigest()
+    cache = watch.get('llm_evaluation_cache') or {}
+    if cache_key in cache:
+        logger.debug(f"LLM cache hit for {watch.get('uuid')} key={cache_key[:8]}")
+        return cache[cache_key]
+
+    # Check global monthly budget before making the call
+    if is_global_token_budget_exceeded(datastore):
+        budget = get_global_token_budget_month(datastore)
+        llm_cfg = datastore.data['settings']['application'].get('llm') or {}
+        used = llm_cfg.get('tokens_this_month', 0)
+        logger.warning(
+            f"LLM evaluate_change skipped for {watch.get('uuid')}: monthly budget {budget:,} reached "
+            f"({used:,} used this month) — passing change through as important"
+        )
+        # Fail open: don't suppress notifications when budget is exhausted
+        return {'important': True, 'summary': ''}
+
+    # Check per-watch cumulative budget before making the call
+    if not _check_token_budget(watch, cfg):
+        # Already over budget — fail open (don't suppress notification)
+        return {'important': True, 'summary': ''}
+
+    url = watch.get('url', '')
+    title = watch.get('page_title') or watch.get('title') or ''
+
+    system_prompt = build_eval_system_prompt()
+    user_prompt = build_eval_prompt(
+        intent=intent,
+        diff=diff,
+        current_snapshot=current_snapshot,
+        url=url,
+        title=title,
+    )
+
+    settings = get_llm_settings(datastore)
+    try:
+        _resp = llm_client.completion(
+            model=cfg['model'],
+            messages=[
+                _cached_system(system_prompt, model=cfg['model']),
+                {'role': 'user', 'content': user_prompt},
+            ],
+            api_key=cfg.get('api_key'),
+            api_base=cfg.get('api_base'),
+            max_tokens=apply_local_token_multiplier(JSON_RESPONSE_MAX_TOKENS, cfg),
+            extra_body=_thinking_extra_body(cfg['model'], settings.thinking_budget),
+            debug=settings.debug,
+        )
+        raw, tokens = _resp[0], _resp[1]
+        input_tokens  = _resp[2] if len(_resp) > 2 else 0
+        output_tokens = _resp[3] if len(_resp) > 3 else 0
+        result = parse_eval_response(raw)
+    except Exception as e:
+        logger.warning(f"LLM evaluation failed for {watch.get('uuid')}: {e}")
+        # On failure: don't suppress the notification — pass through as important
+        watch['llm_last_tokens_used'] = 0
+        return {'important': True, 'summary': ''}
+
+    # Accumulate token usage: per-watch limit and global monthly budget
+    _check_token_budget(watch, cfg, tokens)
+    watch['llm_last_tokens_used'] = tokens
+    accumulate_global_tokens(datastore, tokens,
+                             input_tokens=input_tokens,
+                             output_tokens=output_tokens,
+                             model=cfg['model'])
+
+    # Store in cache
+    if 'llm_evaluation_cache' not in watch or watch['llm_evaluation_cache'] is None:
+        watch['llm_evaluation_cache'] = {}
+    watch['llm_evaluation_cache'][cache_key] = result
+
+    logger.debug(
+        f"LLM eval {watch.get('uuid')} (intent from {source}): "
+        f"important={result['important']} tokens={tokens} summary={result['summary'][:80]}"
+    )
+    return result
@@ -0,0 +1,216 @@
+"""
+Prompt construction for LLM evaluation calls.
+Pure functions — no side effects, fully testable.
+"""
+
+import re
+
+from .bm25_trim import trim_to_relevant
+
+_AGO_RE = re.compile(r'^\d+\s+\w+\s+ago$', re.IGNORECASE)
+
+SNAPSHOT_CONTEXT_CHARS = 3_000   # current page state excerpt sent alongside the diff
+
+
+def _annotate_moved_lines(diff_text: str) -> str:
+    """
+    Pre-process a unified diff to mark lines that appear on both the + and - sides
+    as [MOVED] rather than genuinely added/removed. This prevents the LLM from
+    incorrectly classifying repositioned content as new or deleted.
+
+    Lines are compared after stripping leading +/- and whitespace so that
+    indentation changes don't prevent matching.
+    """
+    lines = diff_text.splitlines()
+    added_texts   = {l[1:].strip().lower() for l in lines if l.startswith('+') and l[1:].strip()}
+    removed_texts = {l[1:].strip().lower() for l in lines if l.startswith('-') and l[1:].strip()}
+    moved_texts   = added_texts & removed_texts
+
+    if not moved_texts:
+        return diff_text
+
+    result = []
+    for line in lines:
+        if line.startswith(('+', '-')):
+            bare = line[1:].strip().lower()
+            if bare in moved_texts or _AGO_RE.match(line[1:].strip()):
+                result.append(f'~{line[1:]}')  # ~ prefix = moved/reordered/trivial, skip
+                continue
+        result.append(line)
+    return '\n'.join(result)
+
+
+def build_eval_prompt(intent: str, diff: str, current_snapshot: str = '',
+                      url: str = '', title: str = '') -> str:
+    """
+    Build the user message for a diff evaluation call.
+    The system prompt is kept separate (see build_eval_system_prompt).
+    """
+    parts = []
+
+    if url:
+        parts.append(f"URL: {url}")
+    if title:
+        parts.append(f"Page title: {title}")
+
+    parts.append(f"Intent: {intent}")
+
+    if current_snapshot:
+        excerpt = trim_to_relevant(current_snapshot, intent, max_chars=SNAPSHOT_CONTEXT_CHARS)
+        if excerpt:
+            parts.append(f"\nCurrent page state (relevant excerpt):\n{excerpt}")
+
+    parts.append(f"\nWhat changed (diff):\n{diff}")
+
+    return '\n'.join(parts)
+
+
+def build_eval_system_prompt() -> str:
+    return (
+        "You are a precise, reliable website-change evaluator for a monitoring tool.\n"
+        "Your job is to read a unified diff and decide whether it matches a user's stated intent.\n"
+        "Accuracy is critical — false positives waste the user's attention; false negatives miss what they care about.\n\n"
+        "Diff format:\n"
+        "- Lines starting with '+' are newly ADDED content\n"
+        "- Lines starting with '-' are REMOVED content\n"
+        "- Lines starting with ' ' (space) are unchanged context\n\n"
+        "Respond with ONLY a JSON object — no markdown, no explanation outside it:\n"
+        '{"important": true/false, "summary": "one sentence describing the relevant change, or why it doesn\'t match"}\n\n'
+        "Rules:\n"
+        "- important=true ONLY when the diff clearly and specifically matches the intent — be strict\n"
+        "- Pay close attention to direction: an intent about price drops means removed (-) prices and added (+) lower prices\n"
+        "- The user's intent always wins. If the intent explicitly asks about timestamps, numbers, counters, "
+        "thresholds, or any specific value (e.g. 'when the timestamp is greater than 1778599592', "
+        "'when stock count > 5'), evaluate the diff against that intent — do NOT dismiss it as cosmetic.\n"
+        "- Otherwise: empty, trivial, or genuinely cosmetic diffs (heartbeat timestamps, view counters, "
+        "whitespace, navigation tweaks) default to important=false\n"
+        "- For numeric comparisons in the intent, parse the values explicitly and compare them — "
+        "do not eyeball or round\n"
+        "- If the same text appears in both removed (-) and added (+) lines the content has likely just "
+        "shifted or been reordered. Treat pure reordering as important=false unless the intent "
+        "explicitly asks about order or position.\n"
+        "- Use OR logic when the intent lists multiple triggers — any one matching is sufficient\n"
+        "- When uncertain whether a change truly matches, prefer important=false and explain why in the summary\n"
+        "- Summary must be in the same language as the intent\n"
+        "- If important=false, the summary must clearly explain what changed and why it does not match"
+    )
+
+
+def build_preview_prompt(intent: str, content: str, url: str = '', title: str = '') -> str:
+    """
+    Build the user message for a live-preview extraction call.
+    Unlike build_eval_prompt (which analyses a diff), this asks the LLM to
+    extract relevant information from the *current* page content — giving the
+    user a direct answer to their intent so they can verify it makes sense
+    before saving.
+    """
+    parts = []
+    if url:
+        parts.append(f"URL: {url}")
+    if title:
+        parts.append(f"Page title: {title}")
+    parts.append(f"Intent / question: {intent}")
+    parts.append(f"\nPage content:\n{content[:6_000]}")
+    return '\n'.join(parts)
+
+
+def build_preview_system_prompt() -> str:
+    return (
+        "You are a precise, detail-oriented web page content analyst for a website monitoring tool.\n"
+        "Given the user's intent or question and the current page content, extract and directly answer "
+        "what the intent is looking for. Never guess or paraphrase — report only what the page actually contains.\n\n"
+        "Respond with ONLY a JSON object — no markdown, no explanation outside it:\n"
+        '{"found": true/false, "answer": "concise direct answer or extraction"}\n\n'
+        "Rules:\n"
+        "- found=true when the page clearly contains something relevant to the intent\n"
+        "- answer must directly address the intent with specific values where possible "
+        "(e.g. for 'current price?' → '$149.99', not 'a price is shown')\n"
+        "- answer must be in the same language as the intent\n"
+        "- Keep answer brief — one or two sentences maximum\n"
+        "- If found=false, briefly state what the page contains instead"
+    )
+
+
+def build_change_summary_prompt(diff: str, custom_prompt: str,
+                                current_snapshot: str = '', url: str = '', title: str = '') -> str:
+    """
+    Build the user message for an AI Change Summary call.
+    The user supplies their own instructions (custom_prompt); this wraps them
+    with the diff (which carries its own surrounding context via unified_diff's
+    n=3 context lines, marked '~' by _annotate_moved_lines).
+
+    NOTE: current_snapshot is accepted for caller compatibility but intentionally
+    unused. A wholesale page excerpt caused the LLM to report unchanged page
+    content (e.g. old release-note bullets) as "what changed" — hallucinations
+    drawn from the excerpt rather than the diff. The in-diff context lines give
+    the model enough surrounding text to describe each change accurately.
+    """
+    parts = []
+    if url:
+        parts.append(f"URL: {url}")
+    if title:
+        parts.append(f"Page title: {title}")
+    parts.append(f"Instructions: {custom_prompt}")
+    parts.append(f"\nWhat changed (diff):\n{_annotate_moved_lines(diff)}")
+    return '\n'.join(parts)
+
+
+def build_change_summary_system_prompt() -> str:
+    """
+    Universal, format-agnostic instructions: how to READ a diff and accuracy rules.
+    All output-format choices (prose vs JSON, sections, bullets, language, length)
+    are owned by the user prompt — including the default in
+    DEFAULT_CHANGE_SUMMARY_PROMPT — so that a user replacing the user-prompt
+    (e.g. asking for raw JSON) is not overridden by hard-coded format rules here.
+    """
+    return (
+        "You analyse a unified-diff document showing how a monitored web page changed, "
+        "and produce exactly the output the user asks for.\n\n"
+        "Rules for reading the diff:\n"
+        "- Lines starting with + are genuinely new content.\n"
+        "- Lines starting with - are genuinely removed content.\n"
+        "- Lines starting with ~ have been PRE-IDENTIFIED as moved/reordered or trivial — "
+        "the same text exists on both sides of the diff, or the line is a standalone timestamp. "
+        "Do NOT treat ~ lines as added or removed.\n\n"
+        "Accuracy: only report what the +/- lines actually contain. Never invent details, "
+        "never speculate, never add information that isn't in the diff.\n\n"
+        "Follow the user's instructions exactly — including the requested output format "
+        "(plain text, JSON, Markdown, single value, etc.), structure, language, and length. "
+        "Do not add preamble, meta-commentary, or self-introduction. Produce only the output "
+        "the user asked for — nothing before it, nothing after it."
+    )
+
+
+def build_setup_prompt(intent: str, snapshot_text: str, url: str = '') -> str:
+    """
+    Build the prompt for the one-time setup call that decides whether
+    a CSS pre-filter would improve evaluation precision.
+    """
+    excerpt = trim_to_relevant(snapshot_text, intent, max_chars=4_000)
+
+    parts = []
+    if url:
+        parts.append(f"URL: {url}")
+    parts.append(f"Intent: {intent}")
+    parts.append(f"\nPage content excerpt:\n{excerpt}")
+
+    return '\n'.join(parts)
+
+
+def build_setup_system_prompt() -> str:
+    return (
+        "You help configure a website change monitor.\n"
+        "Given a monitoring intent and a sample of the page content, decide if a CSS pre-filter "
+        "would improve evaluation precision by scoping the content to a specific structural section.\n\n"
+        "Respond with ONLY a JSON object:\n"
+        '{"needs_prefilter": true/false, "selector": "CSS selector or null", "reason": "one sentence"}\n\n'
+        "Rules:\n"
+        "- Only recommend a pre-filter when the intent references a specific structural section "
+        "(e.g. 'footer', 'sidebar', 'nav', 'header', 'main', 'article') OR the page clearly "
+        "has high-noise sections unrelated to the intent\n"
+        "- Use ONLY semantic element selectors: footer, nav, header, main, article, aside, "
+        "or attribute-based like [id*='price'], [class*='sidebar'] — NEVER positional selectors "
+        "like div:nth-child(3) or //*[2]\n"
+        "- Default to needs_prefilter=false — most intents don't need one\n"
+        "- selector must be null when needs_prefilter=false"
+    )
@@ -0,0 +1,84 @@
+"""
+Parse and validate LLM JSON responses.
+Pure functions — no side effects, fully testable.
+
+LLMs occasionally return JSON wrapped in markdown fences or with trailing
+text. This module handles those cases gracefully.
+"""
+
+import json
+import re
+
+# Positional selectors are fragile — reject them even if the LLM generates them
+_POSITIONAL_SELECTOR_RE = re.compile(
+    r'nth-child|nth-of-type|:eq\(|\[\d+\]|\/\/\*\[\d',
+    re.IGNORECASE
+)
+
+
+def _extract_json(raw: str) -> str:
+    """Strip markdown fences and extract the first JSON object."""
+    raw = raw.strip()
+    # Remove ```json ... ``` or ``` ... ``` fences
+    raw = re.sub(r'^```(?:json)?\s*', '', raw, flags=re.MULTILINE)
+    raw = re.sub(r'\s*```$', '', raw, flags=re.MULTILINE)
+    # Find the first { ... } block
+    match = re.search(r'\{.*\}', raw, re.DOTALL)
+    return match.group(0) if match else raw
+
+
+def parse_eval_response(raw: str) -> dict:
+    """
+    Parse a diff evaluation response.
+    Returns {'important': bool, 'summary': str}.
+    Falls back to important=False on any parse error.
+    """
+    try:
+        data = json.loads(_extract_json(raw))
+        return {
+            'important': bool(data.get('important', False)),
+            'summary': str(data.get('summary', '')).strip(),
+        }
+    except (json.JSONDecodeError, AttributeError):
+        return {'important': False, 'summary': ''}
+
+
+def parse_preview_response(raw: str) -> dict:
+    """
+    Parse a live-preview extraction response.
+    Returns {'found': bool, 'answer': str}.
+    Falls back to found=False on any parse error.
+    """
+    try:
+        data = json.loads(_extract_json(raw))
+        return {
+            'found': bool(data.get('found', False)),
+            'answer': str(data.get('answer', '')).strip(),
+        }
+    except (json.JSONDecodeError, AttributeError):
+        return {'found': False, 'answer': ''}
+
+
+def parse_setup_response(raw: str) -> dict:
+    """
+    Parse a setup/pre-filter decision response.
+    Returns {'needs_prefilter': bool, 'selector': str|None, 'reason': str}.
+    Rejects positional selectors even if the LLM generates them.
+    """
+    try:
+        data = json.loads(_extract_json(raw))
+        needs = bool(data.get('needs_prefilter', False))
+        selector = data.get('selector') or None
+
+        # Sanitise: reject positional selectors
+        if selector and _POSITIONAL_SELECTOR_RE.search(selector):
+            selector = None
+            needs = False
+
+        return {
+            'needs_prefilter': needs,
+            'selector': selector if needs else None,
+            'reason': str(data.get('reason', '')).strip(),
+        }
+    except (json.JSONDecodeError, AttributeError):
+        return {'needs_prefilter': False, 'selector': None, 'reason': ''}
@@ -0,0 +1,18 @@
+"""
+Shared UI placeholder strings for LLM fields.
+
+Used by WTForms field definitions in forms.py and blueprint/tags/form.py.
+Templates use their own _()-translated variants but should stay in sync with these.
+"""
+
+# llm_intent field — placeholder text for per-watch context
+LLM_INTENT_WATCH_PLACEHOLDER = (
+    "e.g. Alert me when the price drops below $300, or a new product is launched. "
+    "Ignore footer and navigation changes."
+)
+
+# llm_intent field — placeholder text for tag/group context
+LLM_INTENT_TAG_PLACEHOLDER = (
+    "e.g. Flag price changes or new product launches across all watches in this group"
+)
+
@@ -2,6 +2,7 @@ from os import getenv
 from copy import deepcopy

 from changedetectionio.blueprint.rss import RSS_FORMAT_TYPES, RSS_CONTENT_FORMAT_DEFAULT
+from changedetectionio.model.Tags import TagsDict

 from changedetectionio.notification import (
    default_notification_body,
@@ -68,7 +69,10 @@ class model(dict):
                    'schema_version' : 0,
                    'shared_diff_access': False,
                    'strip_ignored_lines': False,
-                    'tags': {}, #@todo use Tag.model initialisers
+                    'tags': None,  # Initialized in __init__ with real datastore_path
+                    # All LLM settings now live nested under application.llm.* (post-migration update_31).
+                    # Defaults come from LLMSettings.model_validate({}) at read time —
+                    # no need to pre-seed an empty {} here.
                    'webdriver_delay': None , # Extra delay in seconds before extracting text
                    'ui': {
                        'use_page_title_in_list': True,
@@ -80,10 +84,16 @@ class model(dict):
            }
        }

-    def __init__(self, *arg, **kw):
+    def __init__(self, *arg, datastore_path=None, **kw):
        super(model, self).__init__(*arg, **kw)
+        # Capture any tags data passed in before base_config overwrites the structure
+        existing_tags = self.get('settings', {}).get('application', {}).get('tags') or {}
        # CRITICAL: deepcopy to avoid sharing mutable objects between instances
        self.update(deepcopy(self.base_config))
+        # TagsDict requires the real datastore_path at runtime (cannot be set at class-definition time)
+        if datastore_path is None:
+            raise ValueError("App.model() requires 'datastore_path' keyword argument")
+        self['settings']['application']['tags'] = TagsDict(existing_tags, datastore_path=datastore_path)


 def parse_headers_from_text_file(filepath):
@@ -0,0 +1,65 @@
+"""
+Validation/typing layer for the LLM config dict stored at
+    datastore.data['settings']['application']['llm']
+
+Storage stays a plain dict (orjson-serialized). This model is hydrated on read
+(model_validate) and dumped on write (model_dump). WTForms field names match
+the storage field names exactly — no aliases needed.
+"""
+from typing import ClassVar, Tuple
+
+from pydantic import BaseModel, ConfigDict
+
+
+LLM_DEFAULT_THINKING_BUDGET = 0
+LLM_DEFAULT_MAX_SUMMARY_TOKENS = 3000
+LLM_DEFAULT_LOCAL_TOKEN_MULTIPLIER = 5
+LLM_DEFAULT_MAX_INPUT_CHARS = 100_000
+LLM_DEFAULT_BUDGET_ACTION = 'skip_llm'
+
+
+class LLMSettings(BaseModel):
+    # extra='forbid' rejects any key that isn't a declared field with a
+    # ValidationError. Loud failure forces new form fields to be declared here
+    # before they can land in storage — closes the CWE-915 mass-assignment class
+    # of bugs (see GHSA-h3x5-5j56-hm2j for the canonical example).
+    model_config = ConfigDict(extra='forbid')
+
+    enabled: bool = True
+    debug: bool = False
+    override_diff_with_summary: bool = True
+    restock_use_fallback_extract: bool = True
+    thinking_budget: int = LLM_DEFAULT_THINKING_BUDGET
+    max_summary_tokens: int = LLM_DEFAULT_MAX_SUMMARY_TOKENS
+    budget_action: str = LLM_DEFAULT_BUDGET_ACTION
+    change_summary_default: str = ''
+    token_budget_month: int = 0
+    max_input_chars: int = LLM_DEFAULT_MAX_INPUT_CHARS
+    # Per-watch per-period token cap; read by _check_token_budget() in evaluator.py.
+    # 0 means unlimited. Once a watch's usage within the current period hits this cap,
+    # AI evaluation is skipped for it until the period rolls over. Period is currently
+    # hard-coded to month (matches the global counter rollover); name is period-agnostic
+    # to leave room for a configurable period (day/week/month) later.
+    max_tokens_per_count_period: int = 0
+
+    model: str = ''
+    api_key: str = ''
+    api_base: str = ''
+    provider_kind: str = ''
+    local_token_multiplier: int = LLM_DEFAULT_LOCAL_TOKEN_MULTIPLIER
+
+    tokens_total_cumulative: int = 0
+    tokens_this_month: int = 0
+    tokens_month_key: str = ''
+    cost_usd_total_cumulative: float = 0.0
+    cost_usd_this_month: float = 0.0
+
+    # Provider-connection fields wiped on /llm/clear and when the model is emptied.
+    CONNECTION_FIELDS: ClassVar[Tuple[str, ...]] = (
+        'model', 'api_key', 'api_base', 'provider_kind', 'local_token_multiplier',
+    )
+    # Runtime-managed counters — form submissions must never overwrite these.
+    PROTECTED_FIELDS: ClassVar[Tuple[str, ...]] = (
+        'tokens_total_cumulative', 'tokens_this_month', 'tokens_month_key',
+        'cost_usd_total_cumulative', 'cost_usd_this_month',
+    )
@@ -0,0 +1,239 @@
+# Pydantic Migration
+
+Plan for incrementally moving the app's storage dicts behind Pydantic models. Driven by
+security (CWE-915 mass-assignment, see [GHSA-h3x5-5j56-hm2j][advisory]) and schema
+enforcement, not just type tidying.
+
+[advisory]: https://github.com/dgtlmoon/changedetection.io/security/advisories/GHSA-h3x5-5j56-hm2j
+
+## The goal
+
+Every form/API endpoint that mutates a stored dict should validate input against a
+declared schema before writing. `extra='forbid'` rejects unknown keys — so an attacker
+POSTing extra fields like `uuid=…`, `last_checked=…`, `history=[…]` can't smuggle them
+into storage. Per-route allowlists work but rot; one declared schema per stored shape
+doesn't.
+
+## Prefer a migration over permanent complexity
+
+If you're about to add a compatibility shim, an alias, a backward-compat fallback, or a
+"handle both old and new shape" branch — stop and ask whether a one-time `update_N`
+migration solves the same problem by *renaming the stored data*. A migration runs once
+per install; the shim lives in the code forever and every future contributor has to
+understand it.
+
+Concrete example from this PR: the original design used `Field(alias='llm_X')` so
+Pydantic could accept both the legacy form-field name (`llm_model`) and the new
+storage name (`model`). That alias survived every read/write for the life of the app
+and introduced a subtle `model_dump(by_alias=True)` merge bug. The simpler answer was
+to rename the form fields to match the storage names (an in-PR rename, no migration
+needed since storage was new), drop the aliases entirely, and delete ~25 lines of
+plumbing. **Pay once with a migration; don't pay forever with complexity.**
+
+Same principle applies the moment you find yourself writing `dict.get(new_key) or
+dict.get(old_key)`. That's a migration in disguise — write the migration instead.
+
+## Architecture choice: validator at the boundary, not domain model
+
+There are two ways to use Pydantic. Pick one per slice — they are not interchangeable.
+
+**Pydantic-as-validator (what we do).** Storage stays a plain dict. A `BaseModel`
+validates input at the boundary, dumps back to a dict. No call-site changes; the
+existing `watch['x']` dict access keeps working everywhere.
+
+**Pydantic-as-domain-model.** Replace `dict` inheritance with `BaseModel`. ~190 call
+sites switch from `watch['x']` to `watch.x`. Much bigger blast radius, defers the
+security win. Not what we're doing right now.
+
+The CWE-915 fix only needs the validator pattern. Domain-model replacement is a
+separate, later project.
+
+## The template (LLMSettings)
+
+The first migrated slice. Use as the reference for the next one.
+
+**Match the WTForms field names to the storage / Pydantic field names** so the
+form-input dict and the storage dict have the same key shape. No aliases, no
+`populate_by_name=True`, no `by_alias=True` merge gymnastics. Only reach for
+`Field(alias=…)` if you genuinely cannot rename the form field (rare).
+
+`model/LLMSettings.py`:
+
+```python
+class LLMSettings(BaseModel):
+    model_config = ConfigDict(extra='forbid')
+
+    enabled: bool = True
+    model: str = ''
+    ...
+
+    # System-managed counters
+    tokens_total_cumulative: int = 0
+    ...
+
+    # Field groups
+    CONNECTION_FIELDS: ClassVar[Tuple[str, ...]] = ('model', 'api_key', ...)
+    PROTECTED_FIELDS:  ClassVar[Tuple[str, ...]] = ('tokens_total_cumulative', ...)
+```
+
+Boundary pattern at the route handler:
+
+```python
+# Read
+settings = LLMSettings.model_validate(
+    datastore.data['settings']['application'].get('llm') or {}
+)
+
+# Merge form input
+form_input = dict(form.data.get('llm') or {})
+for protected in LLMSettings.PROTECTED_FIELDS:
+    form_input.pop(protected, None)  # counters never come from form
+merged = LLMSettings.model_validate({**settings.model_dump(), **form_input})
+
+# Write — re-validates the schema on every write
+datastore.data['settings']['application']['llm'] = merged.model_dump()
+```
+
+## Unresolved architectural decisions
+
+Two decisions need answers before the `WatchInput` slice. They're not blockers for `App.py`.
+
+### OpenAPI spec vs Pydantic model — who's source of truth?
+
+Today: `docs/api-spec.yaml` declares the Watch/Tag shape; `model/schema_utils.py` reads
+it to compute readonly fields; the API layer validates against it; the model layer is a
+plain dict that doesn't know about either. When `WatchInput` lands, that's a third
+shape declaration.
+
+Two ways to live:
+- **Pydantic is source.** Generate / sync `api-spec.yaml` from the model
+  (e.g. via `model_json_schema()`). One declaration, multiple consumers. Long-term
+  right answer; needs tooling.
+- **Parallel sources with discipline.** Hand-keep them aligned. Faster to ship but
+  drift is inevitable — that's the bug class we're already trying to close.
+
+Recommendation: start parallel (keep `api-spec.yaml` for now), but write Watch's
+Pydantic model so it could be the eventual single source. Don't *invent* a new
+field shape — match the spec.
+
+### Plugin / processor_config_* extensibility
+
+`processor_config_restock_diff` (and future processor configs) are written by
+plugins, not the core. `extra='forbid'` on a Watch input model would reject them.
+
+Options:
+- **Per-processor sub-models.** Each plugin owns its `<Processor>Settings` Pydantic
+  model; Watch input validates only core fields, processor configs validate
+  separately at their own boundary (the per-watch `restock_diff.json`, etc.).
+- **Opaque pass-through.** Watch input model treats `processor_config_*` as a
+  declared dict-typed field. Loses per-key validation but preserves the
+  plugin-extensibility contract.
+
+Recommendation: per-processor sub-models. Matches the file split already done in
+`update_30` (separate `restock_diff.json` per watch).
+
+## Migration order
+
+| Target | Difficulty | Value | Status |
+|---|---|---|---|
+| `LLMSettings` | low | medium | done (this PR) |
+| `App.py` → `AppSettings` (nested) | low | medium | next |
+| `WatchInput` (form/API validator) | medium | **HIGH — closes [GHSA-h3x5-5j56-hm2j][advisory]** | next-next |
+| `TagInput` (form/API validator) | medium | medium | after Watch |
+| `watch_base(dict)` → `BaseModel` | very high | high | separate multi-PR project, much later |
+
+`Tags.py` (TagsDict), `persistence.py`, `schema_utils.py` are not data models — leave alone.
+
+### Concrete next steps
+
+1. **`App.py`.** Pure dict tree under `settings.{application,requests,headers}`. Define
+   nested `BaseModel`s; `LLMSettings` slots in as the existing sub-tree. No call-site
+   churn — just the global settings POST handler. Sets the pattern for nested models.
+
+2. **`WatchInput` BaseModel** for `blueprint/ui/edit.py:225` and `api/Watch.py`. Replace:
+   ```python
+   datastore.data['watching'][uuid].update(form.data)  # CWE-915
+   ```
+   with:
+   ```python
+   validated = WatchInput.model_validate(form.data)
+   datastore.data['watching'][uuid].update(validated.model_dump())
+   ```
+   Closes the unpatched advisory. Should be a security-tagged commit referencing the GHSA.
+
+3. **`TagInput` BaseModel** — same pattern, smaller.
+
+## Gotchas discovered
+
+These cost real debugging time in the LLMSettings PR. Worth knowing before the next slice.
+
+### `extra='forbid'` is the right default
+
+`extra='ignore'` silently drops unknowns and hides developer mistakes (add a form field,
+forget to declare it on the model, your feature appears to work until you reload). `forbid`
+fails loudly. `allow` defeats the purpose entirely — it's how injection succeeds.
+
+### Don't use Field aliases unless you actually need them
+
+The LLMSettings PR originally used `alias='llm_X'` to bridge llm_-prefixed WTForms
+names to stripped storage names. That created a documented gotcha: with
+`extra='forbid'`, having both `model` and `llm_model` in the same input dict is a
+`ValidationError`, and merging existing-storage-dump with form input required
+`by_alias=True` to keep both sides on the alias shape. We fixed it by renaming the
+form fields to match the storage field names. **Match the form to the model
+upfront and you avoid the whole class of merge bugs.**
+
+### Round-trip counters through the model, don't mutate the dict
+
+If runtime code (e.g. a token accumulator) writes to the storage dict directly, the
+schema is bypassed. Load → mutate instance attributes → `model_dump()` → write back.
+This re-validates on every write and prevents drift.
+
+### Per-call validation needs strict + tolerant modes? Don't.
+
+You might be tempted to validate form input strictly but allow extras in storage
+hydration. Don't — `extra='forbid'` everywhere means storage drift is impossible. If
+something put unknown keys in storage, you want loud failure, not silent acceptance.
+
+### Migrations are convention-based by accident if you let them be
+
+`for k in list(d) if k.startswith('llm_')` is shorter than an explicit list but
+silently catches any future flat `llm_*` key. Migrations are forever — prefer an
+explicit allowlist of keys to move, even if it's verbose.
+
+## What NOT to do
+
+- Don't add custom helper methods (`dump_without_connection()`, `clear_X()`) when stock
+  `model_dump(exclude=set(FIELDS))` works. The standard idiom is more readable and
+  zero-line.
+- Don't push security/business logic into the model (e.g. SSRF guards, credential-exfil
+  checks). The model owns field shape and validation. Route handlers own
+  policy. Mixing them dilutes both.
+- Don't make `get_X_config()` return a Pydantic instance if callers do dict-style access.
+  Either migrate all call sites (high-touch) or keep returning a dict and let the model
+  be the validation/dump layer only.
+- Don't `model_copy(update=...)` without re-validating. It doesn't coerce types or
+  enforce `extra='forbid'`. Use `model_validate({**old.model_dump(), **updates})` for
+  strict merges.
+
+## Required for each new slice
+
+Each migration PR should ship:
+
+- `model/<Thing>Settings.py` (or input model) — declared schema, `extra='forbid'`,
+  field aliases if there's a name mismatch between form and storage.
+- `store/updates.py:update_N` if the storage shape changes. Pure dict-shuffling, no
+  Pydantic import (migrations should not depend on the model — model evolves
+  independently).
+- `tests/unit/test_<thing>.py` — unit coverage of the model itself: defaults,
+  alias merge, type coercion, `extra='forbid'` rejection, dump shapes.
+- All runtime callers updated to go through `get_<thing>_settings(datastore)` or
+  equivalent, not raw dict reads.
+
+## Reference
+
+- `model/LLMSettings.py` — the template
+- `tests/unit/test_llm_settings.py` — model unit-test template
+- `store/updates.py:update_31` — schema migration template
+- `blueprint/settings/__init__.py` (POST handler) — boundary-validation template
+- `llm/evaluator.py:accumulate_global_tokens` — instance-mutation-then-dump-back template
@@ -46,11 +46,26 @@ class model(EntityPersistenceMixin, watch_base):
        super(model, self).__init__(*arg, **kw)

        self['overrides_watch'] = kw.get('default', {}).get('overrides_watch')
+        self['url_match_pattern'] = kw.get('default', {}).get('url_match_pattern', '')

        if kw.get('default'):
            self.update(kw['default'])
            del kw['default']

+    def matches_url(self, url: str) -> bool:
+        """Return True if this tag should be auto-applied to the given watch URL.
+
+        Wildcard patterns (*,?,[ ) use fnmatch; anything else is a case-insensitive
+        substring match. Returns False if no pattern is configured.
+        """
+        import fnmatch
+        pattern = self.get('url_match_pattern', '').strip()
+        if not pattern or not url:
+            return False
+        if any(c in pattern for c in ('*', '?', '[')):
+            return fnmatch.fnmatch(url.lower(), pattern.lower())
+        return pattern.lower() in url.lower()
+
    # _save_to_disk() method provided by EntityPersistenceMixin
    # commit() and _get_commit_data() methods inherited from watch_base
    # Tag uses default _get_commit_data() (includes all keys)
@@ -0,0 +1,39 @@
+import os
+import shutil
+from pathlib import Path
+from loguru import logger
+
+_SENTINEL = object()
+
+
+class TagsDict(dict):
+    """Dict subclass that removes the corresponding tag.json file when a tag is deleted."""
+
+    def __init__(self, *args, datastore_path: str | os.PathLike, **kwargs) -> None:
+        self._datastore_path = Path(datastore_path)
+        super().__init__(*args, **kwargs)
+
+    def __delitem__(self, key: str) -> None:
+        super().__delitem__(key)
+        tag_dir = self._datastore_path / key
+        tag_json_file = tag_dir / "tag.json"
+        if not os.path.exists(tag_json_file):
+            logger.critical(f"Aborting deletion of directory '{tag_dir}' because '{tag_json_file}' does not exist.")
+            return
+        try:
+            shutil.rmtree(tag_dir)
+            logger.info(f"Deleted tag directory for tag {key!r}")
+        except FileNotFoundError:
+            pass
+        except OSError as e:
+            logger.error(f"Failed to delete tag directory for tag {key!r}: {e}")
+
+    def pop(self, key: str, default=_SENTINEL):
+        """Remove and return tag, deleting its tag.json file. Raises KeyError if missing and no default given."""
+        if key in self:
+            value = self[key]
+            del self[key]
+            return value
+        if default is _SENTINEL:
+            raise KeyError(key)
+        return default
@@ -43,6 +43,11 @@ from ..html_tools import TRANSLATE_WHITESPACE_TABLE
 FAVICON_RESAVE_THRESHOLD_SECONDS=86400
 BROTLI_COMPRESS_SIZE_THRESHOLD = int(os.getenv('SNAPSHOT_BROTLI_COMPRESSION_THRESHOLD', 1024*20))

+# Module-level favicon filename cache: data_dir → basename (or None)
+# Keyed by data_dir so it survives Watch object recreation, deepcopy, and concurrent requests.
+# Invalidated explicitly in bump_favicon() when a new favicon is saved.
+_FAVICON_FILENAME_CACHE: dict = {}
+
 minimum_seconds_recheck_time = int(os.getenv('MINIMUM_SECONDS_RECHECK_TIME', 3))
 mtable = {'seconds': 1, 'minutes': 60, 'hours': 3600, 'days': 86400, 'weeks': 86400 * 7}

@@ -383,6 +388,25 @@ class model(EntityPersistenceMixin, watch_base):

        return self.get('fetch_backend')

+    @property
+    def fetcher_supports_screenshots(self):
+        """Return True if the fetcher configured for this watch supports screenshots.
+
+        Resolves 'system' via self._datastore, then checks supports_screenshots on
+        the actual fetcher class. Works for built-in and plugin fetchers alike.
+        """
+        from changedetectionio import content_fetchers
+
+        fetcher_name = self.get_fetch_backend  # already handles is_pdf → html_requests
+        if not fetcher_name or fetcher_name == 'system':
+            fetcher_name = self._datastore['settings']['application'].get('fetch_backend', 'html_requests')
+
+        fetcher_class = getattr(content_fetchers, fetcher_name, None)
+        if fetcher_class is None:
+            return False
+
+        return bool(getattr(fetcher_class, 'supports_screenshots', False))
+
    @property
    def is_pdf(self):
        url = str(self.get("url") or "").lower()
@@ -441,22 +465,21 @@ class model(EntityPersistenceMixin, watch_base):
                    if ',' in i:
                        k, v = i.strip().split(',', 2)

-                        # The index history could contain a relative path, so we need to make the fullpath
-                        # so that python can read it
-                        # Cross-platform: check for any path separator (works on Windows and Unix)
-                        if os.sep not in v and '/' not in v and '\\' not in v:
-                            # Relative filename only, no path separators
-                            v = os.path.join(self.data_dir, v)
-                        else:
-                            # It's possible that they moved the datadir on older versions
-                            # So the snapshot exists but is in a different path
-                            # Cross-platform: use os.path.basename instead of split('/')
-                            snapshot_fname = os.path.basename(v)
-                            proposed_new_path = os.path.join(self.data_dir, snapshot_fname)
-                            if not os.path.exists(v) and os.path.exists(proposed_new_path):
-                                v = proposed_new_path
+                        # Always resolve history entries to within the watch's own data directory.
+                        # Entries restored from backup could contain absolute or traversal paths —
+                        # never trust them. Use realpath to also block symlink-based escapes.
+                        safe_data_dir = os.path.realpath(self.data_dir)
+                        snapshot_fname = os.path.basename(v.strip())
+                        resolved_path = os.path.realpath(os.path.join(self.data_dir, snapshot_fname))

-                        tmp_history[k] = v
+                        if not resolved_path.startswith(safe_data_dir + os.sep) and resolved_path != safe_data_dir:
+                            logger.warning(f"Skipping unsafe history entry for {self.get('uuid')}: {v!r}")
+                            continue
+
+                        if not os.path.exists(resolved_path):
+                            continue
+
+                        tmp_history[k] = resolved_path

        if len(tmp_history):
            self.__newest_history_key = list(tmp_history.keys())[-1]
@@ -539,6 +562,15 @@ class model(EntityPersistenceMixin, watch_base):
        if not filepath:
            filepath = self.history[timestamp]

+        # Confine every read to the watch's own data directory — defence in depth
+        # against any path that bypasses the history parser (e.g. direct filepath= callers).
+        # Set HISTORY_SNAPSHOT_FILE_ALLOW_OUTSIDE_WATCH_DATADIR=true to disable (not recommended).
+        if self.data_dir and not strtobool(os.getenv('HISTORY_SNAPSHOT_FILE_ALLOW_OUTSIDE_WATCH_DATADIR', 'False')):
+            safe_data_dir = os.path.realpath(self.data_dir)
+            resolved = os.path.realpath(filepath)
+            if not (resolved.startswith(safe_data_dir + os.sep) or resolved == safe_data_dir):
+                raise PermissionError(f"Snapshot path {filepath!r} is outside the watch data directory")
+
        # Check if binary file (image, PDF, etc.)
        # Binary files are NEVER saved with .br compression, only text files are
        binary_extensions = ('.png', '.jpg', '.jpeg', '.gif', '.webp', '.pdf', '.bin', '.jfif')
@@ -774,24 +806,50 @@ class model(EntityPersistenceMixin, watch_base):
        # Also in the case that the file didnt exist
        return True

-    def bump_favicon(self, url, favicon_base_64: str) -> None:
+    def bump_favicon(self, url, favicon_base_64: str, mime_type: str = None) -> None:
        from urllib.parse import urlparse
        import base64
        import binascii
-        decoded = None
+        import re

-        if url:
+        MAX_FAVICON_BYTES = 1 * 1024 * 1024  # 1 MB
+
+        MIME_TO_EXT = {
+            'image/png': 'png',
+            'image/x-icon': 'ico',
+            'image/vnd.microsoft.icon': 'ico',
+            'image/jpeg': 'jpg',
+            'image/gif': 'gif',
+            'image/svg+xml': 'svg',
+            'image/webp': 'webp',
+            'image/bmp': 'bmp',
+        }
+
+        extension = None
+
+        # If the caller already resolved the MIME type (e.g. from blob.type or a data URI),
+        # use that directly — it's more reliable than guessing from a URL path.
+        if mime_type:
+            extension = MIME_TO_EXT.get(mime_type.lower().split(';')[0].strip(), None)
+
+        # Fall back to extracting extension from URL path, unless it's a data URI.
+        if not extension and url and not url.startswith('data:'):
            try:
                parsed = urlparse(url)
                filename = os.path.basename(parsed.path)
-                (base, extension) = filename.lower().strip().rsplit('.', 1)
+                (_base, ext) = filename.lower().strip().rsplit('.', 1)
+                extension = ext
            except ValueError:
-                logger.error(f"UUID: {self.get('uuid')} Cant work out file extension from '{url}'")
-                return None
-        else:
-            # Assume favicon.ico
-            base = "favicon"
-            extension = "ico"
+                logger.warning(f"UUID: {self.get('uuid')} Cant work out file extension from '{url}', defaulting to ico")
+
+        # Handle data URIs: extract MIME type from the URI itself when not already known
+        if not extension and url and url.startswith('data:'):
+            m = re.match(r'^data:([^;]+);base64,', url)
+            if m:
+                extension = MIME_TO_EXT.get(m.group(1).lower(), None)
+
+        if not extension:
+            extension = 'ico'

        fname = os.path.join(self.data_dir, f"favicon.{extension}")

@@ -800,58 +858,50 @@ class model(EntityPersistenceMixin, watch_base):
            decoded = base64.b64decode(favicon_base_64, validate=True)
        except (binascii.Error, ValueError) as e:
            logger.warning(f"UUID: {self.get('uuid')} FavIcon save data (Base64) corrupt? {str(e)}")
-        else:
-            if decoded:
-                try:
-                    with open(fname, 'wb') as f:
-                        f.write(decoded)
+            return None

-                    # Invalidate favicon filename cache
-                    if hasattr(self, '_favicon_filename_cache'):
-                        delattr(self, '_favicon_filename_cache')
+        if len(decoded) > MAX_FAVICON_BYTES:
+            logger.warning(f"UUID: {self.get('uuid')} Favicon too large ({len(decoded)} bytes), skipping")
+            return None

-                    # A signal that could trigger the socket server to update the browser also
-                    watch_check_update = signal('watch_favicon_bump')
-                    if watch_check_update:
-                        watch_check_update.send(watch_uuid=self.get('uuid'))
+        try:
+            with open(fname, 'wb') as f:
+                f.write(decoded)

-                except Exception as e:
-                    logger.warning(f"UUID: {self.get('uuid')} error saving FavIcon to {fname} - {str(e)}")
+            # Invalidate module-level favicon filename cache for this watch
+            _FAVICON_FILENAME_CACHE.pop(self.data_dir, None)
+
+            # A signal that could trigger the socket server to update the browser also
+            watch_check_update = signal('watch_favicon_bump')
+            if watch_check_update:
+                watch_check_update.send(watch_uuid=self.get('uuid'))
+
+        except Exception as e:
+            logger.warning(f"UUID: {self.get('uuid')} error saving FavIcon to {fname} - {str(e)}")
+            return None

        # @todo - Store some checksum and only write when its different
        logger.debug(f"UUID: {self.get('uuid')} updated favicon to at {fname}")

    def get_favicon_filename(self) -> str | None:
        """
-        Find any favicon.* file in the current working directory
-        and return the contents of the newest one.
+        Find any favicon.* file in the watch data directory.

-        MEMORY LEAK FIX: Cache the result to avoid repeated glob.glob() operations.
-        glob.glob() causes millions of fnmatch allocations when called for every watch on page load.
+        Uses a module-level cache keyed by data_dir to survive Watch object recreation,
+        deepcopy (which drops instance attrs), and concurrent request races.
+        Invalidated by bump_favicon() when a new favicon is saved.

        Returns:
-            str: Basename of the newest favicon file, or None if not found.
+            str: Basename of the favicon file, or None if not found.
        """
-        # Check cache first (prevents 26M+ allocations from repeated glob operations)
-        cache_key = '_favicon_filename_cache'
-        if hasattr(self, cache_key):
-            return getattr(self, cache_key)
+        if self.data_dir in _FAVICON_FILENAME_CACHE:
+            return _FAVICON_FILENAME_CACHE[self.data_dir]

        import glob
-
-        # Search for all favicon.* files
        files = glob.glob(os.path.join(self.data_dir, "favicon.*"))
-
-        if not files:
-            result = None
-        else:
-            # Find the newest by modification time
-            newest_file = max(files, key=os.path.getmtime)
-            result = os.path.basename(newest_file)
-
-        # Cache the result
-        setattr(self, cache_key, result)
-        return result
+        fname = os.path.basename(files[0]) if files else None
+        _FAVICON_FILENAME_CACHE[self.data_dir] = fname
+        return fname

    def get_screenshot_as_thumbnail(self, max_age=3200):
        """Return path to a square thumbnail of the most recent screenshot.
@@ -959,6 +1009,40 @@ class model(EntityPersistenceMixin, watch_base):
        return False


+    @staticmethod
+    def _llm_summary_prompt_hash(prompt: str) -> str:
+        """8-char hex hash of the prompt — used to detect when the prompt changes."""
+        import hashlib
+        return hashlib.md5(prompt.encode('utf-8', errors='replace')).hexdigest()[:8]
+
+    def get_llm_diff_summary(self, from_version, to_version, prompt: str = '') -> str:
+        """Return the cached AI Change Summary for this from→to + prompt combination, or ''.
+
+        The prompt hash is embedded in the filename so that a changed prompt
+        automatically produces a cache miss and triggers regeneration.
+        """
+        prompt_hash = self._llm_summary_prompt_hash(prompt)
+        fname = os.path.join(self.data_dir, f'change-summary-{from_version}-to-{to_version}-{prompt_hash}.txt')
+        if not os.path.isfile(fname):
+            logger.debug(f"LLM cached diff summary '{fname}' NOT found")
+            return ''
+        with open(fname, 'r', encoding='utf-8') as f:
+            logger.debug(f"LLM cached diff summary '{fname}' FOUND")
+            return f.read().strip()
+
+    def save_llm_diff_summary(self, summary: str, from_version, to_version, prompt: str = ''):
+        """Persist the AI Change Summary keyed by version pair + prompt hash."""
+        self.ensure_data_dir_exists()
+        prompt_hash = self._llm_summary_prompt_hash(prompt)
+        fname = os.path.join(self.data_dir, f'change-summary-{from_version}-to-{to_version}-{prompt_hash}.txt')
+        tmp = fname + '.tmp'
+        try:
+            with open(tmp, 'w', encoding='utf-8') as f:
+                f.write(summary)
+            os.replace(tmp, fname)
+        except OSError as e:
+            logger.warning(f"Could not write LLM summary cache {fname}: {e}")
+
    def pause(self):
        self['paused'] = True

@@ -982,6 +1066,7 @@ class model(EntityPersistenceMixin, watch_base):
        Prepare watch data for commit.

        Excludes processor_config_* keys (stored in separate files).
+        Excludes __-prefixed keys (transient in-memory state — must not persist to disk).
        Normalizes browser_steps to empty list if no meaningful steps.
        """
        import copy
@@ -995,8 +1080,11 @@ class model(EntityPersistenceMixin, watch_base):
        else:
            snapshot = dict(self)

-        # Exclude processor config keys (stored separately)
-        watch_dict = {k: copy.deepcopy(v) for k, v in snapshot.items() if not k.startswith('processor_config_')}
+        # Exclude processor config keys (stored separately) and __-prefixed transient keys
+        watch_dict = {
+            k: copy.deepcopy(v) for k, v in snapshot.items()
+            if not k.startswith('processor_config_') and not k.startswith('__')
+        }

        # Normalize browser_steps: if no meaningful steps, save as empty list
        if not self.has_browser_steps:
@@ -1182,18 +1270,13 @@ class model(EntityPersistenceMixin, watch_base):
    def compile_error_texts(self, has_proxies=None):
        """Compile error texts for this watch.
        Accepts has_proxies parameter to ensure it works even outside app context"""
-        from flask import url_for
+        from flask import url_for, has_request_context
        from markupsafe import Markup

        output = []  # Initialize as list since we're using append
        last_error = self.get('last_error','')

-        try:
-            url_for('settings.settings_page')
-        except Exception as e:
-            has_app_context = False
-        else:
-            has_app_context = True
+        has_app_context = has_request_context()

        # has app+request context, we can use url_for()
        if has_app_context:
@@ -186,7 +186,13 @@ class watch_base(dict):
            'consecutive_filter_failures': 0,  # Every time the CSS/xPath filter cannot be located, reset when all is fine.
            'content-type': None,
            'date_created': None,
+            'extract_lines_containing': [],  # Keep only lines containing these substrings (plain text, case-insensitive)
            'extract_text': [],  # Extract text by regex after filters
+            # LLM intent-based evaluation
+            'llm_intent': '',                # Plain-English description of what the user cares about (change filter)
+            'llm_change_summary': '',        # Prompt for AI Change Summary — replaces {{ diff }} in notifications
+            'llm_prefilter': None,           # CSS selector derived at setup time (semantic only, e.g. "footer")
+            'llm_evaluation_cache': {},      # {sha256(intent+diff): {important, summary}} - evaluated once, cached
            'fetch_backend': 'system',  # plaintext, playwright etc
            'fetch_time': 0.0,
            'filter_failure_notification_send': strtobool(os.getenv('FILTER_FAILURE_NOTIFICATION_SEND_DEFAULT', 'True')),
@@ -329,20 +335,22 @@ class watch_base(dict):
        if self.__watch_was_edited:
            return  # Already marked as edited

+        # __-prefixed keys are transient in-memory state (e.g. __check_status set by
+        # set_watch_minitext_status). They never persist to disk and must not trigger
+        # the edited flag — otherwise just observing a check in progress would force
+        # the next run to bypass the unchanged-content skip.
+        if isinstance(key, str) and key.startswith('__'):
+            return
+
        # Import from shared schema utilities (no circular dependency)
-        from .schema_utils import get_readonly_watch_fields
-        readonly_fields = get_readonly_watch_fields()
+        from .schema_utils import get_readonly_watch_fields, SYSTEM_MANAGED_NON_SPEC_FIELDS

-        # Additional system-managed fields not in OpenAPI spec (yet)
-        # These are set by processors/workers and should not trigger edited flag
-        additional_system_fields = {
-            'last_check_status',  # Set by processors
-            'restock',  # Set by restock processor
-            'last_viewed',  # Set by mark_all_viewed endpoint
-        }
-
-        # Only mark as edited if this is a user-writable field
-        if key not in readonly_fields and key not in additional_system_fields:
+        # `last_viewed` is set internally by mark_all_viewed and shouldn't flag the watch as
+        # edited, but is not in SYSTEM_MANAGED_NON_SPEC_FIELDS because it IS user-writable via
+        # the UpdateWatch schema (the API path).
+        if (key not in get_readonly_watch_fields()
+                and key != 'last_viewed'
+                and key not in SYSTEM_MANAGED_NON_SPEC_FIELDS):
            self.__watch_was_edited = True

    def __setitem__(self, key, value):
@@ -8,6 +8,35 @@ Shared by both the model layer and API layer to avoid circular dependencies.
 import functools


+# Watch fields written by workers/processors that are NOT part of the public OpenAPI spec.
+#
+# These fields exist on a watch dict at runtime but are internal implementation details
+# (skip-cache hashes, last-check status strings, LLM runtime state, etc.). Used by:
+#   - model/__init__.py: don't trigger the "edited" flag when these are written internally
+#   - api/Watch.py: strip from GET responses and silently discard from PUT/POST inputs
+#                   so that a GET → PUT round trip doesn't trip the unknown-field validator
+#
+# `last_viewed` is intentionally NOT included: it's set internally by mark_all_viewed BUT
+# is also explicitly writable via the UpdateWatch schema (see api/Watch.py valid_fields).
+SYSTEM_MANAGED_NON_SPEC_FIELDS = frozenset({
+    'last_check_status',           # Set by processors
+    'last_filter_config_hash',     # text_json_diff internal skip-cache
+    'restock',                     # Set by restock processor
+    '_llm_result',                 # LLM runtime — populated by evaluator
+    '_llm_intent',
+    '_llm_change_summary',
+    'llm_prefilter',
+    'llm_evaluation_cache',
+    'llm_last_tokens_used',
+    'llm_tokens_used_cumulative',
+})
+
+
+def get_system_managed_non_spec_fields():
+    """Return the set of internal fields not in the public OpenAPI spec."""
+    return SYSTEM_MANAGED_NON_SPEC_FIELDS
+
+
@functools.cache
 def get_openapi_schema_dict():
    """
@@ -48,8 +48,9 @@ To verify this works:
 """

 import json
+import os
 import re
-from urllib.parse import unquote_plus
+from urllib.parse import unquote_plus, urlparse

 import requests
 from apprise import plugins
@@ -59,6 +60,8 @@ from apprise.utils.logic import dict_full_update
 from loguru import logger
 from requests.structures import CaseInsensitiveDict

+from changedetectionio.validate_url import is_private_hostname, is_url_private_or_parser_confused
+
 SUPPORTED_HTTP_METHODS = {"get", "post", "put", "delete", "patch", "head"}


@@ -195,6 +198,17 @@ def apprise_http_custom_handler(

    url = re.sub(rf"^{schema}", "https" if schema.endswith("s") else "http", parsed_url.get("url"))

+    # SSRF protection — block private/loopback addresses unless explicitly allowed.
+    # Uses parser-agnostic check so urlparse/urllib3 differentials (GHSA-rph4-96w6-q594)
+    # can't smuggle an internal target past the gate.
+    if not os.getenv('ALLOW_IANA_RESTRICTED_ADDRESSES', '').lower() in ('true', '1', 'yes'):
+        if is_url_private_or_parser_confused(url):
+            raise ValueError(
+                f"Notification target '{url}' is a private/reserved address "
+                f"or contains a parser-differential payload. "
+                f"Set ALLOW_IANA_RESTRICTED_ADDRESSES=true to allow."
+            )
+
    response = requests.request(
        method=method,
        url=url,
@@ -65,6 +65,9 @@ def notification_format_align_with_apprise(n_format : str):
    :return:
    """

+    if not n_format:
+        return NotifyFormat.TEXT.value
+
    if n_format.startswith('html'):
        # Apprise only knows 'html' not 'htmlcolor' etc, which shouldnt matter here
        n_format = NotifyFormat.HTML.value
@@ -259,9 +262,12 @@ def apply_service_tweaks(url, n_body, n_title, requested_output_format):
    elif (url.startswith('discord://') or url.startswith('https://discordapp.com/api/webhooks')
          or url.startswith('https://discord.com/api'))\
            and 'html' in requested_output_format:
-        # Discord doesn't support HTML, replace <br> with newlines
+        # Discord doesn't render HTML — convert markup to plain text equivalents.
+        # &nbsp; is injected upstream to preserve double-spaces for HTML email clients;
+        # Discord displays it as the literal string "&nbsp;" so strip it here.
        n_body = n_body.strip().replace('<br>', '\n')
        n_body = n_body.replace('</br>', '\n')
+        n_body = n_body.replace('&nbsp;', ' ')
        n_body = newline_re.sub('\n', n_body)

        # Don't replace placeholders or truncate here - let the custom Discord plugin handle it
@@ -358,9 +364,55 @@ def process_notification(n_object: NotificationContextData, datastore):
        # Should always be false for 'text' mode or its too hard to read
        # But otherwise, this could be some setting
        word_diff=False if requested_output_format_original == 'text' else True,
+        # HTML-format notifications must escape diff content (GHSA-q8xq-qg4x-wphg).
+        # FormattableDiff/Extract escape internally so {{ diff(...) }} stays callable —
+        # the post-Jinja escape loop below would otherwise convert them to plain str.
+        escape_output='html' in requested_output_format,
        )
    )

+    # {{ raw_diff }} always holds the actual diff regardless of AI Change Summary
+    n_object['raw_diff'] = n_object.get('diff', '')
+
+    # AI Change Summary: optionally replace {{ diff }} with the AI summary
+    _llm_change_summary = (n_object.get('_llm_change_summary') or '').strip()
+    from changedetectionio.llm.evaluator import get_llm_settings
+    _override_diff = get_llm_settings(datastore).override_diff_with_summary
+    if _llm_change_summary and _override_diff:
+        n_object['diff'] = _llm_change_summary
+
+    # Lazily populate llm_summary / llm_intent if used in notification template
+    scan_text = n_object.get('notification_body', '') + n_object.get('notification_title', '')
+    if 'llm_summary' in scan_text or 'llm_intent' in scan_text or 'raw_diff' in scan_text:
+        n_object['llm_summary'] = _llm_change_summary or (n_object.get('_llm_result') or {}).get('summary', '')
+        n_object['llm_intent'] = n_object.get('_llm_intent', '')
+
+    # Escape diff/snapshot variables before Jinja renders them into an HTML notification.
+    # GHSA-q8xq-qg4x-wphg: inscriptis decodes HTML entities when converting text/html
+    # pages to snapshot text, so a page that visibly displays "&lt;a href...&gt;" yields
+    # literal "<a href...>" in the snapshot — which would otherwise render as live
+    # markup in HTML emails / Telegram (parse_mode=html) / Discord embeds, letting a
+    # watched page inject phishing links into the operator's notification channel.
+    # Also covers #3529 — raw '<' chars from text/plain pages breaking HTML email layout.
+    # The operator's own template HTML (e.g. <a href="{{watch_url}}">) is outside the
+    # variable values so it stays untouched. Diff placemarkers contain no HTML chars,
+    # so they survive escape and are still replaced with <span> tags later.
+    if 'html' in requested_output_format:
+        from markupsafe import escape as html_escape
+        from changedetectionio.notification_service import FormattableDiff, FormattableExtract
+        _page_content_keys = {'raw_diff', 'current_snapshot', 'prev_snapshot', 'triggered_text'}
+        for key in [k for k in notification_parameters if k.startswith('diff') or k in _page_content_keys]:
+            value = notification_parameters.get(key)
+            if not value:
+                continue
+            # FormattableDiff / FormattableExtract are callable str subclasses — {{ diff(lines=5) }}
+            # etc. relies on __call__. Wrapping them with str(html_escape(...)) here would lose
+            # __call__ and break those tokens. They escape internally via escape_output=True
+            # (set by add_rendered_diff_to_notification_vars above) for both __str__ and __call__.
+            if isinstance(value, (FormattableDiff, FormattableExtract)):
+                continue
+            notification_parameters[key] = str(html_escape(str(value)))
+
    with (apprise.LogCapture(level=apprise.logging.DEBUG) as logs):
        for url in n_object['notification_urls']:

@@ -378,13 +430,6 @@ def process_notification(n_object: NotificationContextData, datastore):
            logger.info(f">> Process Notification: AppRise start notifying '{url}'")
            url = jinja_render(template_str=url, **notification_parameters)

-            # If it's a plaintext document, and they want HTML type email/alerts, so it needs to be escaped
-            watch_mime_type = n_object.get('watch_mime_type')
-            if watch_mime_type and 'text/' in watch_mime_type.lower() and not 'html' in watch_mime_type.lower():
-                if 'html' in requested_output_format:
-                    from markupsafe import escape
-                    n_body = str(escape(n_body))
-
            if 'html' in requested_output_format:
                # Since the n_body is always some kind of text from the 'diff' engine, attempt to preserve whitespaces that get sent to the HTML output
                # But only where its more than 1 consecutive whitespace, otherwise "and this" becomes "and&nbsp;this" etc which is too much.
@@ -29,7 +29,7 @@ def _check_cascading_vars(datastore, var_name, watch):
    v = watch.get(var_name)
    if v and not watch.get('notification_muted'):
        if var_name == 'notification_format' and v == USE_SYSTEM_DEFAULT_NOTIFICATION_FORMAT_FOR_WATCH:
-            return datastore.data['settings']['application'].get('notification_format')
+            return datastore.data['settings']['application'].get('notification_format') or default_notification_format

        return v

@@ -54,34 +54,174 @@ def _check_cascading_vars(datastore, var_name, watch):
    return None


+class FormattableTimestamp(str):
+    """
+    A str subclass representing a formatted datetime. As a plain string it renders
+    with the default format, but can also be called with a custom format argument
+    in Jinja2 templates:
+
+        {{ change_datetime }}                        → '2024-01-15 10:30:00 UTC'
+        {{ change_datetime(format='%Y') }}           → '2024'
+        {{ change_datetime(format='%A') }}           → 'Monday'
+        {{ change_datetime(format='%Y-%m-%d') }}     → '2024-01-15'
+
+    Being a str subclass means it is natively JSON serializable.
+    """
+    _DEFAULT_FORMAT = '%Y-%m-%d %H:%M:%S %Z'
+
+    def __new__(cls, timestamp):
+        dt = datetime.datetime.fromtimestamp(int(timestamp), tz=pytz.UTC)
+        local_tz = datetime.datetime.now().astimezone().tzinfo
+        dt_local = dt.astimezone(local_tz)
+        try:
+            formatted = dt_local.strftime(cls._DEFAULT_FORMAT)
+        except Exception:
+            formatted = dt_local.isoformat()
+        instance = super().__new__(cls, formatted)
+        instance._dt = dt_local
+        return instance
+
+    def __call__(self, format=_DEFAULT_FORMAT):
+        try:
+            return self._dt.strftime(format)
+        except Exception:
+            return self._dt.isoformat()
+
+
+class FormattableExtract(str):
+    """
+    A str subclass that holds only the extracted changed fragments from a diff.
+    Used for {{diff_changed_from}} and {{diff_changed_to}} tokens.
+
+        {{ diff_changed_from }}   → old value(s) only, e.g. "$99.99"
+        {{ diff_changed_to }}     → new value(s) only, e.g. "$109.99"
+
+    Multiple changed fragments are joined with newlines.
+    Being a str subclass means it is natively JSON serializable.
+    """
+    def __new__(cls, prev_snapshot, current_snapshot, extract_fn, escape_output=False):
+        if prev_snapshot or current_snapshot:
+            from changedetectionio import diff as diff_module
+            # word_diff=True is required — placemarker extraction regexes only exist in word-diff output
+            raw = diff_module.render_diff(prev_snapshot or '', current_snapshot or '', word_diff=True)
+            extracted = extract_fn(raw)
+        else:
+            extracted = ''
+        if escape_output and extracted:
+            # Placemarkers (@removed_PLACEMARKER_OPEN etc) contain no HTML chars,
+            # so html_escape leaves them intact — they still get swapped to <span>
+            # tags later by apply_service_tweaks. See GHSA-q8xq-qg4x-wphg.
+            from markupsafe import escape as html_escape
+            extracted = str(html_escape(extracted))
+        instance = super().__new__(cls, extracted)
+        return instance
+
+
+class FormattableDiff(str):
+    """
+    A str subclass representing a rendered diff. As a plain string it renders
+    with the default options for that variant, but can be called with custom
+    arguments in Jinja2 templates:
+
+        {{ diff }}                                    → default diff output
+        {{ diff(lines=5) }}                           → truncate to 5 lines
+        {{ diff(added_only=true) }}                   → only show added lines
+        {{ diff(removed_only=true) }}                 → only show removed lines
+        {{ diff(context=3) }}                         → 3 lines of context around changes
+        {{ diff(word_diff=false) }}                   → line-level diff instead of word-level
+        {{ diff(lines=10, added_only=true) }}         → combine args
+        {{ diff_added(lines=5) }}                     → works on any diff_* variant too
+
+    Being a str subclass means it is natively JSON serializable.
+    """
+    def __new__(cls, prev_snapshot, current_snapshot, escape_output=False, **base_kwargs):
+        if prev_snapshot or current_snapshot:
+            from changedetectionio import diff as diff_module
+            rendered = diff_module.render_diff(prev_snapshot, current_snapshot, **base_kwargs)
+        else:
+            rendered = ''
+        if escape_output and rendered:
+            # Placemarkers (@removed_PLACEMARKER_OPEN etc) contain no HTML chars,
+            # so html_escape leaves them intact — they still get swapped to <span>
+            # tags later by apply_service_tweaks. See GHSA-q8xq-qg4x-wphg.
+            from markupsafe import escape as html_escape
+            rendered = str(html_escape(rendered))
+        instance = super().__new__(cls, rendered)
+        instance._prev = prev_snapshot
+        instance._current = current_snapshot
+        instance._base_kwargs = base_kwargs
+        instance._escape_output = escape_output
+        return instance
+
+    def __call__(self, lines=None, added_only=False, removed_only=False, context=0,
+                 word_diff=None, case_insensitive=False, ignore_junk=False):
+        from changedetectionio import diff as diff_module
+        kwargs = dict(self._base_kwargs)
+
+        if added_only:
+            kwargs['include_removed'] = False
+        if removed_only:
+            kwargs['include_added'] = False
+        if context:
+            kwargs['context_lines'] = int(context)
+        if word_diff is not None:
+            kwargs['word_diff'] = bool(word_diff)
+        if case_insensitive:
+            kwargs['case_insensitive'] = True
+        if ignore_junk:
+            kwargs['ignore_junk'] = True
+
+        result = diff_module.render_diff(self._prev or '', self._current or '', **kwargs)
+
+        if lines is not None:
+            result = '\n'.join(result.splitlines()[:int(lines)])
+
+        if self._escape_output and result:
+            from markupsafe import escape as html_escape
+            result = str(html_escape(result))
+
+        return result
+
+
+
 # What is passed around as notification context, also used as the complete list of valid {{ tokens }}
 class NotificationContextData(dict):
    def __init__(self, initial_data=None, **kwargs):
+        # ValidateJinja2Template() validates against the keynames of this dict to check for valid tokens in the body (user submission)
        super().__init__({
            'base_url': None,
+            'change_datetime': FormattableTimestamp(time.time()),
            'current_snapshot': None,
-            'diff': None,
-            'diff_clean': None,
-            'diff_added': None,
-            'diff_added_clean': None,
-            'diff_full': None,
-            'diff_full_clean': None,
-            'diff_patch': None,
-            'diff_removed': None,
-            'diff_removed_clean': None,
+            'diff': FormattableDiff('', ''),
+            'diff_clean': FormattableDiff('', '', include_change_type_prefix=False),
+            'diff_added': FormattableDiff('', '', include_removed=False),
+            'diff_added_clean': FormattableDiff('', '', include_removed=False, include_change_type_prefix=False),
+            'diff_full': FormattableDiff('', '', include_equal=True),
+            'diff_full_clean': FormattableDiff('', '', include_equal=True, include_change_type_prefix=False),
+            'diff_patch': FormattableDiff('', '', patch_format=True),
+            'diff_removed': FormattableDiff('', '', include_added=False),
+            'diff_removed_clean': FormattableDiff('', '', include_added=False, include_change_type_prefix=False),
+            'diff_changed_from': FormattableExtract('', '', extract_fn=lambda x: x),
+            'diff_changed_to': FormattableExtract('', '', extract_fn=lambda x: x),
            'diff_url': None,
+            # Always the raw +/- diff regardless of LLM summary override (populated in handler.py from {{diff}})
+            'raw_diff': FormattableDiff('', ''),
            'markup_text_links_to_html_links': False, # If automatic conversion of plaintext to HTML should happen
            'notification_timestamp': time.time(),
+            'prev_snapshot': None,
            'preview_url': None,
            'screenshot': None,
-            'triggered_text': None,
            'timestamp_from': None,
            'timestamp_to': None,
+            'triggered_text': None,
+            'llm_summary': None,     # AI plain-English summary of what changed (requires AI intent to be configured)
+            'llm_intent': None,      # The intent that was evaluated (watch-level or inherited from tag)
            'uuid': 'XXXXXXXX-XXXX-XXXX-XXXX-XXXXXXXXXXXX',  # Converted to 'watch_uuid' in create_notification_parameters
            'watch_mime_type': None,
            'watch_tag': None,
            'watch_title': None,
            'watch_url': 'https://WATCH-PLACE-HOLDER/',
+            'watch_uuid': 'XXXXXXXX-XXXX-XXXX-XXXX-XXXXXXXXXXXX',  # Converted to 'watch_uuid' in create_notification_parameters
        })

        # Apply any initial data passed in
@@ -103,7 +243,7 @@ class NotificationContextData(dict):
        So we can test the output in the notification body
        """
        for key in self.keys():
-            if key in ['uuid', 'time', 'watch_uuid']:
+            if key in ['uuid', 'time', 'watch_uuid', 'change_datetime'] or key.startswith('diff'):
                continue
            rand_str = 'RANDOM-PLACEHOLDER-'+''.join(random.choices(string.ascii_letters + string.digits, k=12))
            self[key] = rand_str
@@ -115,25 +255,7 @@ class NotificationContextData(dict):

        super().__setitem__(key, value)

-def timestamp_to_localtime(timestamp):
-    # Format the date using locale-aware formatting with timezone
-    dt = datetime.datetime.fromtimestamp(int(timestamp))
-    dt = dt.replace(tzinfo=pytz.UTC)
-
-    # Get local timezone-aware datetime
-    local_tz = datetime.datetime.now().astimezone().tzinfo
-    local_dt = dt.astimezone(local_tz)
-
-    # Format date with timezone - using strftime for locale awareness
-    try:
-        formatted_date = local_dt.strftime('%Y-%m-%d %H:%M:%S %Z')
-    except:
-        # Fallback if locale issues
-        formatted_date = local_dt.isoformat()
-
-    return formatted_date
-
-def add_rendered_diff_to_notification_vars(notification_scan_text:str, prev_snapshot:str, current_snapshot:str, word_diff:bool):
+def add_rendered_diff_to_notification_vars(notification_scan_text:str, prev_snapshot:str, current_snapshot:str, word_diff:bool, escape_output:bool=False):
    """
    Efficiently renders only the diff placeholders that are actually used in the notification text.

@@ -146,17 +268,19 @@ def add_rendered_diff_to_notification_vars(notification_scan_text:str, prev_snap
        prev_snapshot: Previous version of content for diff comparison
        current_snapshot: Current version of content for diff comparison
        word_diff: Whether to use word-level (True) or line-level (False) diffing
+        escape_output: If True, the rendered diff output is HTML-escaped. Used for HTML-format
+            notifications so attacker-controlled page content can't inject live markup.
+            Both the cached str representation and the result of {{ diff(...) }} calls are escaped.

    Returns:
        dict: Only the diff placeholders that were found in notification_scan_text, with rendered content
    """
-    from changedetectionio import diff
    import re
-    from functools import lru_cache

    now = time.time()

-    # Define specifications for each diff variant
+    # Define base kwargs for each diff variant — these become the stored defaults
+    # on the FormattableDiff object, so {{ diff(lines=5) }} overrides on top of them
    diff_specs = {
        'diff': {'word_diff': word_diff},
        'diff_clean': {'word_diff': word_diff, 'include_change_type_prefix': False},
@@ -169,23 +293,27 @@ def add_rendered_diff_to_notification_vars(notification_scan_text:str, prev_snap
        'diff_removed_clean': {'word_diff': word_diff, 'include_added': False, 'include_change_type_prefix': False},
    }

-    # Memoize render_diff to avoid duplicate renders with same kwargs
-    @lru_cache(maxsize=4)
-    def cached_render(kwargs_tuple):
-        return diff.render_diff(prev_snapshot, current_snapshot, **dict(kwargs_tuple))
+    from changedetectionio.diff import extract_changed_from, extract_changed_to
+    extract_specs = {
+        'diff_changed_from': extract_changed_from,
+        'diff_changed_to':   extract_changed_to,
+    }

    ret = {}
    rendered_count = 0
-    # Only check and render diff keys that exist in NotificationContextData
+    # Only create FormattableDiff/FormattableExtract objects for diff keys actually used in the notification text
    for key in NotificationContextData().keys():
-        if key.startswith('diff') and key in diff_specs:
-            # Check if this placeholder is actually used in the notification text
-            pattern = rf"(?<![A-Za-z0-9_]){re.escape(key)}(?![A-Za-z0-9_])"
-            if re.search(pattern, notification_scan_text, re.IGNORECASE):
-                kwargs = diff_specs[key]
-                # Convert dict to sorted tuple for cache key (handles duplicate kwarg combinations)
-                ret[key] = cached_render(tuple(sorted(kwargs.items())))
-                rendered_count += 1
+        if not key.startswith('diff'):
+            continue
+        pattern = rf"(?<![A-Za-z0-9_]){re.escape(key)}(?![A-Za-z0-9_])"
+        if not re.search(pattern, notification_scan_text, re.IGNORECASE):
+            continue
+        if key in diff_specs:
+            ret[key] = FormattableDiff(prev_snapshot, current_snapshot, escape_output=escape_output, **diff_specs[key])
+            rendered_count += 1
+        elif key in extract_specs:
+            ret[key] = FormattableExtract(prev_snapshot, current_snapshot, extract_fn=extract_specs[key], escape_output=escape_output)
+            rendered_count += 1

    if rendered_count:
        logger.trace(f"Rendered {rendered_count} diff placeholder(s) {sorted(ret.keys())} in {time.time() - now:.3f}s")
@@ -198,7 +326,7 @@ def set_basic_notification_vars(current_snapshot, prev_snapshot, watch, triggere
        'current_snapshot': current_snapshot,
        'prev_snapshot': prev_snapshot,
        'screenshot': watch.get_screenshot() if watch and watch.get('notification_screenshot') else None,
-        'change_datetime': timestamp_to_localtime(timestamp_changed) if timestamp_changed else None,
+        'change_datetime': FormattableTimestamp(timestamp_changed) if timestamp_changed else None,
        'triggered_text': triggered_text,
        'uuid': watch.get('uuid') if watch else None,
        'watch_url': watch.get('url') if watch else None,
@@ -307,6 +435,11 @@ class NotificationService:
        n_object['notification_body'] = _check_cascading_vars(self.datastore,'notification_body', watch)
        n_object['notification_format'] = _check_cascading_vars(self.datastore,'notification_format', watch)

+        # Attach LLM results so notification tokens render correctly
+        n_object['_llm_result'] = watch.get('_llm_result')
+        n_object['_llm_intent'] = watch.get('_llm_intent', '')
+        n_object['_llm_change_summary'] = watch.get('_llm_change_summary', '')
+
        # (Individual watch) Only prepare to notify if the rules above matched
        queued = False
        if n_object and n_object.get('notification_urls'):
@@ -346,7 +479,7 @@ Thanks - Your omniscient changedetection.io installation.
            'notification_body': body,
            'notification_format': _check_cascading_vars(self.datastore, 'notification_format', watch),
        })
-        n_object['markup_text_links_to_html_links'] = n_object.get('notification_format').startswith('html')
+        n_object['markup_text_links_to_html_links'] = (n_object.get('notification_format') or '').startswith('html')

        if len(watch['notification_urls']):
            n_object['notification_urls'] = watch['notification_urls']
@@ -393,9 +526,9 @@ Thanks - Your omniscient changedetection.io installation.
        n_object = NotificationContextData({
            'notification_title': f"Changedetection.io - Alert - Browser step at position {step} could not be run",
            'notification_body': body,
-            'notification_format': self._check_cascading_vars('notification_format', watch),
+            'notification_format': _check_cascading_vars(self.datastore, 'notification_format', watch),
        })
-        n_object['markup_text_links_to_html_links'] = n_object.get('notification_format').startswith('html')
+        n_object['markup_text_links_to_html_links'] = (n_object.get('notification_format') or '').startswith('html')

        if len(watch['notification_urls']):
            n_object['notification_urls'] = watch['notification_urls']
@@ -61,7 +61,7 @@ class ChangeDetectionSpec:
        pass

    @hookspec
-    def get_itemprop_availability_override(self, content, fetcher_name, fetcher_instance, url):
+    def get_itemprop_availability_override(self, content, fetcher_name, fetcher_instance, url, llm_intent=None):
        """Provide custom implementation of get_itemprop_availability for a specific fetcher.

        This hook allows plugins to provide their own product availability detection
@@ -73,6 +73,7 @@ class ChangeDetectionSpec:
            fetcher_name: The name of the fetcher being used (e.g., 'html_js_zyte')
            fetcher_instance: The fetcher instance that generated the content
            url: The URL being watched/checked
+            llm_intent: Optional user-supplied intent string (e.g. "alert when price drops below $300")

        Returns:
            dict or None: Dictionary with availability data:
@@ -129,6 +130,109 @@ class ChangeDetectionSpec:
        """
        pass

+    @hookspec
+    def update_handler_alter(update_handler, watch, datastore):
+        """Modify or wrap the update_handler before it processes a watch.
+
+        This hook is called after the update_handler (perform_site_check instance) is created
+        but before it calls call_browser() and run_changedetection(). Plugins can use this to:
+        - Wrap the handler to add logging/metrics
+        - Modify handler configuration
+        - Add custom preprocessing logic
+
+        Args:
+            update_handler: The perform_site_check instance that will process the watch
+            watch: The watch dict being processed
+            datastore: The application datastore
+
+        Returns:
+            object or None: Return a modified/wrapped handler, or None to keep the original.
+                           If multiple plugins return handlers, they are chained in registration order.
+        """
+        pass
+
+    @hookspec
+    def update_finalize(update_handler, watch, datastore, processing_exception):
+        """Called after watch processing completes (success or failure).
+
+        This hook is called in the finally block after all processing is complete,
+        allowing plugins to perform cleanup, update metrics, or log final status.
+
+        The plugin can access update_handler.last_logging_insert_id if it was stored
+        during update_handler_alter, and use processing_exception to determine if
+        the processing succeeded or failed.
+
+        Args:
+            update_handler: The perform_site_check instance (may be None if creation failed)
+            watch: The watch dict that was processed (may be None if not loaded)
+            datastore: The application datastore
+            processing_exception: The exception from the main processing block, or None if successful.
+                                 This does NOT include cleanup exceptions - only exceptions from
+                                 the actual watch processing (fetch, diff, etc).
+
+        Returns:
+            None: This hook doesn't return a value
+        """
+        pass
+
+    @hookspec
+    def get_html_head_extras():
+        """Return HTML to inject into the <head> of every page via base.html.
+
+        Plugins can use this to add <script>, <style>, or <link> tags that should
+        be present on all pages.  Return a raw HTML string or None.
+
+        IMPORTANT: Always use Flask's url_for() for any src/href URLs so that
+        sub-path deployments (nginx reverse proxy with USE_X_SETTINGS / X-Forwarded-Prefix)
+        work correctly.  This hook is called inside a request context so url_for() is
+        always available.
+
+        For small amounts of CSS/JS, return them inline — no file-serving needed::
+
+            from changedetectionio.pluggy_interface import hookimpl
+
+            @hookimpl
+            def get_html_head_extras(self):
+                return (
+                    '<style>.my-module-banner { color: red; }</style>\\n'
+                    '<script>console.log("my_module_content loaded");</script>'
+                )
+
+        For larger assets, register your own lightweight Flask routes in the plugin
+        module and point to them with url_for() so the sub-path prefix is handled
+        automatically::
+
+            from flask import url_for, Response
+            from changedetectionio.pluggy_interface import hookimpl
+            from changedetectionio.flask_app import app as _app
+
+            MY_CSS = ".my-module-example { color: red; }"
+            MY_JS  = "console.log('my_module_content loaded');"
+
+            @_app.route('/my_module_content/css')
+            def my_module_content_css():
+                return Response(MY_CSS, mimetype='text/css',
+                                headers={'Cache-Control': 'max-age=3600'})
+
+            @_app.route('/my_module_content/js')
+            def my_module_content_js():
+                return Response(MY_JS, mimetype='application/javascript',
+                                headers={'Cache-Control': 'max-age=3600'})
+
+            @hookimpl
+            def get_html_head_extras(self):
+                css = url_for('my_module_content_css')
+                js  = url_for('my_module_content_js')
+                return (
+                    f'<link rel="stylesheet" href="{css}">\\n'
+                    f'<script src="{js}" defer></script>'
+                )
+
+        Returns:
+            str or None: Raw HTML string to inject inside <head>, or None
+        """
+        pass
+

 # Set up Plugin Manager
 plugin_manager = pluggy.PluginManager(PLUGIN_NAMESPACE)
@@ -138,24 +242,27 @@ plugin_manager.add_hookspecs(ChangeDetectionSpec)

 # Load plugins from subdirectories
 def load_plugins_from_directories():
-    # Dictionary of directories to scan for plugins
-    plugin_dirs = {
-        'conditions': os.path.join(os.path.dirname(__file__), 'conditions', 'plugins'),
-        # Add more plugin directories here as needed
-    }
-    
-    # Note: Removed the direct import of example_word_count_plugin as it's now in the conditions/plugins directory
-    
-    for dir_name, dir_path in plugin_dirs.items():
+    # List of (python_package_prefix, filesystem_path) pairs to scan for plugins.
+    # NOTE: processors/restock_diff/plugins is intentionally excluded here — those
+    # plugins are registered via register_builtin_restock_plugins() to avoid the
+    # circular import: restock_diff/__init__.py → model.Watch → content_fetchers → pluggy_interface.
+    plugin_dirs = [
+        (
+            'changedetectionio.conditions.plugins',
+            os.path.join(os.path.dirname(__file__), 'conditions', 'plugins'),
+        ),
+    ]
+
+    for module_prefix, dir_path in plugin_dirs:
        if not os.path.exists(dir_path):
            continue
-            
+
        # Get all Python files (excluding __init__.py)
        for filename in os.listdir(dir_path):
            if filename.endswith(".py") and filename != "__init__.py":
                module_name = filename[:-3]  # Remove .py extension
-                module_path = f"changedetectionio.{dir_name}.plugins.{module_name}"
-                
+                module_path = f"{module_prefix}.{module_name}"
+
                try:
                    module = importlib.import_module(module_path)
                    # Register the plugin with pluggy
@@ -207,6 +314,24 @@ def register_builtin_fetchers():
    if hasattr(webdriver_selenium, 'webdriver_selenium_plugin'):
        plugin_manager.register(webdriver_selenium.webdriver_selenium_plugin, 'builtin_webdriver_selenium')

+
+def register_builtin_restock_plugins():
+    """Register built-in restock processor plugins after all imports are complete.
+
+    Called from content_fetchers/__init__.py alongside register_builtin_fetchers()
+    to avoid the circular import that occurs when loading via load_plugins_from_directories()
+    (restock_diff/__init__.py → model.Watch → content_fetchers → pluggy_interface).
+    """
+    import importlib
+    module_path = 'changedetectionio.processors.restock_diff.plugins.llm_restock'
+    try:
+        module = importlib.import_module(module_path)
+        if not plugin_manager.is_registered(module):
+            plugin_manager.register(module, 'llm_restock')
+            logger.debug("Registered built-in restock plugin: llm_restock")
+    except Exception as e:
+        logger.error(f"Failed to register llm_restock plugin: {e}")
+
 # Helper function to collect UI stats extras from all plugins
 def collect_ui_edit_stats_extras(watch):
    """Collect and combine HTML content from all plugins that implement ui_edit_stats_extras"""
@@ -243,7 +368,7 @@ def collect_fetcher_status_icons(fetcher_name):

    return None

-def get_itemprop_availability_from_plugin(content, fetcher_name, fetcher_instance, url):
+def get_itemprop_availability_from_plugin(content, fetcher_name, fetcher_instance, url, llm_intent=None):
    """Get itemprop availability data from plugins as a fallback.

    This is called when the built-in get_itemprop_availability doesn't find good data.
@@ -253,6 +378,7 @@ def get_itemprop_availability_from_plugin(content, fetcher_name, fetcher_instanc
        fetcher_name: The name of the fetcher being used (e.g., 'html_js_zyte')
        fetcher_instance: The fetcher instance that generated the content
        url: The URL being watched (watch.link - includes Jinja2 evaluation)
+        llm_intent: Optional user-supplied intent string passed through to plugins

    Returns:
        dict or None: Availability data dictionary from first matching plugin, or None
@@ -262,7 +388,8 @@ def get_itemprop_availability_from_plugin(content, fetcher_name, fetcher_instanc
        content=content,
        fetcher_name=fetcher_name,
        fetcher_instance=fetcher_instance,
-        url=url
+        url=url,
+        llm_intent=llm_intent,
    )

    # Return first non-None result with actual data
@@ -499,4 +626,82 @@ def get_plugin_template_paths():
                template_paths.append(templates_dir)
                logger.debug(f"Added plugin template path: {templates_dir}")

-    return template_paths
+    return template_paths
+
+
+def apply_update_handler_alter(update_handler, watch, datastore):
+    """Apply update_handler_alter hooks from all plugins.
+
+    Allows plugins to wrap or modify the update_handler before it processes a watch.
+    Multiple plugins can chain modifications - each plugin receives the result from
+    the previous plugin.
+
+    Args:
+        update_handler: The perform_site_check instance to potentially modify
+        watch: The watch dict being processed
+        datastore: The application datastore
+
+    Returns:
+        object: The (potentially modified/wrapped) update_handler
+    """
+    # Get all plugins that implement the update_handler_alter hook
+    results = plugin_manager.hook.update_handler_alter(
+        update_handler=update_handler,
+        watch=watch,
+        datastore=datastore
+    )
+
+    # Chain results - each plugin gets the result from the previous one
+    current_handler = update_handler
+    if results:
+        for result in results:
+            if result is not None:
+                logger.debug(f"Plugin modified update_handler for watch {watch.get('uuid')}")
+                current_handler = result
+
+    return current_handler
+
+
+def apply_update_finalize(update_handler, watch, datastore, processing_exception):
+    """Apply update_finalize hooks from all plugins.
+
+    Called in the finally block after watch processing completes, allowing plugins
+    to perform cleanup, update metrics, or log final status.
+
+    Args:
+        update_handler: The perform_site_check instance (may be None)
+        watch: The watch dict that was processed (may be None)
+        datastore: The application datastore
+        processing_exception: The exception from processing, or None if successful
+
+    Returns:
+        None
+    """
+    try:
+        # Call all plugins that implement the update_finalize hook
+        plugin_manager.hook.update_finalize(
+            update_handler=update_handler,
+            watch=watch,
+            datastore=datastore,
+            processing_exception=processing_exception
+        )
+    except Exception as e:
+        # Don't let plugin errors crash the worker
+        logger.error(f"Error in update_finalize hook: {e}")
+        logger.exception(f"update_finalize hook exception details:")
+
+
+def collect_html_head_extras():
+    """Collect and combine HTML head extras from all plugins.
+
+    Called from a Flask template global so it always runs inside a request context.
+    This means url_for() works correctly in plugin implementations, including when the
+    app is deployed under a sub-path via USE_X_SETTINGS / X-Forwarded-Prefix (ProxyFix
+    sets SCRIPT_NAME so url_for() automatically prepends the prefix).
+
+    Returns:
+        str: Combined HTML string to inject inside <head>, or empty string
+    """
+    results = plugin_manager.hook.get_html_head_extras()
+    parts = [r for r in results if r]
+    return "\n".join(parts) if parts else ""
@@ -9,6 +9,15 @@ Some suggestions for the future

 - `graphical` 

+## API schema extension (`api.yaml`)
+
+A processor can extend the Watch/Tag API schema by placing an `api.yaml` alongside its `__init__.py`.
+Define a `components.schemas.processor_config_<name>` entry and it will be merged into `WatchBase` at startup,
+making `processor_config_<name>` a valid field on all watch create/update API calls.
+The fully merged spec is served live at `/api/v1/full-spec`.
+
+See `restock_diff/api.yaml` for a working example.
+
 ## Todo

 - Make each processor return a extra list of sub-processed (so you could configure a single processor in different ways)
@@ -341,6 +341,18 @@ def get_processor_descriptions():
    return descriptions


+def wcag_text_color(hex_bg: str) -> str:
+    """Return #000000 or #ffffff for maximum WCAG contrast against hex_bg."""
+    hex_bg = hex_bg.lstrip('#')
+    if len(hex_bg) != 6:
+        return '#000000'
+    r, g, b = (int(hex_bg[i:i+2], 16) / 255 for i in (0, 2, 4))
+    def lin(c):
+        return c / 12.92 if c <= 0.04045 else ((c + 0.055) / 1.055) ** 2.4
+    L = 0.2126 * lin(r) + 0.7152 * lin(g) + 0.0722 * lin(b)
+    return '#000000' if L > 0.179 else '#ffffff'
+
+
 def generate_processor_badge_colors(processor_name):
    """
    Generate consistent colors for a processor badge based on its name.
@@ -1,12 +1,15 @@
+import asyncio
 import re
 import hashlib

 from changedetectionio.browser_steps.browser_steps import browser_steps_get_valid_steps
 from changedetectionio.content_fetchers.base import Fetcher
 from changedetectionio.strtobool import strtobool
+from changedetectionio.validate_url import is_private_hostname, is_url_private_or_parser_confused
 from copy import deepcopy
 from abc import abstractmethod
 import os
+from urllib.parse import urlparse
 from loguru import logger

 SCREENSHOT_FORMAT_JPEG = 'JPEG'
@@ -94,6 +97,22 @@ class difference_detection_processor():
            logger.warning(f"Failed to read checksum file for {self.watch_uuid}: {e}")
            self.last_raw_content_checksum = None

+    async def validate_iana_url(self):
+        """Pre-flight SSRF check — runs DNS lookup in executor to avoid blocking the event loop.
+        Covers all fetchers (requests, playwright, puppeteer, plugins) since every fetch goes
+        through call_browser().
+        """
+        if strtobool(os.getenv('ALLOW_IANA_RESTRICTED_ADDRESSES', 'false')):
+            return
+        loop = asyncio.get_running_loop()
+        # Use the parser-agnostic check so urlparse/urllib3 differentials (GHSA-rph4-96w6-q594)
+        # can't slip a private/internal hostname past this pre-flight gate.
+        if await loop.run_in_executor(None, is_url_private_or_parser_confused, self.watch.link):
+            raise Exception(
+                f"Fetch blocked: '{self.watch.link}' resolves to a private/reserved IP address "
+                f"or contains a parser-differential payload. "
+                f"Set ALLOW_IANA_RESTRICTED_ADDRESSES=true to allow."
+            )

    async def call_browser(self, preferred_proxy_id=None):

@@ -108,6 +127,8 @@ class difference_detection_processor():
                    "file:// type access is denied for security reasons."
                )

+        await self.validate_iana_url()
+
        # Requests, playwright, other browser via wss:// etc, fetch_extra_something
        prefer_fetch_backend = self.watch.get('fetch_backend', 'system')

@@ -238,6 +259,16 @@ class difference_detection_processor():
        # @todo .quit here could go on close object, so we can run JS if change-detected
        await self.fetcher.quit(watch=self.watch)

+        # Sanitize lone surrogates - these can appear when servers return malformed/mixed-encoding
+        # content that gets decoded into surrogate characters (e.g. \udcad). Without this,
+        # encode('utf-8') raises UnicodeEncodeError downstream in checksums, diffs, file writes, etc.
+        # Covers all fetchers (requests, playwright, puppeteer, selenium) in one place.
+        # Also note: By this point we SHOULD know the original encoding so it can safely convert to utf-8 for the rest of the app.
+        # See: https://github.com/dgtlmoon/changedetection.io/issues/3952
+
+        if self.fetcher.content and isinstance(self.fetcher.content, str):
+            self.fetcher.content = self.fetcher.content.encode('utf-8', errors='replace').decode('utf-8')
+
        # After init, call run_changedetection() which will do the actual change-detection

    def get_extra_watch_config(self, filename):
@@ -42,10 +42,7 @@ def render_form(watch, datastore, request, url_for, render_template, flash, redi
    # Get error information for the template
    screenshot_url = watch.get_screenshot()

-    system_uses_webdriver = datastore.data['settings']['application']['fetch_backend'] == 'html_webdriver'
-    is_html_webdriver = False
-    if (watch.get('fetch_backend') == 'system' and system_uses_webdriver) or watch.get('fetch_backend') == 'html_webdriver' or watch.get('fetch_backend', '').startswith('extra_browser_'):
-        is_html_webdriver = True
+    is_html_webdriver = watch.fetcher_supports_screenshots

    password_enabled_and_share_is_off = False
    if datastore.data['settings']['application'].get('password') or os.getenv("SALTED_PASS", False):
@@ -64,7 +61,7 @@ def render_form(watch, datastore, request, url_for, render_template, flash, redi
        screenshot=screenshot_url,
        is_html_webdriver=is_html_webdriver,
        password_enabled_and_share_is_off=password_enabled_and_share_is_off,
-        extra_title=f" - {watch.label} - Extract Data",
+        extra_title=f" - {watch.label} - {gettext('Extract Data')}",
        extra_stylesheets=[url_for('static_content', group='styles', filename='diff.css')],
        pure_menu_fixed=False
    )
@@ -100,7 +100,13 @@ class guess_stream_type():
        if any(s in http_content_header for s in RSS_XML_CONTENT_TYPES):
            self.is_rss = True
        elif any(s in http_content_header for s in JSON_CONTENT_TYPES):
-            self.is_json = True
+            # JSONP detection: server claims application/json but content is actually JSONP (e.g. cb({...}))
+            # A JSONP response starts with an identifier followed by '(' - not valid JSON
+            if re.match(r'^\w[\w.]*\s*\(', test_content):
+                logger.warning(f"Content-Type header claims JSON but content looks like JSONP (starts with identifier+parenthesis) - treating as plaintext")
+                self.is_plaintext = True
+            else:
+                self.is_json = True
        elif 'pdf' in magic_content_header:
            self.is_pdf = True
        # magic will call a rss document 'xml'
@@ -1,6 +1,7 @@

 from babel.numbers import parse_decimal
 from changedetectionio.model.Watch import model as BaseWatch
+from decimal import Decimal, InvalidOperation
 from typing import Union
 import re

@@ -10,6 +11,8 @@ supports_browser_steps = True
 supports_text_filters_and_triggers = True
 supports_text_filters_and_triggers_elements = True
 supports_request_type = True
+_price_re = re.compile(r"Price:\s*(\d+(?:\.\d+)?)", re.IGNORECASE)
+

 class Restock(dict):

@@ -31,6 +34,7 @@ class Restock(dict):

        if standardized_value:
            # Convert to float
+            # @todo locale needs to be the locale of the webpage
            return float(parse_decimal(standardized_value, locale='en'))

        return None
@@ -62,15 +66,22 @@ class Restock(dict):

        super().__setitem__(key, value)

+def get_price_from_history_str(history_str):
+    m = _price_re.search(history_str)
+    if not m:
+        return None
+
+    try:
+        return str(Decimal(m.group(1)))
+    except InvalidOperation:
+        return None
+
+
 class Watch(BaseWatch):
    def __init__(self, *arg, **kw):
        super().__init__(*arg, **kw)
        self['restock'] = Restock(kw['default']['restock']) if kw.get('default') and kw['default'].get('restock') else Restock()

-        self['restock_settings'] = kw['default']['restock_settings'] if kw.get('default',{}).get('restock_settings') else {
-            'follow_price_changes': True,
-            'in_stock_processing' : 'in_stock_only'
-        } #@todo update

    def clear_watch(self):
        super().clear_watch()
@@ -79,13 +90,27 @@ class Watch(BaseWatch):
    def extra_notification_token_values(self):
        values = super().extra_notification_token_values()
        values['restock'] = self.get('restock', {})
+
+        values['restock']['previous_price'] = None
+        if self.history_n >= 2:
+            history = self.history
+            if history and len(history) >=2:
+                """Unfortunately for now timestamp is stored as string key"""
+                sorted_keys = sorted(list(history), key=lambda x: int(x))
+                sorted_keys.reverse()
+
+                price_str = self.get_history_snapshot(timestamp=sorted_keys[-1])
+                if price_str:
+                    values['restock']['previous_price'] = get_price_from_history_str(price_str)
        return values

    def extra_notification_token_placeholder_info(self):
        values = super().extra_notification_token_placeholder_info()

        values.append(('restock.price', "Price detected"))
+        values.append(('restock.in_stock', "In stock status"))
        values.append(('restock.original_price', "Original price at first check"))
+        values.append(('restock.previous_price', "Previous price in history"))

        return values

@@ -0,0 +1,149 @@
+components:
+  schemas:
+    processor_config_restock_diff:
+      type: object
+      description: Configuration for the restock_diff processor (restock and price tracking)
+      properties:
+        in_stock_processing:
+          type: string
+          enum: [in_stock_only, all_changes, 'off']
+          default: in_stock_only
+          description: |
+            When to trigger on stock changes:
+            - `in_stock_only`: Only trigger on Out Of Stock -> In Stock transitions
+            - `all_changes`: Trigger on any availability change
+            - `off`: Disable stock/availability tracking
+        follow_price_changes:
+          type: boolean
+          default: true
+          description: Monitor and track price changes
+        price_change_min:
+          type: [number, 'null']
+          description: Trigger a notification when the price drops below this value
+        price_change_max:
+          type: [number, 'null']
+          description: Trigger a notification when the price rises above this value
+        price_change_threshold_percent:
+          type: [number, 'null']
+          minimum: 0
+          maximum: 100
+          description: Minimum price change percentage since the original price to trigger a notification
+
+paths:
+  /watch:
+    post:
+      x-code-samples:
+        - lang: 'curl'
+          label: 'Restock & price tracking'
+          source: |
+            curl -X POST "http://localhost:5000/api/v1/watch" \
+              -H "x-api-key: YOUR_API_KEY" \
+              -H "Content-Type: application/json" \
+              -d '{
+                "url": "https://example.com/product",
+                "processor": "restock_diff",
+                "processor_config_restock_diff": {
+                  "in_stock_processing": "in_stock_only",
+                  "follow_price_changes": true,
+                  "price_change_threshold_percent": 5
+                }
+              }'
+        - lang: 'Python'
+          label: 'Restock & price tracking'
+          source: |
+            import requests
+
+            headers = {
+                'x-api-key': 'YOUR_API_KEY',
+                'Content-Type': 'application/json'
+            }
+            data = {
+                'url': 'https://example.com/product',
+                'processor': 'restock_diff',
+                'processor_config_restock_diff': {
+                    'in_stock_processing': 'in_stock_only',
+                    'follow_price_changes': True,
+                    'price_change_threshold_percent': 5,
+                }
+            }
+            response = requests.post('http://localhost:5000/api/v1/watch',
+                                     headers=headers, json=data)
+            print(response.json())
+
+  /watch/{uuid}:
+    put:
+      x-code-samples:
+        - lang: 'curl'
+          label: 'Update restock config'
+          source: |
+            curl -X PUT "http://localhost:5000/api/v1/watch/YOUR-UUID" \
+              -H "x-api-key: YOUR_API_KEY" \
+              -H "Content-Type: application/json" \
+              -d '{
+                "processor_config_restock_diff": {
+                  "in_stock_processing": "all_changes",
+                  "follow_price_changes": true,
+                  "price_change_min": 10.00,
+                  "price_change_max": 500.00
+                }
+              }'
+        - lang: 'Python'
+          label: 'Update restock config'
+          source: |
+            import requests
+
+            headers = {
+                'x-api-key': 'YOUR_API_KEY',
+                'Content-Type': 'application/json'
+            }
+            uuid = 'YOUR-UUID'
+            data = {
+                'processor_config_restock_diff': {
+                    'in_stock_processing': 'all_changes',
+                    'follow_price_changes': True,
+                    'price_change_min': 10.00,
+                    'price_change_max': 500.00,
+                }
+            }
+            response = requests.put(f'http://localhost:5000/api/v1/watch/{uuid}',
+                                    headers=headers, json=data)
+            print(response.text)
+
+  /tag/{uuid}:
+    put:
+      x-code-samples:
+        - lang: 'curl'
+          label: 'Set restock config on group/tag'
+          source: |
+            curl -X PUT "http://localhost:5000/api/v1/tag/YOUR-TAG-UUID" \
+              -H "x-api-key: YOUR_API_KEY" \
+              -H "Content-Type: application/json" \
+              -d '{
+                "overrides_watch": true,
+                "processor_config_restock_diff": {
+                  "in_stock_processing": "in_stock_only",
+                  "follow_price_changes": true,
+                  "price_change_threshold_percent": 10
+                }
+              }'
+        - lang: 'Python'
+          label: 'Set restock config on group/tag'
+          source: |
+            import requests
+
+            headers = {
+                'x-api-key': 'YOUR_API_KEY',
+                'Content-Type': 'application/json'
+            }
+            tag_uuid = 'YOUR-TAG-UUID'
+            data = {
+                'overrides_watch': True,
+                'processor_config_restock_diff': {
+                    'in_stock_processing': 'in_stock_only',
+                    'follow_price_changes': True,
+                    'price_change_threshold_percent': 10,
+                }
+            }
+            response = requests.put(f'http://localhost:5000/api/v1/tag/{tag_uuid}',
+                                    headers=headers, json=data)
+            print(response.text)
@@ -31,7 +31,7 @@ class RestockSettingsForm(Form):
    follow_price_changes = BooleanField(_l('Follow price changes'), default=True)

 class processor_settings_form(processor_text_json_diff_form):
-    restock_settings = FormField(RestockSettingsForm)
+    processor_config_restock_diff = FormField(RestockSettingsForm)

    def extra_tab_content(self):
        return _l('Restock & Price Detection')
@@ -48,34 +48,34 @@ class processor_settings_form(processor_text_json_diff_form):

        output += """
        {% from '_helpers.html' import render_field, render_checkbox_field, render_button %}
-        <script>        
+        <script>
            $(document).ready(function () {
-                toggleOpacity('#restock_settings-follow_price_changes', '.price-change-minmax', true);
+                toggleOpacity('#processor_config_restock_diff-follow_price_changes', '.price-change-minmax', true);
            });
        </script>

        <fieldset id="restock-fieldset-price-group">
            <div class="pure-control-group">
                <fieldset class="pure-group inline-radio">
-                    {{ render_field(form.restock_settings.in_stock_processing) }}
+                    {{ render_field(form.processor_config_restock_diff.in_stock_processing) }}
                </fieldset>
                <fieldset class="pure-group">
-                    {{ render_checkbox_field(form.restock_settings.follow_price_changes) }}
+                    {{ render_checkbox_field(form.processor_config_restock_diff.follow_price_changes) }}
                    <span class="pure-form-message-inline">Changes in price should trigger a notification</span>
                </fieldset>
-                <fieldset class="pure-group price-change-minmax">               
-                    {{ render_field(form.restock_settings.price_change_min, placeholder=watch.get('restock', {}).get('price')) }}
+                <fieldset class="pure-group price-change-minmax">
+                    {{ render_field(form.processor_config_restock_diff.price_change_min, placeholder=watch.get('restock', {}).get('price')) }}
                    <span class="pure-form-message-inline">Minimum amount, Trigger a change/notification when the price drops <i>below</i> this value.</span>
                </fieldset>
                <fieldset class="pure-group price-change-minmax">
-                    {{ render_field(form.restock_settings.price_change_max, placeholder=watch.get('restock', {}).get('price')) }}
+                    {{ render_field(form.processor_config_restock_diff.price_change_max, placeholder=watch.get('restock', {}).get('price')) }}
                    <span class="pure-form-message-inline">Maximum amount, Trigger a change/notification when the price rises <i>above</i> this value.</span>
                </fieldset>
                <fieldset class="pure-group price-change-minmax">
-                    {{ render_field(form.restock_settings.price_change_threshold_percent) }}
+                    {{ render_field(form.processor_config_restock_diff.price_change_threshold_percent) }}
                    <span class="pure-form-message-inline">Price must change more than this % to trigger a change since the first check.</span><br>
                    <span class="pure-form-message-inline">For example, If the product is $1,000 USD originally, <strong>2%</strong> would mean it has to change more than $20 since the first check.</span><br>
-                </fieldset>                
+                </fieldset>
            </div>
        </fieldset>
        """
@@ -0,0 +1,300 @@
+"""
+LLM fallback plugin for price and restock info extraction.
+
+When the built-in structured-metadata extraction (JSON-LD, microdata, OpenGraph)
+fails to produce both a price and availability, this plugin is called as a last
+resort.  It sends a trimmed, HTML-stripped version of the page to the configured
+LLM and asks it to return a structured JSON answer.
+
+The module-level `datastore` variable is injected at startup by
+`inject_datastore_into_plugins()` in pluggy_interface.py.
+"""
+import json
+import re
+from loguru import logger
+from changedetectionio.pluggy_interface import hookimpl
+from changedetectionio.llm.evaluator import apply_local_token_multiplier
+
+# Injected at startup by inject_datastore_into_plugins()
+datastore = None
+
+SYSTEM_PROMPT = (
+    'You are an expert price and restock extraction utility. '
+    'Your task is to analyse a product page and determine the price and stock status of the MAIN product only.\n\n'
+
+    'AVAILABILITY — treat as "in stock":\n'
+    '- Action buttons near the product: "Add to cart", "Add to basket", "Buy now", '
+    '"Order now", "Purchase", "Import", "Add to bag", "Add to trolley", "In stock", '
+    '"Available", "Ships in X days/weeks", "In store", "Pick up today".\n'
+    '- "Pre-order" or "Reserve" — the item is orderable, treat as "in stock".\n'
+    '- "Only X left", "Almost gone", "Low stock", "Limited availability" — still in stock.\n'
+    '- "Request a quote" or "Contact us for pricing" — item is available, price is null.\n'
+    '- IMPORTANT: Ignore cart/basket/bag links in the page HEADER or navigation bar '
+    '(e.g. a shopping cart icon showing item count). That reflects what is already in '
+    'the visitor\'s cart — it says nothing about whether THIS product is available.\n\n'
+
+    'PRICE — what NOT to use:\n'
+    '- A "$0.00" or "0" that appears near header/nav links such as "Login", "Wishlist", '
+    '"Contact Us", "My Account" is an empty shopping-cart indicator, NOT the product price. '
+    'Ignore it entirely — return null for price rather than 0 in this situation.\n'
+    '- Only return 0 (free) when the page clearly states the product itself costs nothing '
+    '(e.g. "Free", "Free download", "Price: $0").\n\n'
+
+    'AVAILABILITY — treat as "out of stock":\n'
+    '- "Out of stock", "Sold out", "Unavailable", "Currently unavailable", '
+    '"Temporarily out of stock", "Discontinued", "No longer available", '
+    '"Notify me when available", "Email me when back", "Join waitlist".\n\n'
+
+    'AVAILABILITY — return null when uncertain:\n'
+    '- The page asks the user to select a size, colour, or other variant first '
+    '("Select an option", "Choose a size") — availability depends on the variant, so return null.\n'
+    '- You cannot clearly tell from the page content whether the item is available.\n\n'
+
+    'PRICE rules:\n'
+    '- Extract the main selling price as a plain number, no currency symbol.\n'
+    '- Prices may use any popular locale format — interpret them all correctly and return a plain decimal number. '
+    'Examples: "10 000 Kč" = 10000, "1.299,95 €" = 1299.95, "1,299.95" = 1299.95, '
+    '"10 000,50" = 10000.50, "£1.299" = 1299, "¥10000" = 10000.\n'
+    '- If both an original (crossed-out) price and a sale/current price appear, use the sale price.\n'
+    '- "From $X" or "Starting at $X" are teaser prices — prefer a definite price or return null.\n'
+    '- A price of 0 (free) is valid — return 0, not null.\n'
+    '- If pricing requires a quote or login, return null for price.\n'
+    '- Ignore prices shown in search/filter UI elements (e.g. "Price from: — to:").\n'
+    '- IMPORTANT: Ignore ALL prices that appear inside or below recommendation/discovery blocks '
+    'such as: "Similar items", "You may also like", "Customers also bought", '
+    '"Based on your browsing", "Based on your shopping", "Frequently bought together", '
+    '"People also viewed", "Related products", "Sponsored products", "More like this", '
+    '"Other sellers", "Compare with similar items". '
+    'These sections contain prices for OTHER products, not the main product.\n'
+    '- When multiple prices appear on the page, prefer the price that is positioned '
+    'earliest/highest in the page content — it is almost always the main product price. '
+    'Prices appearing after large blocks of descriptive text or review sections are '
+    'likely from recommendation widgets and should be ignored.\n\n'
+
+    'CLASSIFIEDS AND LISTING PAGES:\n'
+    '- On classifieds or marketplace sites (e.g. eBay listings, Craigslist, Bazoš, Gumtree), '
+    'if a price is shown alongside seller contact details or a "Contact seller" link, '
+    'treat the item as "instock" — the listing being active means it is available.\n\n'
+
+    'Return ONLY a JSON object with exactly these three keys:\n'
+    '  "price"        — number or null\n'
+    '  "currency"     — ISO-4217 code (USD, EUR, GBP …) or null\n'
+    '  "availability" — exactly one of: "instock", "outofstock", or null\n'
+    '                   Use "instock" when the product can be ordered/purchased.\n'
+    '                   Use "outofstock" when it cannot.\n'
+    '                   Use null when you genuinely cannot tell.\n'
+    'No markdown, no backticks, no explanation — pure JSON only.'
+)
+
+_MAX_CONTENT_CHARS = 8_000
+
+
+def _extract_jsonld(html_content: str) -> str:
+    """Extract JSON-LD blocks — these contain reliable structured product data."""
+    blocks = re.findall(
+        r'<script[^>]+type=["\']application/ld\+json["\'][^>]*>(.*?)</script>',
+        html_content, flags=re.DOTALL | re.IGNORECASE
+    )
+    if not blocks:
+        return ''
+    combined = ' '.join(b.strip() for b in blocks)
+    return combined[:2000]
+
+
+# Semantic tags always treated as chrome (nav/header/footer)
+_CHROME_TAGS = {'nav', 'header', 'footer', 'aside'}
+
+# id/class fragments that strongly indicate navigation or site-chrome
+_CHROME_PATTERNS = re.compile(
+    r'\b(nav|navigation|navbar|menu|mega-menu|breadcrumb|breadcrumbs?|'
+    r'site-header|page-header|top-bar|top-nav|top-header|mobile-nav|header-bar|'
+    r'site-footer|page-footer|footer-links|related|similar|'
+    r'you-?may-?also|customers?-?also|frequently-?bought|'
+    r'people-?also|sponsored|recommendation|widget|sidebar|'
+    r'cross-?sell|up-?sell)\b',
+    re.IGNORECASE,
+)
+
+
+def _remove_chrome(html_content: str) -> str:
+    """Use BS4 to strip navigation, header, footer and recommendation noise.
+
+    Uses html.parser (built-in, no lxml) to avoid memory leak issues.
+    Falls back to the original HTML string if BS4 fails for any reason.
+    """
+    try:
+        from bs4 import BeautifulSoup, Tag
+        soup = BeautifulSoup(html_content, 'html.parser')
+
+        # Snapshot the full tag list before any decompositions so we don't
+        # mutate the tree while iterating it.  After a parent is decomposed
+        # its children become orphans (parent=None) — skip those.
+        for tag in list(soup.find_all(True)):
+            if not isinstance(tag, Tag) or tag.parent is None:
+                continue
+            name = tag.name or ''
+            if name in _CHROME_TAGS:
+                tag.decompose()
+                continue
+            try:
+                cls_list = tag.get('class') or []
+                cls_str = ' '.join(cls_list) if isinstance(cls_list, list) else str(cls_list)
+                id_str = tag.get('id') or ''
+            except Exception:
+                continue
+            if _CHROME_PATTERNS.search(cls_str + ' ' + id_str):
+                tag.decompose()
+
+        return str(soup)
+    except Exception as e:
+        logger.debug(f"BS4 chrome removal failed ({e}), using raw HTML")
+        return html_content
+
+
+def _strip_html(html_content: str) -> str:
+    """HTML-to-text for LLM consumption.
+
+    1. Extracts JSON-LD (structured product data) to prepend.
+    2. Strips nav/header/footer/recommendation blocks via BS4.
+    3. Removes all remaining tags and collapses whitespace.
+    JSON-LD is prepended so reliable price/availability data is always visible
+    to the LLM regardless of how deep it sits in the page.
+    """
+    jsonld = _extract_jsonld(html_content)
+
+    # Remove site-chrome before generic tag stripping
+    cleaned = _remove_chrome(html_content)
+
+    # Drop HTML comments (can contain large disabled markup blocks)
+    text = re.sub(r'<!--.*?-->', ' ', cleaned, flags=re.DOTALL)
+    # Drop all <script> and <style> blocks
+    text = re.sub(r'<(script|style)[^>]*>.*?</(script|style)>', ' ', text, flags=re.DOTALL | re.IGNORECASE)
+    # Strip remaining tags
+    text = re.sub(r'<[^>]+>', ' ', text)
+    # Decode common entities
+    text = (text
+            .replace('&nbsp;', ' ')
+            .replace('&amp;', '&')
+            .replace('&lt;', '<')
+            .replace('&gt;', '>')
+            .replace('&quot;', '"')
+            .replace('&#39;', "'"))
+    text = re.sub(r'\s+', ' ', text).strip()
+
+    if jsonld:
+        budget = _MAX_CONTENT_CHARS - len(jsonld) - 1
+        return (jsonld + ' ' + text[:budget]).strip()
+    return text[:_MAX_CONTENT_CHARS]
+
+
+@hookimpl
+def get_itemprop_availability_override(content, fetcher_name, fetcher_instance, url, llm_intent=None):
+    """Use an LLM as a last-resort fallback for price and restock extraction."""
+    global datastore
+
+    if datastore is None:
+        logger.debug("LLM restock fallback: no datastore injected yet, skipping")
+        return None
+
+    try:
+        from changedetectionio.llm.evaluator import _runtime_llm_config, accumulate_global_tokens, get_llm_settings
+        from changedetectionio.llm import client as llm_client
+    except ImportError as e:
+        logger.debug(f"LLM restock fallback: LLM libraries not available ({e})")
+        return None
+
+    # Gate on the user setting (default True — enabled out of the box)
+    if not get_llm_settings(datastore).restock_use_fallback_extract:
+        logger.debug("LLM restock fallback: disabled in settings")
+        return None
+
+    # _runtime_llm_config returns None (with a debug log) when the master 'llm_enabled'
+    # toggle is off, so this path is gated for free.
+    llm_cfg = _runtime_llm_config(datastore)
+    if not llm_cfg or not llm_cfg.get('model'):
+        logger.debug("LLM restock fallback: no LLM model configured or LLM disabled, skipping")
+        return None
+
+    text_content = _strip_html(content) if content else ''
+    logger.debug(f"LLM restock fallback: stripped HTML to {len(text_content)} chars for {url}")
+    if not text_content.strip():
+        logger.debug("LLM restock fallback: no text content after stripping HTML")
+        return None
+
+    logger.info(f"LLM restock fallback: using LLM ({llm_cfg['model']}) for price/stock extraction - {url}")
+
+    user_prompt = f'URL: {url or "unknown"}\n\nPage content:\n{text_content}'
+    if llm_intent:
+        user_prompt += f'\n\nUser notification intent: {llm_intent}'
+
+    try:
+        raw, tokens, input_tokens, output_tokens = llm_client.completion(
+            model=llm_cfg['model'],
+            messages=[
+                {'role': 'system', 'content': SYSTEM_PROMPT},
+                {'role': 'user', 'content': user_prompt},
+            ],
+            api_key=llm_cfg.get('api_key'),
+            api_base=llm_cfg.get('api_base'),
+            # 80 fits a {price, currency, availability} JSON answer comfortably for cloud
+            # models. Local reasoning models burn most of that on chain-of-thought before
+            # the JSON lands — the multiplier scales it up only when provider_kind says so.
+            max_tokens=apply_local_token_multiplier(80, llm_cfg),
+        )
+
+        accumulate_global_tokens(
+            datastore, tokens,
+            input_tokens=input_tokens,
+            output_tokens=output_tokens,
+            model=llm_cfg['model'],
+        )
+
+        # Strip optional markdown fences the model might add
+        raw = raw.strip()
+        if raw.startswith('```'):
+            raw = re.sub(r'^```[a-z]*\n?', '', raw)
+            raw = raw.rstrip('`').strip()
+
+        logger.debug(f"LLM restock fallback raw response: {raw!r}")
+
+        result = json.loads(raw)
+
+        price = result.get('price')
+        currency = result.get('currency') or None
+        availability = result.get('availability') or None
+
+        # Normalise price to float
+        if price is not None:
+            try:
+                if isinstance(price, str):
+                    price = float(re.sub(r'[^\d.]', '', price))
+                else:
+                    price = float(price)
+            except (ValueError, TypeError):
+                logger.warning(f"LLM restock fallback: could not convert price {price!r} to float, ignoring")
+                price = None
+
+        if price is None and not availability:
+            logger.info(f"LLM restock fallback: LLM returned no usable price or availability for {url} (raw: {raw!r})")
+            return None
+
+        logger.info(
+            f"LLM restock fallback result: price={price} currency={currency} "
+            f"availability={availability!r} url={url}"
+        )
+        return {
+            'price': price,
+            'currency': currency,
+            'availability': availability,
+            '_tokens': tokens,
+            '_input_tokens': input_tokens,
+            '_output_tokens': output_tokens,
+            '_model': llm_cfg['model'],
+        }
+
+    except json.JSONDecodeError as e:
+        logger.warning(f"LLM restock fallback: JSON parse failed ({e}) - raw response was: {raw!r}")
+        return None
+    except Exception as e:
+        logger.warning(f"LLM restock fallback: extraction failed for {url}: {e}")
+        return None
@@ -437,26 +437,32 @@ class perform_site_check(difference_detection_processor):

        # Only try to process restock information (like scraping for keywords) if the page was actually rendered correctly.
        # Otherwise it will assume "in stock" because nothing suggesting the opposite was found
-        from ...html_tools import html_to_text
-        text = html_to_text(self.fetcher.content)
-        logger.debug(f"Length of text after conversion: {len(text)}")
-        if not len(text):
-            from ...content_fetchers.exceptions import ReplyWithContentButNoText
-            raise ReplyWithContentButNoText(url=watch.link,
-                                            status_code=self.fetcher.get_last_status_code(),
-                                            screenshot=self.fetcher.screenshot,
-                                            html_content=self.fetcher.content,
-                                            xpath_data=self.fetcher.xpath_data
-                                            )
+#useless
+#        from ...html_tools import html_to_text
+#        text = html_to_text(self.fetcher.content)
+#        logger.debug(f"Length of text after conversion: {len(text)}")
+#        if not len(text):
+#            from ...content_fetchers.exceptions import ReplyWithContentButNoText
+#            raise ReplyWithContentButNoText(url=watch.link,
+#                                            status_code=self.fetcher.get_last_status_code(),
+#                                            screenshot=self.fetcher.screenshot,
+#                                            html_content=self.fetcher.content,
+#                                            xpath_data=self.fetcher.xpath_data
+#                                            )

        # Which restock settings to compare against?
-        restock_settings = watch.get('restock_settings', {})
+        # Settings are stored in restock_diff.json (migrated from watch.json by update_30).
+        _extra_config = self.get_extra_watch_config('restock_diff.json')
+        restock_settings = _extra_config.get('restock_diff') or {
+            'follow_price_changes': True,
+            'in_stock_processing': 'in_stock_only',
+        }

        # See if any tags have 'activate for individual watches in this tag/group?' enabled and use the first we find
        for tag_uuid in watch.get('tags'):
            tag = self.datastore.data['settings']['application']['tags'].get(tag_uuid, {})
            if tag.get('overrides_watch'):
-                restock_settings = tag.get('restock_settings', {})
+                restock_settings = tag.get('processor_config_restock_diff') or {}
                logger.info(f"Watch {watch.get('uuid')} - Tag '{tag.get('title')}' selected for restock settings override")
                break

@@ -480,8 +486,7 @@ class perform_site_check(difference_detection_processor):
        has_price = itemprop_availability.get('price') is not None
        has_availability = itemprop_availability.get('availability') is not None

-        # @TODO !!! some setting like "Use as fallback" or "always use", "t
-        if not (has_price and has_availability) or True:
+        if not (has_price and has_availability):
            from changedetectionio.pluggy_interface import get_itemprop_availability_from_plugin
            fetcher_name = watch.get('fetch_backend', 'html_requests')

@@ -500,9 +505,23 @@ class perform_site_check(difference_detection_processor):
            # Try plugin override - plugins can decide if they support this fetcher
            if fetcher_name:
                logger.debug(f"Calling extra plugins for getting item price/availability (fetcher: {fetcher_name})")
-                plugin_availability = get_itemprop_availability_from_plugin(self.fetcher.content, fetcher_name, self.fetcher, watch.link)
+                from changedetectionio.llm.evaluator import resolve_intent
+                _llm_intent, _ = resolve_intent(watch, self.datastore)
+                plugin_availability = get_itemprop_availability_from_plugin(self.fetcher.content, fetcher_name, self.fetcher, watch.link, llm_intent=_llm_intent or None)

                if plugin_availability:
+                    # Extract and strip LLM token metadata before using as Restock data
+                    _plugin_tokens = plugin_availability.pop('_tokens', 0)
+                    _plugin_input_tokens = plugin_availability.pop('_input_tokens', 0)
+                    _plugin_output_tokens = plugin_availability.pop('_output_tokens', 0)
+                    _plugin_model = plugin_availability.pop('_model', '')
+
+                    # Update per-watch token counters directly on the watch (same
+                    # pattern as evaluator.py) so they're committed when update_watch runs
+                    if _plugin_tokens:
+                        watch['llm_last_tokens_used'] = _plugin_tokens
+                        watch['llm_tokens_used_cumulative'] = (watch.get('llm_tokens_used_cumulative') or 0) + _plugin_tokens
+
                    # Plugin provided better data, use it
                    plugin_has_price = plugin_availability.get('price') is not None
                    plugin_has_availability = plugin_availability.get('availability') is not None
@@ -283,4 +283,7 @@ def query_price_availability(extracted_data):
    if not result.get('availability') and 'availability' in microdata:
        result['availability'] = microdata['availability']

+    # result['price'] could be float or str here, depending on the website, for example it might contain "1,00" commas, etc.
+    # using something like babel you need to know the locale of the website and even then it can be problematic
+    # we dont really do anything with the price data so far.. so just accept it the way it comes.
    return result
@@ -3,11 +3,11 @@
 {% block content %}
    <div class="tabs">
    <ul>
-        {% if last_error_text %}<li class="tab" id="error-text-tab"><a href="{{ url_for('ui.ui_diff.diff_history_page', uuid=uuid)}}#error-text">Error Text</a></li> {% endif %}
-        {% if last_error_screenshot %}<li class="tab" id="error-screenshot-tab"><a href="{{ url_for('ui.ui_diff.diff_history_page', uuid=uuid)}}#error-screenshot">Error Screenshot</a></li> {% endif %}
-        <li class="tab" id=""><a href="{{ url_for('ui.ui_diff.diff_history_page', uuid=uuid)}}#text">Text</a></li>
-        <li class="tab" id="screenshot-tab"><a href="{{ url_for('ui.ui_diff.diff_history_page', uuid=uuid)}}#screenshot">Screenshot</a></li>
-        <li class="tab active" id="extract-tab"><a href="{{ url_for('ui.ui_diff.diff_history_page_extract_GET', uuid=uuid)}}">Extract Data</a></li>
+        {% if last_error_text %}<li class="tab" id="error-text-tab"><a href="{{ url_for('ui.ui_diff.diff_history_page', uuid=uuid)}}#error-text">{{ _('Error Text') }}</a></li> {% endif %}
+        {% if last_error_screenshot %}<li class="tab" id="error-screenshot-tab"><a href="{{ url_for('ui.ui_diff.diff_history_page', uuid=uuid)}}#error-screenshot">{{ _('Error Screenshot') }}</a></li> {% endif %}
+        <li class="tab" id=""><a href="{{ url_for('ui.ui_diff.diff_history_page', uuid=uuid)}}#text">{{ _('Text') }}</a></li>
+        <li class="tab" id="screenshot-tab"><a href="{{ url_for('ui.ui_diff.diff_history_page', uuid=uuid)}}#screenshot">{{ _('Screenshot') }}</a></li>
+        <li class="tab active" id="extract-tab"><a href="{{ url_for('ui.ui_diff.diff_history_page_extract_GET', uuid=uuid)}}">{{ _('Extract Data') }}</a></li>
    </ul>
 </div>

@@ -17,23 +17,23 @@
        <form id="extract-data-form" class="pure-form pure-form-stacked edit-form"  action="{{ url_for('ui.ui_diff.diff_history_page_extract_POST', uuid=uuid) }}"  method="POST">
            <input type="hidden" name="csrf_token" value="{{ csrf_token() }}">

-            <p>This tool will extract text data from all of the watch history.</p>
+            <p>{{ _('This tool will extract text data from all of the watch history.') }}</p>

            <div class="pure-control-group">
                {{ render_field(extract_form.extract_regex) }}
                <span class="pure-form-message-inline">
-                    A <strong>RegEx</strong> is a pattern that identifies exactly which part inside of the text that you want to extract.<br>
+                    {{ _('A <strong>RegEx</strong> is a pattern that identifies exactly which part inside of the text that you want to extract.')|safe }}<br>

                    <p>
-                        For example, to extract only the numbers from text &dash;<br>
-                        <strong>Raw text</strong>: <code>Temperature <span style="color: red">5.5</span>°C in Sydney</code><br>
-                        <strong>RegEx to extract:</strong> <code>Temperature <span style="color: red">([0-9\.]+)</span></code><br>
+                        {{ _('For example, to extract only the numbers from text') }} &dash;<br>
+                        <strong>{{ _('Raw text') }}</strong>: <code>Temperature <span style="color: red">5.5</span>°C in Sydney</code><br>
+                        <strong>{{ _('RegEx to extract:') }}</strong> <code>Temperature <span style="color: red">([0-9\.]+)</span></code><br>
                    </p>
                    <p>
-                        <a href="https://RegExr.com/">Be sure to test your RegEx here.</a>
+                        <a href="https://RegExr.com/">{{ _('Be sure to test your RegEx here.') }}</a>
                    </p>
                    <p>
-                        Each RegEx group bracket <code>()</code> will be in its own column, the first column value is always the date.
+                        {{ _('Each RegEx group bracket') }} <code>()</code> {{ _('will be in its own column, the first column value is always the date.') }}
                    </p>
                </span>
            </div>
@@ -35,8 +35,52 @@ def _task(watch, update_handler):
    return text_after_filter


+def _compute_ignore_line_numbers_for_preview(text_pre_extract, ignore_patterns, extract_patterns):
+    """1-indexed output line numbers in the post-extract display that correspond
+    to input lines matching ignore_text patterns.
+
+    Needed because extract_text (#4138) transforms line content — e.g. "0.54.10"
+    becomes ".54.10" — so a substring match for "0.54.10" against the post-extract
+    text fails and the preview UI can no longer mark the line as ignored. We find
+    the ignored line numbers in the pre-extract text and replay extract_by_regex
+    line-by-line to map them forward.
+    """
+    from changedetectionio import html_tools
+    from changedetectionio.processors.text_json_diff.processor import ContentTransformer
+
+    if not text_pre_extract or not ignore_patterns:
+        return []
+
+    ignored_input_lines = set(
+        html_tools.strip_ignore_text(
+            content=text_pre_extract,
+            wordlist=ignore_patterns,
+            mode='line numbers'
+        )
+    )
+    if not ignored_input_lines:
+        return []
+
+    if not extract_patterns:
+        return sorted(ignored_input_lines)
+
+    # Replay extract_by_regex per-line. Each emitted match ends with exactly one
+    # '\n', so counting newlines tells us how many output lines this input produced.
+    output_line_counter = 0
+    result = []
+    for input_idx, line in enumerate(text_pre_extract.splitlines()):
+        is_ignored = (input_idx + 1) in ignored_input_lines
+        matches_in_line = ContentTransformer.extract_by_regex(line, extract_patterns).count('\n')
+        for _ in range(matches_in_line):
+            output_line_counter += 1
+            if is_ignored:
+                result.append(output_line_counter)
+
+    return result
+
+
 def prepare_filter_prevew(datastore, watch_uuid, form_data):
-    '''Used by @app.route("/edit/<string:uuid>/preview-rendered", methods=['POST'])'''
+    '''Used by @app.route("/edit/<uuid_str:uuid>/preview-rendered", methods=['POST'])'''
    from changedetectionio import forms, html_tools
    from changedetectionio.model.Watch import model as watch_model
    from concurrent.futures import ThreadPoolExecutor
@@ -50,6 +94,7 @@ def prepare_filter_prevew(datastore, watch_uuid, form_data):

    text_after_filter = ''
    text_before_filter = ''
+    text_pre_extract = ''
    trigger_line_numbers = []
    ignore_line_numbers = []
    blocked_line_numbers = []
@@ -65,6 +110,12 @@ def prepare_filter_prevew(datastore, watch_uuid, form_data):
        # Only update vars that came in via the AJAX post
        p = {k: v for k, v in form.data.items() if k in form_data.keys()}
        tmp_watch.update(p)
+
+        # Apply llm_intent from form directly — it's not part of processor_text_json_diff_form
+        # but the AJAX sends all visible inputs, so it arrives in form_data
+        if hasattr(form_data, 'get') and 'llm_intent' in form_data:
+            tmp_watch['llm_intent'] = (form_data.get('llm_intent') or '').strip()
+
        blank_watch_no_filters = watch_model(datastore_path=datastore.datastore_path, __datastore=datastore.data)
        blank_watch_no_filters['url'] = tmp_watch.get('url')

@@ -83,15 +134,22 @@ def prepare_filter_prevew(datastore, watch_uuid, form_data):
            update_handler.fetcher.content = str(decompressed_data) # str() because playwright/puppeteer/requests return string
            update_handler.fetcher.headers['content-type'] = tmp_watch.get('content-type')

-            # Process our watch with filters and the HTML from disk, and also a blank watch with no filters but also with the same HTML from disk
+            # Process our watch with filters and the HTML from disk, and also a blank watch with no filters but also with the same HTML from disk.
+            # The third task runs with extract_text cleared so we can compute ignore_line_numbers
+            # against the pre-extract text (extract_text transforms lines so post-extract substring
+            # matching for ignore patterns would otherwise fail — see #4138 follow-up).
            # Do this as parallel threads (not processes) to avoid pickle issues with Lock objects
+            tmp_watch_no_extract = deepcopy(tmp_watch)
+            tmp_watch_no_extract['extract_text'] = []
            try:
-                with ThreadPoolExecutor(max_workers=2) as executor:
+                with ThreadPoolExecutor(max_workers=3) as executor:
                    future1 = executor.submit(_task, tmp_watch, update_handler)
                    future2 = executor.submit(_task, blank_watch_no_filters, update_handler)
+                    future3 = executor.submit(_task, tmp_watch_no_extract, update_handler)

                    text_after_filter = future1.result()
                    text_before_filter = future2.result()
+                    text_pre_extract = future3.result()
            except Exception as e:
                x=1

@@ -105,10 +163,11 @@ def prepare_filter_prevew(datastore, watch_uuid, form_data):

    try:
        text_to_ignore = tmp_watch.get('ignore_text', []) + datastore.data['settings']['application'].get('global_ignore_text', [])
-        ignore_line_numbers = html_tools.strip_ignore_text(content=text_after_filter,
-                                                           wordlist=text_to_ignore,
-                                                           mode='line numbers'
-                                                           )
+        ignore_line_numbers = _compute_ignore_line_numbers_for_preview(
+            text_pre_extract=text_pre_extract,
+            ignore_patterns=text_to_ignore,
+            extract_patterns=tmp_watch.get('extract_text', [])
+        )
    except Exception as e:
        text_before_filter = f"Error: {str(e)}"

@@ -120,6 +179,18 @@ def prepare_filter_prevew(datastore, watch_uuid, form_data):
    except Exception as e:
        text_before_filter = f"Error: {str(e)}"

+    # LLM preview extraction — asks the LLM to directly answer the intent
+    # against the current filtered content (no diff comparison).
+    # e.g. intent "how many articles?" → answer "30 articles listed"
+    # Results are NOT cached back to the real watch.
+    llm_evaluation = None
+    try:
+        from changedetectionio.llm.evaluator import preview_extract
+        if text_after_filter and text_after_filter.strip() not in ('', 'Empty content'):
+            llm_evaluation = preview_extract(tmp_watch, datastore, content=text_after_filter)
+    except Exception as e:
+        logger.warning(f"LLM preview evaluation failed for {watch_uuid}: {e}")
+
    logger.trace(f"Parsed in {time.time() - now:.3f}s")

    return ({
@@ -128,6 +199,7 @@ def prepare_filter_prevew(datastore, watch_uuid, form_data):
        'blocked_line_numbers': blocked_line_numbers,
        'duration': time.time() - now,
        'ignore_line_numbers': ignore_line_numbers,
+        'llm_evaluation': llm_evaluation,
        'trigger_line_numbers': trigger_line_numbers,
        })

@@ -7,6 +7,7 @@ a side-by-side or unified diff view with syntax highlighting and change markers.

 import os
 import time
+from flask_babel import gettext
 from loguru import logger

 from changedetectionio import diff, strtobool
@@ -97,6 +98,7 @@ DIFF_PREFERENCES_CONFIG = {
    'added': {'default': True, 'type': 'bool'},
    'replaced': {'default': True, 'type': 'bool'},
    'type': {'default': 'diffLines', 'type': 'value'},
+    'llm_all_changes': {'default': False, 'type': 'bool'},
 }

 def render(watch, datastore, request, url_for, render_template, flash, redirect, extract_form=None):
@@ -154,11 +156,7 @@ def render(watch, datastore, request, url_for, render_template, flash, redirect,

    screenshot_url = watch.get_screenshot()

-    system_uses_webdriver = datastore.data['settings']['application']['fetch_backend'] == 'html_webdriver'
-
-    is_html_webdriver = False
-    if (watch.get('fetch_backend') == 'system' and system_uses_webdriver) or watch.get('fetch_backend') == 'html_webdriver' or watch.get('fetch_backend', '').startswith('extra_browser_'):
-        is_html_webdriver = True
+    is_html_webdriver = watch.fetcher_supports_screenshots

    password_enabled_and_share_is_off = False
    if datastore.data['settings']['application'].get('password') or os.getenv("SALTED_PASS", False):
@@ -202,6 +200,36 @@ def render(watch, datastore, request, url_for, render_template, flash, redirect,
    if str(from_version) != str(dates[-2]) or str(to_version) != str(dates[-1]):
        note = 'Note: You are not viewing the latest changes.'

+    llm_configured = bool(
+        datastore.data.get('settings', {}).get('application', {}).get('llm', {}).get('model')
+    )
+
+    # Load cached AI diff summary for this exact from→to + prompt combination
+    viewing_latest = str(to_version) == str(dates[-1])
+    llm_diff_summary = ''
+    llm_summary_prompt = ''
+    if llm_configured:
+        try:
+            from changedetectionio.llm.evaluator import (
+                get_effective_summary_prompt, build_summary_cache_prompt,
+            )
+            _prompt = get_effective_summary_prompt(watch, datastore)
+            llm_summary_prompt = _prompt
+            # Must match the cache_prompt the worker writes and the UI ajax route reads —
+            # using UI default diff prefs so the initial render finds the worker's pre-cache.
+            from changedetectionio.llm.evaluator import get_llm_settings
+            _ls = get_llm_settings(datastore)
+            _max_summary_tokens = _ls.max_summary_tokens
+            _llm_model = _ls.model
+            _cache_prompt = build_summary_cache_prompt(
+                effective_prompt=_prompt,
+                max_summary_tokens=_max_summary_tokens,
+                model=_llm_model,
+            )
+            llm_diff_summary = watch.get_llm_diff_summary(from_version, to_version, prompt=_cache_prompt)
+        except Exception as e:
+            logger.warning(f"Could not load llm-diff-summary for {uuid}: {e}")
+
    output = render_template("diff.html",
                             #initial_scroll_line_number=100,
                             bottom_horizontal_offscreen_contents=offscreen_content,
@@ -209,9 +237,9 @@ def render(watch, datastore, request, url_for, render_template, flash, redirect,
                             current_diff_url=watch['url'],
                             diff_cell_grid=diff_cell_grid,
                             diff_prefs=diff_prefs,
-                             extra_classes='difference-page',
+                             extra_classes=' '.join(filter(None, ['difference-page', 'llm-configured' if llm_configured else ''])),
                             extra_stylesheets=extra_stylesheets,
-                             extra_title=f" - {watch.label} - History",
+                             extra_title=f" - {watch.label} - {gettext('History')}",
                             extract_form=extract_form,
                             from_version=str(from_version),
                             is_html_webdriver=is_html_webdriver,
@@ -228,5 +256,9 @@ def render(watch, datastore, request, url_for, render_template, flash, redirect,
                             uuid=uuid,
                             versions=dates,  # All except current/last
                             watch_a=watch,
+                             llm_configured=llm_configured,
+                             llm_diff_summary=llm_diff_summary,
+                             llm_summary_prompt=llm_summary_prompt,
+                             viewing_latest=viewing_latest,
                             )
    return output
@@ -85,6 +85,10 @@ class FilterConfig:
            self._subtractive_selectors_cache = [*tag_selectors, *watch_selectors, *global_selectors]
        return self._subtractive_selectors_cache

+    @property
+    def extract_lines_containing(self):
+        return self._get_merged_rules('extract_lines_containing')
+
    @property
    def extract_text(self):
        return self._get_merged_rules('extract_text')
@@ -101,6 +105,30 @@ class FilterConfig:
    def text_should_not_be_present(self):
        return self._get_merged_rules('text_should_not_be_present')

+    def get_filter_config_hash(self):
+        """
+        Stable hash of the effective filter configuration.
+
+        Used by the skip-logic in run_changedetection() so that any change to
+        global settings, tag overrides, or watch filters automatically invalidates
+        the raw-content-unchanged shortcut — without needing scattered
+        clear_all_last_checksums() calls at every settings mutation site.
+        """
+        app = self.datastore.data['settings']['application']
+        config = {
+            'extract_lines_containing': sorted(self.extract_lines_containing),
+            'extract_text':             sorted(self.extract_text),
+            'ignore_text':              sorted(self.ignore_text),
+            'include_filters':          sorted(self.include_filters),
+            'subtractive_selectors':    sorted(self.subtractive_selectors),
+            'text_should_not_be_present': sorted(self.text_should_not_be_present),
+            'trigger_text':             sorted(self.trigger_text),
+            # Global processing flags not captured by the filter lists above
+            'ignore_whitespace':        app.get('ignore_whitespace', False),
+            'strip_ignored_lines':      app.get('strip_ignored_lines', False),
+        }
+        return hashlib.md5(json.dumps(config, sort_keys=True).encode()).hexdigest()
+
    @property
    def has_include_filters(self):
        return bool(self.include_filters) and bool(self.include_filters[0].strip())
@@ -135,6 +163,17 @@ class ContentTransformer:
        text = text.replace("\n\n", "\n")
        return '\n'.join(sorted(text.splitlines(), key=lambda x: x.lower()))

+    @staticmethod
+    def extract_lines_containing(text, substrings):
+        """Keep only lines that contain at least one of the given substrings (case-insensitive)."""
+        needles = [s.lower() for s in substrings if s.strip()]
+        if not needles:
+            return text
+        return '\n'.join(
+            line for line in text.splitlines()
+            if any(needle in line.lower() for needle in needles)
+        )
+
    @staticmethod
    def extract_by_regex(text, regex_patterns):
        """Extract text matching regex patterns."""
@@ -347,6 +386,7 @@ class ContentProcessor:
    def extract_text_from_html(self, html_content, stream_content_type):
        """Convert HTML to plain text."""
        do_anchor = self.datastore.data["settings"]["application"].get("render_anchor_tag_content", False)
+
        return html_tools.html_to_text(
            html_content=html_content,
            render_anchor_tag_content=do_anchor,
@@ -376,19 +416,26 @@ class perform_site_check(difference_detection_processor):
            raise Exception("Watch no longer exists.")

        current_raw_document_checksum = self.get_raw_document_checksum()
-        # Skip processing only if BOTH conditions are true:
-        # 1. HTML content unchanged (checksum matches last saved checksum)
-        # 2. Watch configuration was not edited (including trigger_text, filters, etc.)
-        # The was_edited flag handles all watch configuration changes, so we don't need
-        # separate checks for trigger_text or other processing rules.
+
+        # Build filter config up front so we can hash it for the skip check.
+        filter_config = FilterConfig(watch, self.datastore)
+        current_filter_config_hash = filter_config.get_filter_config_hash()
+
+        # Skip only when ALL of these hold:
+        #   1. raw HTML is unchanged
+        #   2. watch config was not edited (was_edited covers per-watch field changes)
+        #   3. effective filter config is unchanged (covers global/tag setting changes that
+        #      bypass was_edited — e.g. global_ignore_text, global_subtractive_selectors)
+        # last_filter_config_hash being False means first run or upgrade: don't skip.
        if (not force_reprocess and
            not watch.was_edited and
            self.last_raw_content_checksum and
-            self.last_raw_content_checksum == current_raw_document_checksum):
+            self.last_raw_content_checksum == current_raw_document_checksum and
+            watch.get('last_filter_config_hash') and
+            watch.get('last_filter_config_hash') == current_filter_config_hash):
            raise checksumFromPreviousCheckWasTheSame()

-        # Initialize components
-        filter_config = FilterConfig(watch, self.datastore)
+        # Initialize remaining components
        content_processor = ContentProcessor(self.fetcher, watch, filter_config, self.datastore)
        transformer = ContentTransformer()
        rule_engine = RuleEngine()
@@ -409,6 +456,7 @@ class perform_site_check(difference_detection_processor):

        # Save the raw content checksum to file (processor implementation detail, not watch config)
        self.update_last_raw_content_checksum(current_raw_document_checksum)
+        update_obj['last_filter_config_hash'] = current_filter_config_hash

        # === CONTENT PREPROCESSING ===
        # Avoid creating unnecessary intermediate string copies by reassigning only when needed
@@ -447,16 +495,17 @@ class perform_site_check(difference_detection_processor):
        # Start with content reference, avoid copy until modification
        html_content = content

-        # Apply include filters (CSS, XPath, JSON)
-        # Except for plaintext (incase they tried to confuse the system, it will HTML escape
-        #if not stream_content_type.is_plaintext:
-        if filter_config.has_include_filters:
-            html_content = content_processor.apply_include_filters(content, stream_content_type)
-
-        # Apply subtractive selectors
+        # Apply subtractive selectors first so include filters operate on already-cleaned content.
+        # Otherwise a subtractive selector that relies on ancestor context (e.g. ".main .ads")
+        # cannot match after the include filter has extracted the inner element and stripped
+        # the parent wrapper.
        if filter_config.has_subtractive_selectors:
            html_content = content_processor.apply_subtractive_selectors(html_content)

+        # Apply include filters (CSS, XPath, JSON)
+        if filter_config.has_include_filters:
+            html_content = content_processor.apply_include_filters(html_content, stream_content_type)
+
        # === TEXT EXTRACTION ===
        if watch.is_source_type_url:
            # For source URLs, keep raw content
@@ -502,26 +551,43 @@ class perform_site_check(difference_detection_processor):

        update_obj["last_check_status"] = self.fetcher.get_last_status_code()

+        # Snapshot an ignore-applied stream BEFORE extract operations so line-level
+        # ignore patterns still match original content (#4138). Otherwise an extract_text
+        # regex like /(\d+\.\d+\.\d+)/ would transform "v.1.2.1" into "1.2.1" and the
+        # ignore_text pattern "v" would no longer match — meaning changes to ignored
+        # lines would incorrectly affect the checksum.
+        text_for_checksuming = None
+        if filter_config.ignore_text:
+            text_for_checksuming = html_tools.strip_ignore_text(stripped_text, filter_config.ignore_text)
+
+        # === LINE FILTER (plain-text substring) ===
+        if filter_config.extract_lines_containing:
+            stripped_text = transformer.extract_lines_containing(stripped_text, filter_config.extract_lines_containing)
+            if text_for_checksuming is not None:
+                text_for_checksuming = transformer.extract_lines_containing(text_for_checksuming, filter_config.extract_lines_containing)
+
        # === REGEX EXTRACTION ===
        if filter_config.extract_text:
-            extracted = transformer.extract_by_regex(stripped_text, filter_config.extract_text)
-            stripped_text = extracted
+            stripped_text = transformer.extract_by_regex(stripped_text, filter_config.extract_text)
+            if text_for_checksuming is not None:
+                text_for_checksuming = transformer.extract_by_regex(text_for_checksuming, filter_config.extract_text)

        # === MORE TEXT TRANSFORMATIONS ===
        if watch.get('remove_duplicate_lines'):
            stripped_text = transformer.remove_duplicate_lines(stripped_text)
+            if text_for_checksuming is not None:
+                text_for_checksuming = transformer.remove_duplicate_lines(text_for_checksuming)

        if watch.get('sort_text_alphabetically'):
            stripped_text = transformer.sort_alphabetically(stripped_text)
+            if text_for_checksuming is not None:
+                text_for_checksuming = transformer.sort_alphabetically(text_for_checksuming)

        # === CHECKSUM CALCULATION ===
-        text_for_checksuming = stripped_text
-
-        # Apply ignore_text for checksum calculation
-        if filter_config.ignore_text:
-            text_for_checksuming = html_tools.strip_ignore_text(stripped_text, filter_config.ignore_text)
-
-            # Optionally remove ignored lines from output
+        if text_for_checksuming is None:
+            text_for_checksuming = stripped_text
+        else:
+            # Optionally remove ignored lines from displayed output too
            strip_ignored_lines = watch.get('strip_ignored_lines')
            if strip_ignored_lines is None:
                strip_ignored_lines = self.datastore.data['settings']['application'].get('strip_ignored_lines')
@@ -535,8 +601,8 @@ class perform_site_check(difference_detection_processor):
        # === BLOCKING RULES EVALUATION ===
        blocked = False

-        # Check trigger_text
-        if rule_engine.evaluate_trigger_text(stripped_text, filter_config.trigger_text):
+        # Check trigger_text - use text_for_checksuming so ignore_text can suppress trigger_text
+        if rule_engine.evaluate_trigger_text(text_for_checksuming, filter_config.trigger_text):
            blocked = True

        # Check text_should_not_be_present
@@ -29,9 +29,11 @@ def register_watch_operation_handlers(socketio, datastore):
            # Perform the operation
            if op == 'pause':
                watch.toggle_pause()
+                watch.commit()
                logger.info(f"Socket.IO: Toggled pause for watch {uuid}")
            elif op == 'mute':
                watch.toggle_mute()
+                watch.commit()
                logger.info(f"Socket.IO: Toggled mute for watch {uuid}")
            elif op == 'recheck':
                # Import here to avoid circular imports
@@ -199,8 +199,31 @@ def handle_watch_update(socketio, **kwargs):
        logger.error(f"Socket.IO error in handle_watch_update: {str(e)}")


+def _suppress_werkzeug_ws_abrupt_disconnect_noise():
+    """Patch BaseWSGIServer.log to suppress the AssertionError traceback that fires when
+    a browser closes a WebSocket connection mid-handshake (e.g. closing a tab).
+    The exception is caught inside run_wsgi and routed to self.server.log() — it never
+    propagates out, so wrapping run_wsgi doesn't help. Patching the log method is the
+    only reliable intercept point. The error is cosmetic: Socket.IO already handles the
+    disconnect correctly via its own disconnect handler and timeout logic."""
+    try:
+        from werkzeug.serving import BaseWSGIServer
+        _original_log = BaseWSGIServer.log
+
+        def _filtered_log(self, type, message, *args):
+            if type == 'error' and 'write() before start_response' in message:
+                return
+            _original_log(self, type, message, *args)
+
+        BaseWSGIServer.log = _filtered_log
+    except Exception:
+        pass
+
+
 def init_socketio(app, datastore):
    """Initialize SocketIO with the main Flask app"""
+    _suppress_werkzeug_ws_abrupt_disconnect_noise()
+
    import platform
    import sys

@@ -345,4 +368,4 @@ def init_socketio(app, datastore):

    logger.info("Socket.IO initialized and attached to main Flask app")
    logger.info(f"Socket.IO: Registered event handlers: {socketio.handlers if hasattr(socketio, 'handlers') else 'No handlers found'}")
-    return socketio
+    return socketio
@@ -44,12 +44,12 @@ data_sanity_test () {
  cd ..
  TMPDIR=$(mktemp -d)
  PORT_N=$((5000 + RANDOM % (6501 - 5000)))
-  ./changedetection.py -p $PORT_N -d $TMPDIR -u "https://localhost?test-url-is-sanity=1" &
+  ALLOW_IANA_RESTRICTED_ADDRESSES=true ./changedetection.py -p $PORT_N -d $TMPDIR -u "https://localhost?test-url-is-sanity=1" &
  PID=$!
  sleep 5
  kill $PID
  sleep 2
-  ./changedetection.py -p $PORT_N -d $TMPDIR &
+  ALLOW_IANA_RESTRICTED_ADDRESSES=true ./changedetection.py -p $PORT_N -d $TMPDIR &
  PID=$!
  sleep 5
  # On a restart the URL should still be there
--- a/Show More
+++ b/Show More