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

.github/workflows/test-only.yml

@@ -31,33 +31,15 @@ jobs:
             echo "Checking $f"
             msgfmt --check-format -o /dev/null "$f"
           done
-      - name: Lint .po/.pot files with dennis (errors only)
+      - name: Lint .pot template with dennis
         run: |
           pip install "$(grep -E '^dennis ?>=' requirements.txt)"
-          dennis-cmd lint --errorsonly changedetectionio/translations/
-      - name: Lint .pot template with dennis (warnings)
+          dennis-cmd lint --strict changedetectionio/translations/messages.pot
+      - name: Lint .po files with dennis
         run: |
-          output=$(dennis-cmd lint changedetectionio/translations/messages.pot)
-          echo "$output"
-          warnings=$(echo "$output" | awk '/Warnings:/ {print $NF; exit}')
-          if (( ${warnings:-0} > 0 )); then
-            echo "ERROR: ${warnings} dennis warning(s) detected in messages.pot"
-            echo "Fix the warning(s)."
-            exit 1
-          fi
-      - name: Lint .po files with dennis (warnings)
+          dennis-cmd lint --strict --excluderules=W302 changedetectionio/translations/*/LC_MESSAGES/messages.po
         # W302 (unchanged) is excluded due to high false-positive rate in this codebase:
         # many msgstrs intentionally match msgid (units like "AI", "LLM", and proper nouns).
-        run: |
-          output=$(dennis-cmd lint --excluderules=W302 \
-            changedetectionio/translations/*/LC_MESSAGES/messages.po)
-          echo "$output"
-          warnings=$(echo "$output" | awk '/Total number of warnings:/ {print $NF; exit}')
-          if (( ${warnings:-0} > 0 )); then
-            echo "ERROR: ${warnings} dennis warning(s) detected in .po files"
-            echo "Fix the warning(s)."
-            exit 1
-          fi
       - name: Check translation catalog is up-to-date
         run: |
           pip install "$(grep -E '^babel==' requirements.txt)"

.pre-commit-config.yaml

@@ -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

changedetectionio/__init__.py

@@ -2,7 +2,7 @@
 
 # Read more https://github.com/dgtlmoon/changedetection.io/wiki
 # Semver means never use .01, or 00. Should be .1.
-__version__ = '0.55.3'
+__version__ = '0.55.6'
 
 from changedetectionio.strtobool import strtobool
 from json.decoder import JSONDecodeError

changedetectionio/api/Watch.py

@@ -1,4 +1,5 @@
 import os
+import re
 import threading
 
 from changedetectionio.validate_url import is_safe_valid_url
@@ -278,8 +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"

changedetectionio/blueprint/settings/__init__.py

@@ -10,12 +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.blueprint.settings.llm import construct_llm_blueprint
+    from changedetectionio.llm.evaluator import is_llm_features_disabled
     settings_blueprint = Blueprint('settings', __name__, template_folder="templates")
-    settings_blueprint.register_blueprint(construct_llm_blueprint(datastore), url_prefix='/llm')
+    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
@@ -30,24 +33,12 @@ def construct_blueprint(datastore: ChangeDetectionStore):
 
         default = deepcopy(datastore.data['settings'])
 
-        # Pre-populate LLM sub-form fields from stored config (text fields only —
-        # PasswordField for api_key is intentionally left blank on GET).
-        _stored_llm = datastore.data['settings']['application'].get('llm') or {}
-        default['llm'] = {
-            'llm_model':                         _stored_llm.get('model', ''),
-            'llm_api_base':                      _stored_llm.get('api_base', ''),
-            'llm_provider_kind':                 _stored_llm.get('provider_kind', ''),
-            'llm_local_token_multiplier':        _stored_llm.get('local_token_multiplier', 5),
-            'llm_change_summary_default':        datastore.data['settings']['application'].get('llm_change_summary_default', ''),
-            'llm_override_diff_with_summary':    datastore.data['settings']['application'].get('llm_override_diff_with_summary', True),
-            'llm_restock_use_fallback_extract':  datastore.data['settings']['application'].get('llm_restock_use_fallback_extract', True),
-            'llm_debug':                         datastore.data['settings']['application'].get('llm_debug', False),
-            'llm_budget_action':                 datastore.data['settings']['application'].get('llm_budget_action', 'skip_llm'),
-            'llm_thinking_budget':               str(datastore.data['settings']['application'].get('llm_thinking_budget', 0)),
-            'llm_max_summary_tokens':            str(datastore.data['settings']['application'].get('llm_max_summary_tokens', 3000)),
-            'llm_token_budget_month':            _stored_llm.get('token_budget_month', 0),
-            'llm_max_input_chars':               _stored_llm.get('max_input_chars', 0),
-        }
+        # 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())
@@ -98,79 +89,43 @@ def construct_blueprint(datastore: ChangeDetectionStore):
 
                 datastore.data['settings']['application'].update(app_update)
 
-                # Save LLM config separately under settings.application.llm.
-                # Token counters (tokens_total_cumulative, tokens_this_month, tokens_month_key)
-                # are system-managed and must never be overwritten by form submissions.
-                _LLM_PROTECTED_FIELDS = {
-                    'tokens_total_cumulative', 'tokens_this_month', 'tokens_month_key',
-                    'cost_usd_total_cumulative', 'cost_usd_this_month',
-                }
-                existing_llm = datastore.data['settings']['application'].get('llm') or {}
-                preserved_counters = {k: v for k, v in existing_llm.items() if k in _LLM_PROTECTED_FIELDS}
-
-                llm_data = form.data.get('llm') or {}
-
-                # PasswordField never re-populates its value on GET, so the submitted value
-                # is only non-empty when the user explicitly typed a new key.
-                # If blank, preserve the existing key so a settings save doesn't accidentally clear it.
-                submitted_api_key = (llm_data.get('llm_api_key') or '').strip()
-                effective_api_key = submitted_api_key if submitted_api_key else existing_llm.get('api_key', '')
-
-                # Application-level LLM settings (survive provider changes)
-                datastore.data['settings']['application']['llm_change_summary_default'] = (
-                    llm_data.get('llm_change_summary_default') or ''
-                ).strip()
-                datastore.data['settings']['application']['llm_override_diff_with_summary'] = (
-                    bool(llm_data.get('llm_override_diff_with_summary', True))
-                )
-                datastore.data['settings']['application']['llm_restock_use_fallback_extract'] = (
-                    bool(llm_data.get('llm_restock_use_fallback_extract', True))
-                )
-                datastore.data['settings']['application']['llm_debug'] = (
-                    bool(llm_data.get('llm_debug', False))
-                )
-                datastore.data['settings']['application']['llm_budget_action'] = (
-                    llm_data.get('llm_budget_action') or 'skip_llm'
-                )
-                datastore.data['settings']['application']['llm_thinking_budget'] = (
-                    int(llm_data.get('llm_thinking_budget') or 0)
-                )
-                datastore.data['settings']['application']['llm_max_summary_tokens'] = (
-                    int(llm_data.get('llm_max_summary_tokens') or 3000)
+                # 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 {}
                 )
 
-                # Monthly token budget — only save if env var is not set
-                import os as _os
-                if not _os.getenv('LLM_TOKEN_BUDGET_MONTH', '').strip():
-                    _budget = llm_data.get('llm_token_budget_month') or 0
-                    existing_llm['token_budget_month'] = int(_budget) if _budget else 0
+                llm_form_input = dict(form.data.get('llm') or {})
 
-                # Max input chars — only save if env var is not set
-                if not _os.getenv('LLM_MAX_INPUT_CHARS', '').strip():
-                    _max_chars = llm_data.get('llm_max_input_chars') or 0
-                    existing_llm['max_input_chars'] = int(_max_chars) if _max_chars else 0
+                # 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}
 
-                llm_config = {
-                    'model': (llm_data.get('llm_model') or '').strip(),
-                    'api_key': effective_api_key,
-                    'api_base': (llm_data.get('llm_api_base') or '').strip(),
-                    # Identifies a self-hosted OpenAI-compatible endpoint so reasoning-friendly
-                    # token caps can be applied conditionally (cloud-LLM defaults stay tight).
-                    'provider_kind': (llm_data.get('llm_provider_kind') or '').strip(),
-                    'local_token_multiplier': int(llm_data.get('llm_local_token_multiplier') or 5),
-                    'token_budget_month': existing_llm.get('token_budget_month', 0),
-                    'max_input_chars': existing_llm.get('max_input_chars', 0),
-                    **preserved_counters,
-                }
-                # Only store if a model is set
-                if llm_config['model']:
-                    datastore.data['settings']['application']['llm'] = llm_config
-                else:
-                    # Remove model config but retain counters for historical record
-                    if preserved_counters:
-                        datastore.data['settings']['application']['llm'] = preserved_counters
-                    else:
-                        datastore.data['settings']['application'].pop('llm', 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)

changedetectionio/blueprint/settings/llm.py

@@ -56,6 +56,7 @@ def construct_llm_blueprint(datastore: ChangeDetectionStore):
     @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()
@@ -66,10 +67,29 @@ def construct_llm_blueprint(datastore: ChangeDetectionStore):
             logger.debug("LLM model list: no provider specified, returning 400")
             return jsonify({'models': [], 'error': 'No provider specified'}), 400
 
-        # Fall back to the stored key if the user hasn't typed one yet
+        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:
-            api_key = (datastore.data['settings']['application'].get('llm') or {}).get('api_key', '')
-            logger.debug("LLM model list: no api_key in request, using stored 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/'}
@@ -115,15 +135,21 @@ def construct_llm_blueprint(datastore: ChangeDetectionStore):
     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':                 (request.args.get('api_key')                 or stored.get('api_key', '')).strip(),
-            'api_base':                (request.args.get('api_base')                or stored.get('api_base', '')).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'),
         }
@@ -140,6 +166,23 @@ def construct_llm_blueprint(datastore: ChangeDetectionStore):
             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
@@ -150,7 +193,7 @@ def construct_llm_blueprint(datastore: ChangeDetectionStore):
             # 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
+            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':
@@ -158,7 +201,7 @@ def construct_llm_blueprint(datastore: ChangeDetectionStore):
                 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=bool(datastore.data['settings']['application'].get('llm_debug', False)),
+                debug=get_llm_settings(datastore).debug,
             )
             reply = text.strip()
             if not reply:
@@ -181,16 +224,30 @@ def construct_llm_blueprint(datastore: ChangeDetectionStore):
             logger.exception("LLM connection test full traceback:")
             return jsonify({'ok': False, 'error': str(e)}), 400
 
-    @llm_blueprint.route("/clear", methods=['GET'])
+    # 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")
-        datastore.data['settings']['application'].pop('llm', None)
+        # 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=['GET'])
+    @llm_blueprint.route("/clear-summary-cache", methods=['POST'])
     @login_optionally_required
     def llm_clear_summary_cache():
         import glob

changedetectionio/blueprint/settings/templates/settings.html

@@ -34,7 +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>
@@ -394,7 +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>

changedetectionio/blueprint/settings/templates/settings_llm_tab.html

@@ -69,6 +69,17 @@
   {% 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>
@@ -114,22 +125,22 @@
   </div>
 
   <div class="pure-control-group">
-    {{ render_field(form.llm.form.llm_api_key) }}
+    {{ 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.llm_api_base) }}
+    {{ 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.llm_provider_kind() }}
+  {{ form.llm.form.provider_kind() }}
 
   <div class="pure-control-group" id="llm-local-advanced-group" style="display:none">
-    <label for="{{ form.llm.form.llm_local_token_multiplier.id }}">{{ form.llm.form.llm_local_token_multiplier.label.text }}</label>
-    {{ form.llm.form.llm_local_token_multiplier() }}
+    <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>
@@ -152,7 +163,7 @@
   </div>
 
   <div class="pure-control-group">
-    {{ render_field(form.llm.form.llm_model,
+    {{ render_field(form.llm.form.model,
                     placeholder=_("Enter API key and click 'Load available models'")) }}
   </div>
 
@@ -163,9 +174,14 @@
       &#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?') }}"
@@ -189,9 +205,11 @@
 
   <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?') }}"
@@ -205,9 +223,9 @@
 
   <div class="pure-control-group">
     <label></label>
-    {{ form.llm.form.llm_debug() }}
-    <label for="{{ form.llm.form.llm_debug.id }}" style="display:inline; font-weight:normal;">
-      {{ form.llm.form.llm_debug.label.text }}
+    {{ 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.') }}
@@ -225,10 +243,10 @@
   <p class="stab-section-title">{{ _('Default AI Change Summary') }}</p>
 
   <div class="pure-control-group">
-    {{ render_field(form.llm.form.llm_change_summary_default) }}
+    {{ 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-llm_change_summary_default'); if(!t.value && t.placeholder) t.value=t.placeholder; return false;">{{ _('Modify default prompt') }}</a>
+      &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>
 
@@ -241,9 +259,9 @@
   {% if llm_config and llm_config.get('model') %}
   <div class="pure-control-group">
     <label></label>
-    {{ form.llm.form.llm_override_diff_with_summary() }}
-    <label for="{{ form.llm.form.llm_override_diff_with_summary.id }}" style="display:inline; font-weight:normal;">
-      {{ form.llm.form.llm_override_diff_with_summary.label.text }}
+    {{ 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.',
@@ -253,9 +271,9 @@
 
   <div class="pure-control-group">
     <label></label>
-    {{ form.llm.form.llm_restock_use_fallback_extract() }}
-    <label for="{{ form.llm.form.llm_restock_use_fallback_extract.id }}" style="display:inline; font-weight:normal;">
-      {{ form.llm.form.llm_restock_use_fallback_extract.label.text }}
+    {{ 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.') }}
@@ -263,21 +281,21 @@
   </div>
 
   <div class="pure-control-group">
-    <label for="{{ form.llm.form.llm_thinking_budget.id }}">{{ form.llm.form.llm_thinking_budget.label.text }}</label>
-    {{ form.llm.form.llm_thinking_budget() }}
+    <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.llm_max_summary_tokens.id }}">{{ form.llm.form.llm_max_summary_tokens.label.text }}</label>
-    {{ form.llm.form.llm_max_summary_tokens() }}
+    <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.llm_budget_action.label.text }}</label>
+    <label>{{ form.llm.form.budget_action.label.text }}</label>
     <div>
-      {% for subfield in form.llm.form.llm_budget_action %}
+      {% 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>
@@ -330,9 +348,9 @@
         {% 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-llm_token_budget_month" value="{{ llm_token_budget_month_env }}">
+          <input type="hidden" name="llm-token_budget_month" value="{{ llm_token_budget_month_env }}">
         {% else %}
-          {{ form.llm.form.llm_token_budget_month(placeholder=_('0 = unlimited'), value=llm_stored.get('token_budget_month', 0) or '') }}
+          {{ 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>
@@ -347,14 +365,21 @@
       <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.llm_max_input_chars(value=llm_max_input_chars_env, readonly=True, style="width:10em;opacity:0.6;cursor:not-allowed;") }}
+          {{ 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.llm_max_input_chars(placeholder='100000', value=llm_stored.get('max_input_chars', 100000) or '') }}
+          {{ 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 %}
@@ -367,9 +392,9 @@
         {% 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-llm_token_budget_month" value="{{ llm_token_budget_month_env }}">
+          <input type="hidden" name="llm-token_budget_month" value="{{ llm_token_budget_month_env }}">
         {% else %}
-          {{ form.llm.form.llm_token_budget_month(placeholder=_('0 = unlimited'), value=llm_stored.get('token_budget_month', 0) or '') }}
+          {{ 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>
@@ -378,14 +403,21 @@
       <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.llm_max_input_chars(value=llm_max_input_chars_env, readonly=True, style="width:10em;opacity:0.6;cursor:not-allowed;") }}
+          {{ 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.llm_max_input_chars(placeholder='100000', value=llm_stored.get('max_input_chars', 100000) or '') }}
+          {{ 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 %}
@@ -416,8 +448,8 @@
     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-llm_api_base"]');
-    const kindField     = document.querySelector('[name="llm-llm_provider_kind"]');
+    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';
@@ -445,8 +477,8 @@
 
   window.llmFetchModels = async function () {
     const provider  = document.getElementById('llm-provider').value;
-    const apiKey    = document.querySelector('[name="llm-llm_api_key"]').value.trim();
-    const apiBase   = document.querySelector('[name="llm-llm_api_base"]').value.trim();
+    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');
@@ -481,7 +513,7 @@
       }
 
       modelSel.innerHTML = '<option value="">{{ _("— choose a model —") }}</option>';
-      const currentModel = document.querySelector('[name="llm-llm_model"]').value.trim();
+      const currentModel = document.querySelector('[name="llm-model"]').value.trim();
       for (const m of data.models) {
         const opt = document.createElement('option');
         opt.value = m;
@@ -503,7 +535,7 @@
   };
 
   window.llmOnModelPick = function (value) {
-    if (value) document.querySelector('[name="llm-llm_model"]').value = value;
+    if (value) document.querySelector('[name="llm-model"]').value = value;
   };
 
   window.llmRunTest = async function () {
@@ -519,11 +551,11 @@
     // 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-llm_model"]')    || {}).value || '';
-    const apiKey  = (document.querySelector('[name="llm-llm_api_key"]')  || {}).value || '';
-    const apiBase = (document.querySelector('[name="llm-llm_api_base"]') || {}).value || '';
-    const kind    = (document.querySelector('[name="llm-llm_provider_kind"]') || {}).value || '';
-    const mult    = (document.querySelector('[name="llm-llm_local_token_multiplier"]') || {}).value || '';
+    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());
@@ -553,7 +585,7 @@
 
   // On page load: detect and pre-select provider from current model
   (function detectCurrentProvider() {
-    const modelField = document.querySelector('[name="llm-llm_model"]');
+    const modelField = document.querySelector('[name="llm-model"]');
     if (!modelField) return;
     const m = modelField.value.trim();
     if (!m) return;
@@ -564,7 +596,7 @@
     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-llm_api_base"]');
+      const baseField = document.querySelector('[name="llm-api_base"]');
       guessed = (baseField && baseField.value.trim()) ? 'openai_compatible' : 'openai';
     }
     else if (m.startsWith('claude'))   guessed = 'anthropic';

changedetectionio/blueprint/ui/diff.py

@@ -272,8 +272,10 @@ def construct_blueprint(datastore: ChangeDetectionStore):
 
         # Diff-pref flags + system prompt + active model are part of the cache key
         # so prompt or model changes bust the cache.
-        _max_summary_tokens = datastore.data['settings']['application'].get('llm_max_summary_tokens', 3000)
-        _llm_model = (datastore.data['settings']['application'].get('llm') or {}).get('model', '')
+        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,

changedetectionio/blueprint/ui/templates/edit.html

@@ -57,7 +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>
@@ -321,9 +323,11 @@ 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>
@@ -503,7 +507,7 @@ Math: {{ 1 + 1 }}") }}
                             <td>{{ _('Server type reply') }}</td>
                             <td>{{ watch.get('remote_server_reply') }}</td>
                         </tr>
-                        {% if settings_application.get('llm', {}).get('model') %}
+                        {% 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>

changedetectionio/content_fetchers/requests.py

@@ -9,7 +9,7 @@ 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
+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!
@@ -87,10 +87,12 @@ class fetcher(Fetcher):
 
         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:
-                parsed_initial = urlparse(url)
-                if parsed_initial.hostname and is_private_hostname(parsed_initial.hostname):
-                    raise Exception(f"Fetch blocked: '{url}' resolves to a private/reserved IP address. "
+                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,
@@ -111,9 +113,9 @@ class fetcher(Fetcher):
                 location = r.headers.get('Location', '')
                 redirect_url = urljoin(current_url, location)
                 if not allow_iana_restricted:
-                    parsed_redirect = urlparse(redirect_url)
-                    if parsed_redirect.hostname and is_private_hostname(parsed_redirect.hostname):
-                        raise Exception(f"Redirect blocked: '{redirect_url}' resolves to a private/reserved IP address.")
+                    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,

changedetectionio/flask_app.py

@@ -522,6 +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():

changedetectionio/forms.py

@@ -584,6 +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
@@ -876,7 +887,6 @@ class processor_text_json_diff_form(commonSettingsForm):
 
     conditions_match_logic = RadioField(_l('Match'), choices=[('ALL', _l('Match all of the following')),('ANY', _l('Match any of the following'))], default='ALL')
     conditions = FieldList(FormField(ConditionFormRow), min_entries=1)  # Add rule logic here
-    # dennis-ignore: W303 - False positive caused by <title>. https://github.com/mozilla/dennis/issues/213
     use_page_title_in_list = TernaryNoneBooleanField(_l('Use page <title> in list'), default=None)
 
     history_snapshot_max_length = IntegerField(_l('Number of history items per watch to keep'), render_kw={"style": "width: 5em;"}, validators=[validators.Optional(), validators.NumberRange(min=2)])
@@ -1025,7 +1035,6 @@ class globalSettingsApplicationUIForm(Form):
     open_diff_in_new_tab = BooleanField(_l("Open 'History' page in a new tab"), default=True, validators=[validators.Optional()])
     socket_io_enabled = BooleanField(_l('Realtime UI Updates Enabled'), default=True, validators=[validators.Optional()])
     favicons_enabled = BooleanField(_l('Favicons Enabled'), default=True, validators=[validators.Optional()])
-    # dennis-ignore: W303 - False positive caused by <title>. https://github.com/mozilla/dennis/issues/213
     use_page_title_in_list = BooleanField(_l('Use page <title> in watch overview list')) #BooleanField=True
 
 # datastore.data['settings']['application']..
@@ -1097,12 +1106,12 @@ class globalSettingsLLMForm(Form):
       gemini/gemini-2.0-flash               → Google Gemini
       azure/gpt-4o                          → Azure OpenAI
     """
-    llm_model = StringField(
+    model = StringField(
         _l('Model'),
         validators=[validators.Optional()],
         render_kw={"placeholder": "gpt-4o-mini", "style": "width: 24em;"},
     )
-    llm_api_key = PasswordField(
+    api_key = PasswordField(
         _l('API Key'),
         validators=[validators.Optional()],
         render_kw={
@@ -1110,9 +1119,9 @@ class globalSettingsLLMForm(Form):
             "style": "width: 24em;",
         },
     )
-    llm_api_base = StringField(
+    api_base = StringField(
         _l('API Base URL'),
-        validators=[validators.Optional()],
+        validators=[validators.Optional(), validateLLMApiBaseSafe()],
         render_kw={
             "placeholder": "http://localhost:11434  (Ollama / custom endpoints only)",
             "style": "width: 24em;",
@@ -1121,7 +1130,7 @@ class globalSettingsLLMForm(Form):
     # 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.
-    llm_provider_kind = HiddenField(
+    provider_kind = HiddenField(
         validators=[validators.Optional()],
         default='',
     )
@@ -1133,13 +1142,13 @@ class globalSettingsLLMForm(Form):
     # 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.
-    llm_local_token_multiplier = IntegerField(
+    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;"},
     )
-    llm_change_summary_default = TextAreaField(
+    change_summary_default = TextAreaField(
         _l('Default AI Change Summary prompt'),
         validators=[validators.Optional(), validators.Length(max=2000)],
         render_kw={
@@ -1149,8 +1158,8 @@ class globalSettingsLLMForm(Form):
         },
         default='',
     )
-    llm_max_tokens_per_check = IntegerField(
-        _l('Max tokens per check'),
+    max_tokens_per_count_period = IntegerField(
+        _l('Max tokens per watch per period'),
         validators=[validators.Optional(), validators.NumberRange(min=0)],
         default=0,
         render_kw={
@@ -1158,22 +1167,13 @@ class globalSettingsLLMForm(Form):
             "style": "width: 8em;",
         },
     )
-    llm_max_tokens_cumulative = IntegerField(
-        _l('Max cumulative tokens (per watch)'),
-        validators=[validators.Optional(), validators.NumberRange(min=0)],
-        default=0,
-        render_kw={
-            "placeholder": "0 = unlimited",
-            "style": "width: 8em;",
-        },
-    )
-    llm_token_budget_month = IntegerField(
+    token_budget_month = IntegerField(
         _l('Monthly token budget'),
         validators=[validators.Optional(), validators.NumberRange(min=0)],
         default=0,
         render_kw={"style": "width: 10em;"},
     )
-    llm_max_input_chars = IntegerField(
+    max_input_chars = IntegerField(
         _l('Max input characters'),
         validators=[validators.Optional(), validators.NumberRange(min=1)],
         default=100000,
@@ -1182,19 +1182,27 @@ class globalSettingsLLMForm(Form):
             "style": "width: 10em;",
         },
     )
-    llm_override_diff_with_summary = BooleanField(
+    # 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,
     )
-    llm_restock_use_fallback_extract = BooleanField(
+    restock_use_fallback_extract = BooleanField(
         _l('Use LLM as a fallback for extracting price and restock info'),
         default=True,
     )
-    llm_debug = BooleanField(
+    debug = BooleanField(
         _l('Enable LLM debug logging'),
         default=False,
     )
-    llm_thinking_budget = SelectField(
+    thinking_budget = SelectField(
         _l('AI thinking budget (tokens)'),
         choices=[
             ('0',    _l('Off (no thinking)')),
@@ -1205,7 +1213,7 @@ class globalSettingsLLMForm(Form):
         default=str(LLM_DEFAULT_THINKING_BUDGET),
         validators=[validators.Optional()],
     )
-    llm_max_summary_tokens = SelectField(
+    max_summary_tokens = SelectField(
         _l('Max AI summary length (tokens)'),
         choices=[
             ('500',   '500'),
@@ -1218,7 +1226,7 @@ class globalSettingsLLMForm(Form):
         default=str(LLM_DEFAULT_MAX_SUMMARY_TOKENS),
         validators=[validators.Optional()],
     )
-    llm_budget_action = RadioField(
+    budget_action = RadioField(
         _l('When monthly token budget is reached'),
         choices=[
             ('skip_llm',   _l('Skip AI summarisation only (watch still checks)')),

changedetectionio/llm/evaluator.py

@@ -20,6 +20,8 @@ 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,
@@ -29,7 +31,29 @@ from .prompt_builder import (
 )
 from .response_parser import parse_eval_response, parse_preview_response, parse_setup_response
 
-_DEFAULT_MAX_INPUT_CHARS = 100_000
+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.
@@ -38,10 +62,9 @@ def _get_max_input_chars(datastore) -> int:
     env_val = os.getenv('LLM_MAX_INPUT_CHARS', '').strip()
     if env_val.isdigit() and int(env_val) > 0:
         return int(env_val)
-    cfg = datastore.data.get('settings', {}).get('application', {}).get('llm') or {}
-    stored = cfg.get('max_input_chars')
-    if stored and int(stored) > 0:
-        return int(stored)
+    stored = get_llm_settings(datastore).max_input_chars
+    if stored and stored > 0:
+        return stored
     return _DEFAULT_MAX_INPUT_CHARS
 
 
@@ -57,14 +80,25 @@ def _check_input_size(text: str, max_chars: int) -> None:
         )
 
 
-LLM_DEFAULT_THINKING_BUDGET = 0  # 0 = thinking disabled by default
-
 def _thinking_extra_body(model: str, budget: int) -> dict | None:
     """Return litellm extra_body to control thinking for models that support it.
-    For Gemini 2.5+: passes thinkingConfig with the given budget (0 = disabled).
-    For all other models: returns None (no-op).
+
+    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/gemini-2.5'):
+    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}}}
 
@@ -80,8 +114,6 @@ def _cached_system(text: str, model: str = '') -> dict:
     return {'role': 'system', 'content': text}
 
 
-LLM_DEFAULT_MAX_SUMMARY_TOKENS = 3000
-
 # 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.
@@ -207,6 +239,8 @@ def get_llm_config(datastore) -> dict | None:
       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:
@@ -225,9 +259,33 @@ def get_llm_config(datastore) -> dict | None:
 
 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
 # ---------------------------------------------------------------------------
@@ -298,25 +356,22 @@ def accumulate_global_tokens(datastore, tokens: int,
 
     current_month = _get_month_key()
     cost = _estimate_cost_usd(model, input_tokens, output_tokens)
-
-    # Work on the live dict in-place (or create a stub if llm key is absent)
-    app_settings = datastore.data['settings']['application']
-    if 'llm' not in app_settings:
-        app_settings['llm'] = {}
-    llm_cfg = app_settings['llm']
+    settings = get_llm_settings(datastore)
 
     # Month rollover: reset monthly counters
-    if llm_cfg.get('tokens_month_key') != current_month:
-        llm_cfg['tokens_this_month'] = 0
-        llm_cfg['cost_usd_this_month'] = 0.0
-        llm_cfg['tokens_month_key'] = current_month
+    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
 
-    llm_cfg['tokens_total_cumulative'] = (llm_cfg.get('tokens_total_cumulative') or 0) + tokens
-    llm_cfg['tokens_this_month']       = (llm_cfg.get('tokens_this_month') or 0) + tokens
-    llm_cfg['cost_usd_total_cumulative'] = (llm_cfg.get('cost_usd_total_cumulative') or 0.0) + cost
-    llm_cfg['cost_usd_this_month']       = (llm_cfg.get('cost_usd_this_month') or 0.0) + cost
+    settings.tokens_total_cumulative += tokens
+    settings.tokens_this_month       += tokens
+    settings.cost_usd_total_cumulative += cost
+    settings.cost_usd_this_month       += cost
 
-    # Persist immediately — token accounting must survive restarts
+    # 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()
 
 
@@ -344,31 +399,44 @@ def is_global_token_budget_exceeded(datastore) -> bool:
 
 def _check_token_budget(watch, cfg, tokens_this_call: int = 0) -> bool:
     """
-    Check token budget limits.  Returns True if within budget, False if exceeded.
-    Also accumulates tokens_this_call into watch['llm_tokens_used_cumulative'].
+    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 = watch.get('llm_tokens_used_cumulative') or 0
-        watch['llm_tokens_used_cumulative'] = current + tokens_this_call
+        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_check = int(cfg.get('max_tokens_per_check') or 0)
-    max_cumulative = int(cfg.get('max_tokens_cumulative') or 0)
-
-    if max_per_check and tokens_this_call > max_per_check:
-        logger.warning(
-            f"LLM token budget exceeded for {watch.get('uuid')}: "
-            f"{tokens_this_call} tokens > per-check limit {max_per_check}"
-        )
-        return False
-
-    if max_cumulative:
-        total = watch.get('llm_tokens_used_cumulative') or 0
-        if total > max_cumulative:
-            logger.warning(
-                f"LLM cumulative token budget exceeded for {watch.get('uuid')}: "
-                f"{total} tokens > limit {max_cumulative}"
-            )
-            return False
+    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
 
@@ -379,7 +447,7 @@ def run_setup(watch, datastore, snapshot_text: str) -> None:
     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 = get_llm_config(datastore)
+    cfg = _runtime_llm_config(datastore)
     if not cfg:
         return
 
@@ -390,6 +458,7 @@ def run_setup(watch, datastore, snapshot_text: str) -> None:
     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(
@@ -401,8 +470,8 @@ def run_setup(watch, datastore, snapshot_text: str) -> None:
             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'], int(datastore.data['settings']['application'].get('llm_thinking_budget', LLM_DEFAULT_THINKING_BUDGET) or 0)),
-            debug=bool(datastore.data['settings']['application'].get('llm_debug', False)),
+            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'])
@@ -426,11 +495,7 @@ def get_effective_summary_prompt(watch, datastore) -> str:
     prompt, _ = resolve_llm_field(watch, datastore, 'llm_change_summary')
     if prompt:
         return prompt
-    global_default = (
-        datastore.data.get('settings', {})
-        .get('application', {})
-        .get('llm_change_summary_default', '') or ''
-    ).strip()
+    global_default = get_llm_settings(datastore).change_summary_default.strip()
     return global_default or DEFAULT_CHANGE_SUMMARY_PROMPT
 
 
@@ -509,7 +574,7 @@ def summarise_change(watch, datastore, diff: str, current_snapshot: str = '') ->
     The result replaces {{ diff }} in notifications so the user gets a
     readable description instead of raw +/- diff lines.
     """
-    cfg = get_llm_config(datastore)
+    cfg = _runtime_llm_config(datastore)
     if not cfg:
         return ''
 
@@ -540,8 +605,8 @@ def summarise_change(watch, datastore, diff: str, current_snapshot: str = '') ->
         title=title,
     )
 
-    _thinking_budget = int(datastore.data['settings']['application'].get('llm_thinking_budget', LLM_DEFAULT_THINKING_BUDGET) or 0)
-    _extra_body = _thinking_extra_body(cfg['model'], _thinking_budget)
+    settings = get_llm_settings(datastore)
+    _extra_body = _thinking_extra_body(cfg['model'], settings.thinking_budget)
 
     try:
         _resp = llm_client.completion(
@@ -553,14 +618,11 @@ def summarise_change(watch, datastore, diff: str, current_snapshot: str = '') ->
             api_key=cfg.get('api_key'),
             api_base=cfg.get('api_base'),
             max_tokens=apply_local_token_multiplier(
-                _summary_max_tokens(
-                    diff,
-                    max_cap=int(datastore.data['settings']['application'].get('llm_max_summary_tokens', LLM_DEFAULT_MAX_SUMMARY_TOKENS) or LLM_DEFAULT_MAX_SUMMARY_TOKENS),
-                ),
+                _summary_max_tokens(diff, max_cap=settings.max_summary_tokens),
                 cfg,
             ),
             extra_body=_extra_body,
-            debug=bool(datastore.data['settings']['application'].get('llm_debug', False)),
+            debug=settings.debug,
         )
         raw, tokens = _resp[0], _resp[1]
         input_tokens  = _resp[2] if len(_resp) > 2 else 0
@@ -597,7 +659,7 @@ def preview_extract(watch, datastore, content: str) -> dict | None:
 
     Returns {'found': bool, 'answer': str} or None if LLM not configured / no intent.
     """
-    cfg = get_llm_config(datastore)
+    cfg = _runtime_llm_config(datastore)
     if not cfg:
         return None
 
@@ -611,6 +673,7 @@ def preview_extract(watch, datastore, content: str) -> dict | None:
 
     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(
@@ -622,8 +685,8 @@ def preview_extract(watch, datastore, content: str) -> dict | None:
             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'], int(datastore.data['settings']['application'].get('llm_thinking_budget', LLM_DEFAULT_THINKING_BUDGET) or 0)),
-            debug=bool(datastore.data['settings']['application'].get('llm_debug', False)),
+            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)
@@ -648,7 +711,7 @@ def evaluate_change(watch, datastore, diff: str, current_snapshot: str = '') ->
 
     Results are cached by (intent, diff) hash — each unique diff is evaluated exactly once.
     """
-    cfg = get_llm_config(datastore)
+    cfg = _runtime_llm_config(datastore)
     if not cfg:
         return None
 
@@ -697,6 +760,7 @@ def evaluate_change(watch, datastore, diff: str, current_snapshot: str = '') ->
         title=title,
     )
 
+    settings = get_llm_settings(datastore)
     try:
         _resp = llm_client.completion(
             model=cfg['model'],
@@ -707,8 +771,8 @@ def evaluate_change(watch, datastore, diff: str, current_snapshot: str = '') ->
             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'], int(datastore.data['settings']['application'].get('llm_thinking_budget', LLM_DEFAULT_THINKING_BUDGET) or 0)),
-            debug=bool(datastore.data['settings']['application'].get('llm_debug', False)),
+            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

changedetectionio/model/App.py

@@ -2,7 +2,6 @@ from os import getenv
 from copy import deepcopy
 
 from changedetectionio.blueprint.rss import RSS_FORMAT_TYPES, RSS_CONTENT_FORMAT_DEFAULT
-from changedetectionio.llm.evaluator import LLM_DEFAULT_MAX_SUMMARY_TOKENS, LLM_DEFAULT_THINKING_BUDGET
 from changedetectionio.model.Tags import TagsDict
 
 from changedetectionio.notification import (
@@ -71,8 +70,9 @@ class model(dict):
                     'shared_diff_access': False,
                     'strip_ignored_lines': False,
                     'tags': None,  # Initialized in __init__ with real datastore_path
-                    'llm_thinking_budget': LLM_DEFAULT_THINKING_BUDGET,
-                    'llm_max_summary_tokens': LLM_DEFAULT_MAX_SUMMARY_TOKENS,
+                    # 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,

changedetectionio/model/LLMSettings.py

@@ -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',
+    )

changedetectionio/model/PYDANTIC_MIGRATION.md

@@ -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

changedetectionio/notification/apprise_plugin/custom_handlers.py

@@ -60,7 +60,7 @@ 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
+from changedetectionio.validate_url import is_private_hostname, is_url_private_or_parser_confused
 
 SUPPORTED_HTTP_METHODS = {"get", "post", "put", "delete", "patch", "head"}
 
@@ -198,12 +198,14 @@ 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
+    # 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'):
-        hostname = urlparse(url).hostname or ''
-        if hostname and is_private_hostname(hostname):
+        if is_url_private_or_parser_confused(url):
             raise ValueError(
-                f"Notification target '{hostname}' is a private/reserved address. "
+                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."
             )
 

changedetectionio/notification/handler.py

@@ -364,6 +364,10 @@ 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,
         )
     )
 
@@ -372,7 +376,8 @@ def process_notification(n_object: NotificationContextData, datastore):
 
     # AI Change Summary: optionally replace {{ diff }} with the AI summary
     _llm_change_summary = (n_object.get('_llm_change_summary') or '').strip()
-    _override_diff = datastore.data['settings']['application'].get('llm_override_diff_with_summary', True)
+    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
 
@@ -394,10 +399,19 @@ def process_notification(n_object: NotificationContextData, datastore):
     # 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]:
-            if notification_parameters.get(key):
-                notification_parameters[key] = str(html_escape(str(notification_parameters[key])))
+            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']:

changedetectionio/notification_service.py

@@ -99,7 +99,7 @@ class FormattableExtract(str):
     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):
+    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
@@ -107,6 +107,12 @@ class FormattableExtract(str):
             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
 
@@ -128,16 +134,23 @@ class FormattableDiff(str):
 
     Being a str subclass means it is natively JSON serializable.
     """
-    def __new__(cls, prev_snapshot, current_snapshot, **base_kwargs):
+    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,
@@ -163,6 +176,10 @@ class FormattableDiff(str):
         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
 
 
@@ -187,6 +204,8 @@ class NotificationContextData(dict):
             '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,
@@ -236,7 +255,7 @@ class NotificationContextData(dict):
 
         super().__setitem__(key, value)
 
-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.
 
@@ -249,6 +268,9 @@ 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
@@ -287,10 +309,10 @@ def add_rendered_diff_to_notification_vars(notification_scan_text:str, prev_snap
         if not re.search(pattern, notification_scan_text, re.IGNORECASE):
             continue
         if key in diff_specs:
-            ret[key] = FormattableDiff(prev_snapshot, current_snapshot, **diff_specs[key])
+            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])
+            ret[key] = FormattableExtract(prev_snapshot, current_snapshot, extract_fn=extract_specs[key], escape_output=escape_output)
             rendered_count += 1
 
     if rendered_count:

changedetectionio/processors/base.py

@@ -5,7 +5,7 @@ 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
+from changedetectionio.validate_url import is_private_hostname, is_url_private_or_parser_confused
 from copy import deepcopy
 from abc import abstractmethod
 import os
@@ -104,13 +104,13 @@ class difference_detection_processor():
         """
         if strtobool(os.getenv('ALLOW_IANA_RESTRICTED_ADDRESSES', 'false')):
             return
-        parsed = urlparse(self.watch.link)
-        if not parsed.hostname:
-            return
         loop = asyncio.get_running_loop()
-        if await loop.run_in_executor(None, is_private_hostname, parsed.hostname):
+        # 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"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."
             )
 

changedetectionio/processors/restock_diff/plugins/llm_restock.py

@@ -196,22 +196,23 @@ def get_itemprop_availability_override(content, fetcher_name, fetcher_instance,
         logger.debug("LLM restock fallback: no datastore injected yet, skipping")
         return None
 
-    # Gate on the user setting (default True — enabled out of the box)
-    app_settings = datastore.data.get('settings', {}).get('application', {})
-    if not app_settings.get('llm_restock_use_fallback_extract', True):
-        logger.debug("LLM restock fallback: disabled in settings")
-        return None
-
     try:
-        from changedetectionio.llm.evaluator import get_llm_config, accumulate_global_tokens
+        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
 
-    llm_cfg = get_llm_config(datastore)
+    # 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, skipping")
+        logger.debug("LLM restock fallback: no LLM model configured or LLM disabled, skipping")
         return None
 
     text_content = _strip_html(content) if content else ''

changedetectionio/processors/text_json_diff/__init__.py

@@ -35,6 +35,50 @@ 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/<uuid_str:uuid>/preview-rendered", methods=['POST'])'''
     from changedetectionio import forms, html_tools
@@ -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 = []
@@ -89,15 +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
 
@@ -111,10 +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)}"
 

changedetectionio/processors/text_json_diff/difference.py

@@ -217,8 +217,10 @@ def render(watch, datastore, request, url_for, render_template, flash, redirect,
             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.
-            _max_summary_tokens = datastore.data['settings']['application'].get('llm_max_summary_tokens', 3000)
-            _llm_model = (datastore.data['settings']['application'].get('llm') or {}).get('model', '')
+            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,

changedetectionio/static/js/modal.js

@@ -187,6 +187,30 @@ $(document).ready(function() {
       confirmText: $element.attr('data-confirm-button') || 'Confirm',
       cancelText: $element.attr('data-cancel-button') || 'Cancel',
       onConfirm: function() {
+        // data-method="POST" — build a body-level hidden form with the CSRF
+        // token and submit it. Avoids nested-form HTML invalidity when the
+        // anchor lives inside an outer <form> (e.g. settings tabs). The CSRF
+        // token comes from the global `csrftoken` set in base.html.
+        // GHSA-g36r-fm2p-87xm: anchors that mutate server state must not fire
+        // on a bare GET, since <img src=...> CSRF relies on GET firing.
+        const method = ($element.attr('data-method') || 'GET').toUpperCase();
+        if (method === 'POST') {
+          const form = document.createElement('form');
+          form.method = 'POST';
+          form.action = url;
+          form.style.display = 'none';
+          if (typeof csrftoken !== 'undefined' && csrftoken) {
+            const tok = document.createElement('input');
+            tok.type = 'hidden';
+            tok.name = 'csrf_token';
+            tok.value = csrftoken;
+            form.appendChild(tok);
+          }
+          document.body.appendChild(form);
+          form.submit();
+          return;
+        }
+
         // If it's a link, navigate to the URL
         if ($element.is('a')) {
           window.location.href = url;

changedetectionio/static/js/watch-settings.js

@@ -9,6 +9,10 @@ function request_textpreview_update() {
     $('textarea:visible, input:visible').each(function () {
         const $element = $(this); // Cache the jQuery object for the current element
         const name = $element.attr('name'); // Get the name attribute of the element
+        // Radios share a name across multiple inputs; .val() returns the value
+        // attribute regardless of checked state, so iterating would let the last
+        // unchecked radio overwrite the user's actual selection. Skip unchecked.
+        if ($element.is(':radio') && !$element.is(':checked')) return;
         data[name] = $element.is(':checkbox') ? ($element.is(':checked') ? $element.val() : false) : $element.val();
     });
 

changedetectionio/store/updates.py

@@ -775,3 +775,53 @@ class DatastoreUpdatesMixin:
             tag.commit()
             logger.info(f"update_30: migrated tag {tag_uuid} restock_settings → processor_config_restock_diff")
 
+    def update_31(self):
+        """Fold any flat application.llm_* key into nested application.llm.<stripped>.
+
+        Before: a handful of LLM settings (llm_enabled, llm_thinking_budget, …) lived
+        directly on settings.application alongside everything else, while the provider
+        config (model, api_key, …) was already nested under settings.application.llm.
+        Unifies them under one parent so the LLMSettings pydantic model has a single
+        home to read/write.
+
+        Flat key wins on conflict (most-recent form-saved value). Idempotent.
+        """
+        application = self.data['settings']['application']
+        present = [k for k in list(application) if k.startswith('llm_')]
+        if not present:
+            return
+
+        nested = application.get('llm') or {}
+        for flat in present:
+            nested[flat.removeprefix('llm_')] = application.pop(flat)
+        application['llm'] = nested
+        logger.info(f"update_31: folded {len(present)} flat llm_* keys into application.llm.* "
+                    f"({', '.join(present)})")
+
+    def update_32(self):
+        """Drop max_tokens_per_check and rename max_tokens_cumulative → max_tokens_per_count_period.
+
+        max_tokens_per_check was never reachable from the UI (form field declared but
+        never rendered or saved) and overlapped with the cumulative cap. Removing it.
+
+        max_tokens_cumulative was misleading — the field was used as a per-watch
+        per-period cap, not lifetime. Renamed so the semantic is clear and so a
+        future configurable period (day/week/month) doesn't force another rename.
+
+        Both keys are unreached from real installs (no UI path on prior releases);
+        this migration is mostly for branches and devs running pre-release commits.
+        """
+        llm = self.data['settings']['application'].get('llm') or {}
+        if not llm:
+            return
+        changed = False
+        if 'max_tokens_per_check' in llm:
+            del llm['max_tokens_per_check']
+            changed = True
+        if 'max_tokens_cumulative' in llm:
+            llm.setdefault('max_tokens_per_count_period', llm.pop('max_tokens_cumulative'))
+            changed = True
+        if changed:
+            self.data['settings']['application']['llm'] = llm
+            logger.info("update_32: cleaned up obsolete max_tokens_per_check / renamed max_tokens_cumulative")
+

changedetectionio/templates/_common_fields.html

@@ -34,7 +34,6 @@
             </tr>
             <tr>
                 <td><code>{{ '{{watch_title}}' }}</code></td>
-                {# TRANSLATORS: dennis-ignore: W303 - False positive caused by <title>. https://github.com/mozilla/dennis/issues/213 #}
                 <td>{{ _('The page title of the watch, uses <title> if not set, falls back to URL') }}</td>
             </tr>
             <tr>
@@ -112,7 +111,7 @@
                 <td><code>{{ '{{triggered_text}}' }}</code></td>
                 <td>{{ _('Text that tripped the trigger from filters') }}</td>
             </tr>
-            {% if settings_application and settings_application.get('llm', {}).get('model') %}
+            {% if not llm_features_disabled and settings_application and settings_application.get('llm', {}).get('model') %}
             <tr>
                 <td><code>{{ '{{diff}}' }}</code> <small style="opacity:0.6">{{ _('(upgraded)') }}</small></td>
                 <td>{{ _('When AI Change Summary is configured, contains the AI-generated description instead of the raw diff. Falls back to raw diff when not configured.') }}</td>

changedetectionio/templates/base.html

@@ -281,6 +281,7 @@
       </div>
     </dialog>
 
+    {% if not llm_features_disabled %}
     <!-- LLM Not Configured Modal -->
     <dialog id="llm-not-configured-modal" class="modal-dialog" aria-labelledby="llm-not-configured-modal-title">
       <div class="modal-header">
@@ -294,6 +295,7 @@
         <button type="button" class="pure-button" id="close-llm-not-configured-modal">{{ _('Close') }}</button>
       </div>
     </dialog>
+    {% endif %}
 
     <!-- Search Modal -->
     {% if current_user.is_authenticated or not has_password %}

changedetectionio/templates/menu.html

@@ -37,10 +37,12 @@
     </li>
 {% endif %}
     <li class="pure-menu-item menu-collapsible" id="inline-menu-extras-group">
+        {% if not llm_features_disabled %}
         <button class="toggle-button toggle-ai-mode" type="button" title="{{ _('Toggle AI Mode') }}" data-llm-configured="{{ 'true' if llm_configured else 'false' }}" data-llm-settings-url="{{ url_for('settings.settings_page') }}#ai">
             <span class="visually-hidden">{{ _('Toggle AI mode') }}</span>
             {% include "svgs/ai-mode-icon.svg" %}<span class="ai-mode-label">LLM</span>
         </button>
+        {% endif %}
         <button class="toggle-button toggle-light-mode " type="button" title="{{ _('Toggle Light/Dark Mode') }}">
             <span class="visually-hidden">{{ _('Toggle light/dark mode') }}</span>
             <span class="icon-light">

changedetectionio/tests/llm/test_evaluator.py

@@ -294,78 +294,82 @@ class TestTokenBudget:
 
         assert _check_token_budget(watch, cfg, tokens_this_call=10_000) is True
 
-    def test_per_check_limit_exceeded_returns_false(self):
-        """Tokens on this call exceeding per-check limit → False."""
-        from changedetectionio.llm.evaluator import _check_token_budget
+    def test_per_period_limit_exceeded_returns_false(self):
+        """Per-period tokens exceeding the cap → False."""
+        from changedetectionio.llm.evaluator import _check_token_budget, _get_month_key
 
         watch = _make_watch()
-        cfg = {'max_tokens_per_check': 100}
-
-        result = _check_token_budget(watch, cfg, tokens_this_call=150)
-        assert result is False
-
-    def test_per_check_limit_not_exceeded_returns_true(self):
-        """Tokens on this call within per-check limit → True."""
-        from changedetectionio.llm.evaluator import _check_token_budget
-
-        watch = _make_watch()
-        cfg = {'max_tokens_per_check': 200}
-
-        result = _check_token_budget(watch, cfg, tokens_this_call=150)
-        assert result is True
-
-    def test_cumulative_limit_exceeded_returns_false(self):
-        """Total accumulated tokens exceeding cumulative limit → False."""
-        from changedetectionio.llm.evaluator import _check_token_budget
-
-        watch = _make_watch()
-        watch['llm_tokens_used_cumulative'] = 900
-        cfg = {'max_tokens_cumulative': 1000}
+        watch['llm_tokens_this_period'] = 900
+        watch['llm_tokens_period_key'] = _get_month_key()
+        cfg = {'max_tokens_per_count_period': 1000}
 
         # This call adds 200 → total 1100 > 1000
         result = _check_token_budget(watch, cfg, tokens_this_call=200)
         assert result is False
 
-    def test_cumulative_limit_not_yet_exceeded_returns_true(self):
-        """Total accumulated tokens within cumulative limit → True."""
-        from changedetectionio.llm.evaluator import _check_token_budget
+    def test_per_period_limit_not_yet_exceeded_returns_true(self):
+        """Per-period tokens within the cap → True."""
+        from changedetectionio.llm.evaluator import _check_token_budget, _get_month_key
 
         watch = _make_watch()
-        watch['llm_tokens_used_cumulative'] = 500
-        cfg = {'max_tokens_cumulative': 1000}
+        watch['llm_tokens_this_period'] = 500
+        watch['llm_tokens_period_key'] = _get_month_key()
+        cfg = {'max_tokens_per_count_period': 1000}
 
         result = _check_token_budget(watch, cfg, tokens_this_call=100)
         assert result is True
 
-    def test_tokens_accumulated_into_watch(self):
-        """tokens_this_call is added to watch['llm_tokens_used_cumulative']."""
-        from changedetectionio.llm.evaluator import _check_token_budget
+    def test_period_rollover_zeroes_counter(self):
+        """Stale period_key triggers rollover: counter resets before this call's tokens are added."""
+        from changedetectionio.llm.evaluator import _check_token_budget, _get_month_key
+
+        watch = _make_watch()
+        watch['llm_tokens_this_period'] = 999_999  # last period's giant total
+        watch['llm_tokens_period_key'] = '1970-01'  # ancient — guaranteed stale
+        cfg = {'max_tokens_per_count_period': 1000}
+
+        # This call adds 100 → after rollover should be 100, under the 1000 cap
+        result = _check_token_budget(watch, cfg, tokens_this_call=100)
+        assert result is True
+        assert watch['llm_tokens_this_period'] == 100
+        assert watch['llm_tokens_period_key'] == _get_month_key()
+
+    def test_tokens_accumulated_into_both_counters(self):
+        """tokens_this_call increments both the lifetime stat and the per-period counter."""
+        from changedetectionio.llm.evaluator import _check_token_budget, _get_month_key
 
         watch = _make_watch()
         watch['llm_tokens_used_cumulative'] = 300
+        watch['llm_tokens_this_period'] = 50
+        watch['llm_tokens_period_key'] = _get_month_key()
         cfg = {}
 
         _check_token_budget(watch, cfg, tokens_this_call=75)
         assert watch['llm_tokens_used_cumulative'] == 375
+        assert watch['llm_tokens_this_period'] == 125
 
-    def test_zero_tokens_call_does_not_change_cumulative(self):
-        """Calling with tokens_this_call=0 (pre-flight check) doesn't modify cumulative."""
-        from changedetectionio.llm.evaluator import _check_token_budget
+    def test_zero_tokens_call_does_not_change_counters(self):
+        """Calling with tokens_this_call=0 (pre-flight check) doesn't modify counters."""
+        from changedetectionio.llm.evaluator import _check_token_budget, _get_month_key
 
         watch = _make_watch()
         watch['llm_tokens_used_cumulative'] = 200
+        watch['llm_tokens_this_period'] = 80
+        watch['llm_tokens_period_key'] = _get_month_key()
         cfg = {}
 
         _check_token_budget(watch, cfg, tokens_this_call=0)
         assert watch['llm_tokens_used_cumulative'] == 200
+        assert watch['llm_tokens_this_period'] == 80
 
-    def test_evaluate_change_skips_call_when_cumulative_over_budget(self):
-        """Pre-flight cumulative check: if already over budget, skip LLM call and fail open."""
-        from changedetectionio.llm.evaluator import evaluate_change
+    def test_evaluate_change_skips_call_when_per_period_over_budget(self):
+        """Pre-flight check: if already over the period cap, skip the LLM call and fail open."""
+        from changedetectionio.llm.evaluator import evaluate_change, _get_month_key
 
-        ds = _make_datastore(llm_cfg={'model': 'gpt-4o-mini', 'max_tokens_cumulative': 100})
+        ds = _make_datastore(llm_cfg={'model': 'gpt-4o-mini', 'max_tokens_per_count_period': 100})
         watch = _make_watch(llm_intent='flag price drops')
-        watch['llm_tokens_used_cumulative'] = 500  # already far over
+        watch['llm_tokens_this_period'] = 500  # already far over
+        watch['llm_tokens_period_key'] = _get_month_key()
 
         with patch('changedetectionio.llm.client.completion') as mock_llm:
             result = evaluate_change(watch, ds, diff='- $500\n+ $400')
@@ -374,23 +378,6 @@ class TestTokenBudget:
         # Fail open: important=True so the notification is NOT suppressed
         assert result == {'important': True, 'summary': ''}
 
-    def test_evaluate_change_per_check_limit_fails_open(self):
-        """Per-check token exceeded after call → result still returned (fail open)."""
-        from changedetectionio.llm.evaluator import evaluate_change
-
-        # max_tokens_per_check is 50, but the call returns 150 tokens
-        ds = _make_datastore(llm_cfg={'model': 'gpt-4o-mini', 'max_tokens_per_check': 50})
-        watch = _make_watch(llm_intent='flag price drops')
-
-        llm_response = '{"important": false, "summary": "Only minor change"}'
-        with patch('changedetectionio.llm.client.completion', return_value=(llm_response, 150)):
-            result = evaluate_change(watch, ds, diff='- $500\n+ $499')
-
-        # LLM said not important, but even with per-check warning the result is returned
-        # (budget warning is logged but evaluation result is still used)
-        assert result is not None
-        assert 'important' in result
-
 
 # ---------------------------------------------------------------------------
 # resolve_llm_field (generic cascade)

changedetectionio/tests/llm/test_llm_features_disabled.py

@@ -0,0 +1,62 @@
+"""
+Smoke test for the LLM_FEATURES_DISABLED env var.
+
+The env var is intended to hide every LLM/AI surface (settings tab, edit tab,
+base-template AI toggle/modal) for hosted deployments. This test renders the
+three primary pages with the env var set and verifies that none of the
+LLM-related markers leak through.
+"""
+from flask import url_for
+
+
+def _llm_markers_absent(body: bytes, where: str = ''):
+    """All of these strings appear in LLM UI surfaces — none should render."""
+    for marker in (b'AI / LLM', b'toggle-ai-mode', b'llm-not-configured-modal',
+                   b'id="ai-llm"', b'#ai-llm', b'href="#ai"'):
+        if marker in body:
+            idx = body.find(marker)
+            context = body[max(0, idx - 80):idx + len(marker) + 80].decode('utf-8', 'replace')
+            raise AssertionError(f"[{where}] {marker!r} found in body, context: ...{context}...")
+
+
+def test_llm_features_disabled_hides_ui(client, live_server, monkeypatch):
+    monkeypatch.setenv('LLM_FEATURES_DISABLED', 'true')
+
+    # Sanity: helper reports the env var is in effect
+    from changedetectionio.llm.evaluator import is_llm_features_disabled, get_llm_config
+    assert is_llm_features_disabled() is True
+    # get_llm_config() must return None so every `if llm_configured` template hides
+    datastore = client.application.config.get('DATASTORE')
+    assert get_llm_config(datastore) is None
+
+    # 1. Watch list (base.html + menu.html surface)
+    res = client.get(url_for('watchlist.index'))
+    assert res.status_code == 200
+    _llm_markers_absent(res.data, where='watchlist')
+
+    # 2. Settings page (should not have an AI / LLM tab or the LLM tab body)
+    res = client.get(url_for('settings.settings_page'))
+    assert res.status_code == 200
+    _llm_markers_absent(res.data, where='settings')
+
+    # 3. Edit page for a watch (should not have an AI / LLM tab or include_llm_intent body)
+    uuid = datastore.add_watch(url='http://example.com', extras={'title': 'Disabled LLM watch'})
+    res = client.get(url_for('ui.ui_edit.edit_page', uuid=uuid))
+    assert res.status_code == 200
+    _llm_markers_absent(res.data, where='edit')
+    # The watch-edit-only intent textarea should also be absent
+    assert b'name="llm_intent"' not in res.data
+    assert b'name="llm_change_summary"' not in res.data
+
+
+def test_llm_features_enabled_by_default(client, live_server, monkeypatch):
+    """When LLM_FEATURES_DISABLED is unset, the AI / LLM surfaces are still rendered."""
+    monkeypatch.delenv('LLM_FEATURES_DISABLED', raising=False)
+
+    from changedetectionio.llm.evaluator import is_llm_features_disabled
+    assert is_llm_features_disabled() is False
+
+    res = client.get(url_for('settings.settings_page'))
+    assert res.status_code == 200
+    # The AI / LLM settings tab anchor should be present when not disabled
+    assert b'href="#ai"' in res.data

changedetectionio/tests/llm/test_llm_restock_plugin.py

@@ -14,8 +14,9 @@ def _make_datastore(llm_model='gpt-4o-mini', enabled=True):
     ds.data = {
         'settings': {
             'application': {
-                'llm_restock_use_fallback_extract': enabled,
                 'llm': {
+                    'enabled': True,
+                    'restock_use_fallback_extract': enabled,
                     'model': llm_model,
                     'api_key': 'test-key',
                     'api_base': '',
@@ -84,8 +85,8 @@ class TestLLMRestockPluginDisabled:
         ds.data = {
             'settings': {
                 'application': {
-                    'llm_restock_use_fallback_extract': True,
-                    # No 'llm' key → get_llm_config returns None
+                    # No 'llm' key → get_llm_config returns None;
+                    # restock_use_fallback_extract still defaults to True via LLMSettings
                 }
             }
         }

changedetectionio/tests/test_api_security.py

@@ -9,7 +9,7 @@ import json
 import threading
 import uuid as uuid_module
 from flask import url_for
-from .util import live_server_setup, wait_for_all_checks, delete_all_watches
+from .util import live_server_setup, wait_for_all_checks, wait_for_watch_history, delete_all_watches
 import os
 
 
@@ -653,6 +653,80 @@ def test_api_history_edge_cases(client, live_server, measure_memory_usage, datas
     delete_all_watches(client)
 
 
+def test_api_history_html_does_not_serve_as_text_html(client, live_server, measure_memory_usage, datastore_path):
+    """
+    GHSA-cgj8-g98g-4p9x: GET /api/v1/watch/<uuid>/history/<timestamp>?html=true
+    must not serve the stored snapshot with Content-Type: text/html. The bytes
+    are an external site's HTML — if the response is labelled text/html, a
+    <script> the attacker planted on that site executes in our origin when an
+    operator opens the URL in a browser (stored XSS).
+
+    The fix is text/plain; charset=utf-8 + X-Content-Type-Options: nosniff so
+    browsers render inert text and can't sniff back to HTML/UTF-7. API clients
+    don't care about Content-Type and still receive the same bytes.
+
+    This test injects the snapshot directly via Watch.save_history_blob() and
+    save_last_fetched_html() so we exercise the API endpoint's response
+    shaping without depending on the live-fetch pipeline.
+    """
+    api_key = live_server.app.config['DATASTORE'].data['settings']['application'].get('api_access_token')
+
+    test_url = url_for('test_endpoint', _external=True)
+    res = client.post(
+        url_for("createwatch"),
+        data=json.dumps({"url": test_url}),
+        headers={'content-type': 'application/json', 'x-api-key': api_key},
+    )
+    watch_uuid = res.json.get('uuid')
+
+    # Plant a payload that would execute if the response were rendered as HTML.
+    malicious_html = (
+        "<html><body>"
+        "<script>window.__CD_XSS_PROBE = 1</script>"
+        "<img src=x onerror=\"window.__CD_XSS_PROBE = 1\">"
+        "</body></html>"
+    )
+    ts = '1700000000'
+    watch = live_server.app.config['DATASTORE'].data['watching'][watch_uuid]
+    watch.save_history_blob(contents=malicious_html, timestamp=ts, snapshot_id=ts)
+    watch.save_last_fetched_html(timestamp=ts, contents=malicious_html)
+
+    # The actual XSS-relevant assertion: how is the snapshot served?
+    res = client.get(
+        url_for("watchsinglehistory", uuid=watch_uuid, timestamp=ts) + '?html=true',
+        headers={'x-api-key': api_key},
+    )
+    assert res.status_code == 200, f"unexpected status {res.status_code}: {res.data!r}"
+
+    ctype = res.headers.get('Content-Type', '')
+    assert 'text/html' not in ctype, \
+        f"snapshot must not be served as text/html (got {ctype!r}) — see GHSA-cgj8-g98g-4p9x"
+    # Explicit utf-8 closes the UTF-7 sniffing bypass — without a charset, some
+    # browsers will auto-detect UTF-7 from byte patterns and a crafted snapshot
+    # can still execute via `+ADw-script+AD4-...`
+    assert 'charset=utf-8' in ctype.lower(), \
+        f"Content-Type must pin charset=utf-8 to defeat UTF-7 sniffing XSS (got {ctype!r})"
+
+    nosniff = res.headers.get('X-Content-Type-Options', '')
+    assert nosniff.lower() == 'nosniff', \
+        f"X-Content-Type-Options: nosniff required to defeat MIME-sniffing (got {nosniff!r})"
+
+    # Download filename should include the timestamp so multiple snapshots from
+    # the same watch don't overwrite each other on disk.
+    disp = res.headers.get('Content-Disposition', '')
+    assert 'attachment' in disp and ts in disp, \
+        f"Content-Disposition should be attachment + per-timestamp filename (got {disp!r})"
+
+    # API contract: the raw bytes must still be the original HTML — programmatic
+    # consumers depend on getting the stored snapshot back.
+    assert b'<script>' in res.data, \
+        "Response body must still contain the raw stored bytes (the API contract)"
+
+    # Cleanup
+    client.delete(url_for("watch", uuid=watch_uuid), headers={'x-api-key': api_key})
+    delete_all_watches(client)
+
+
 def test_api_notification_edge_cases(client, live_server, measure_memory_usage, datastore_path):
     """
     Test notification configuration edge cases.

changedetectionio/tests/test_live_preview.py

@@ -77,3 +77,82 @@ def test_content_filter_live_preview(client, live_server, measure_memory_usage,
     assert reply.get('trigger_line_numbers') == [1]  # Triggers "Awesome" in line 1
 
     delete_all_watches(client)
+
+
+def _setup_version_list_preview(datastore_path, client):
+    """Shared HTML fixture for #4138 preview regressions (version tag list)."""
+    import time
+
+    data = """<html><body>
+0.55.5<br>
+0.55.4<br>
+0.55.3<br>
+0.54.10<br>
+0.54.9<br>
+</body></html>"""
+    with open(os.path.join(datastore_path, "endpoint-content.txt"), "w") as f:
+        f.write(data)
+
+    test_url = url_for('test_endpoint', _external=True)
+    uuid = client.application.config.get('DATASTORE').add_watch(url=test_url)
+    client.get(url_for("ui.form_watch_checknow"), follow_redirects=True)
+    time.sleep(0.5)
+    wait_for_all_checks(client)
+    return test_url, uuid
+
+
+def test_preview_ignore_highlight_with_extract_text(client, live_server, measure_memory_usage, datastore_path):
+    """Regression for #4138 follow-up: when extract_text rewrites a line (e.g. "0.54.10" → ".54.10"),
+    the preview must still highlight that row as 'ignored' even though substring matching against the
+    post-extract text fails."""
+    import json
+
+    test_url, uuid = _setup_version_list_preview(datastore_path, client)
+
+    res = client.post(
+        url_for("ui.ui_edit.watch_get_preview_rendered", uuid=uuid),
+        data={
+            "include_filters": "",
+            "fetch_backend": 'html_requests',
+            "ignore_text": "0.54.10",
+            "extract_text": r"/(.\d+\.\d+)/",
+            "url": test_url,
+        },
+    )
+    reply = json.loads(res.data.decode('utf-8'))
+    # The regex strips the leading "0", so the post-extract line for the ignored input is ".54.10".
+    # The preview should still mark its position (line 4) as ignored.
+    assert reply.get('ignore_line_numbers') == [4], \
+        f"Expected line 4 to be highlighted as ignored, got {reply.get('ignore_line_numbers')!r}"
+
+    delete_all_watches(client)
+
+
+def test_preview_strip_ignored_lines_with_extract_text(client, live_server, measure_memory_usage, datastore_path):
+    """Regression for #4138 follow-up: with strip_ignored_lines enabled, an ignored line must be
+    removed from the preview output even when extract_text would otherwise rewrite it (0.54.10 → .54.10)."""
+    import json
+
+    test_url, uuid = _setup_version_list_preview(datastore_path, client)
+
+    res = client.post(
+        url_for("ui.ui_edit.watch_get_preview_rendered", uuid=uuid),
+        data={
+            "include_filters": "",
+            "fetch_backend": 'html_requests',
+            "ignore_text": "0.54.10",
+            "extract_text": r"/(.\d+\.\d+)/",
+            "strip_ignored_lines": "true",
+            "url": test_url,
+        },
+    )
+    reply = json.loads(res.data.decode('utf-8'))
+    after_filter = reply.get('after_filter', '')
+
+    assert '.54.10' not in after_filter, \
+        f"Stripped ignored line should not appear in preview output, got:\n{after_filter!r}"
+    assert '0.54.10' not in after_filter
+    assert reply.get('ignore_line_numbers') == [], \
+        f"Stripped lines need no highlight, got {reply.get('ignore_line_numbers')!r}"
+
+    delete_all_watches(client)

changedetectionio/tests/test_llm_api_key_security.py

@@ -329,9 +329,9 @@ def test_settings_form_preserves_api_key_when_submitted_blank(
     res = client.post(
         url_for('settings.settings_page'),
         data={
-            'llm-llm_model': 'gpt-4o',
-            'llm-llm_api_key': '',           # blank — PasswordField behaviour
-            'llm-llm_api_base': '',
+            'llm-model': 'gpt-4o',
+            'llm-api_key': '',           # blank — PasswordField behaviour
+            'llm-api_base': '',
             'application-pager_size': '50',
             'application-notification_format': 'System default',
             'requests-time_between_check-days': '0',
@@ -351,3 +351,325 @@ def test_settings_form_preserves_api_key_when_submitted_blank(
         f"Blank PasswordField submission must not clear the existing API key (got '{saved_key}')"
 
     delete_all_watches(client)
+
+
+# ---------------------------------------------------------------------------
+# SSRF — api_base must reject private/loopback/reserved hosts (GHSA-jrxm-qjfh-g54f)
+# ---------------------------------------------------------------------------
+
+# Hosts that is_private_hostname() must classify as restricted.
+# 169.254.169.254 is the cloud metadata service (AWS/GCP IMDSv1).
+_SSRF_PRIVATE_HOSTS = [
+    'http://127.0.0.1:6379',
+    'http://localhost:11434',
+    'http://10.0.0.5:8080',
+    'http://192.168.1.1',
+    'http://169.254.169.254',
+]
+
+
+def test_llm_models_endpoint_blocks_private_api_base(
+        client, live_server, measure_memory_usage, datastore_path, monkeypatch):
+    """GET /settings/llm/models must refuse api_base pointing at private/loopback
+    hosts and must never reach litellm."""
+    # Default state — protection ON
+    monkeypatch.delenv('ALLOW_IANA_RESTRICTED_ADDRESSES', raising=False)
+
+    for bad in _SSRF_PRIVATE_HOSTS:
+        res = client.get(
+            url_for('settings.llm.llm_get_models'),
+            query_string={'provider': 'openai_compatible', 'api_base': bad},
+        )
+        assert res.status_code == 400, \
+            f"api_base={bad!r} should have been rejected by SSRF guard"
+        body = res.get_json()
+        assert body['models'] == []
+        assert 'ALLOW_IANA_RESTRICTED_ADDRESSES' in body['error'], \
+            f"Error message should mention the env-var bypass: {body['error']!r}"
+        # The raw attacker-controlled api_base must never be reflected back
+        # (avoids XSS when JS renders the error into the DOM).
+        assert bad not in body['error']
+
+
+def test_llm_test_endpoint_blocks_private_api_base(
+        client, live_server, measure_memory_usage, datastore_path, monkeypatch):
+    """GET /settings/llm/test must refuse api_base pointing at private/loopback
+    hosts and must never reach litellm.completion()."""
+    monkeypatch.delenv('ALLOW_IANA_RESTRICTED_ADDRESSES', raising=False)
+
+    for bad in _SSRF_PRIVATE_HOSTS:
+        res = client.get(
+            url_for('settings.llm.llm_test'),
+            query_string={'model': 'openai/gpt-4', 'api_base': bad},
+        )
+        assert res.status_code == 400, \
+            f"api_base={bad!r} should have been rejected by SSRF guard"
+        body = res.get_json()
+        assert body['ok'] is False
+        assert 'ALLOW_IANA_RESTRICTED_ADDRESSES' in body['error']
+        assert bad not in body['error']
+
+
+def test_llm_endpoints_allow_api_base_when_iana_bypass_enabled(
+        client, live_server, measure_memory_usage, datastore_path, monkeypatch):
+    """When ALLOW_IANA_RESTRICTED_ADDRESSES=true the SSRF guard is bypassed so
+    operators can intentionally point at a local Ollama / vLLM endpoint.
+    We patch litellm so the test doesn't actually need a live model server —
+    we only need to confirm the guard didn't short-circuit."""
+    monkeypatch.setenv('ALLOW_IANA_RESTRICTED_ADDRESSES', 'true')
+
+    # Stub get_valid_models so the call returns successfully without network.
+    import litellm
+    monkeypatch.setattr(litellm, 'get_valid_models',
+                        lambda **kwargs: ['llama3.2'])
+
+    # Supply api_key explicitly so we aren't tripped by the credential-exfil
+    # guard (which refuses to substitute the stored key for a non-stored api_base).
+    res = client.get(
+        url_for('settings.llm.llm_get_models'),
+        query_string={'provider': 'openai_compatible',
+                      'api_base': 'http://127.0.0.1:11434',
+                      'api_key': 'sk-test-explicit'},
+    )
+    assert res.status_code == 200, \
+        "With ALLOW_IANA_RESTRICTED_ADDRESSES=true, private api_base must be allowed"
+    body = res.get_json()
+    assert body['error'] is None
+    assert body['models'], "Stubbed model list should be returned"
+
+
+def test_settings_form_rejects_private_api_base(
+        client, live_server, measure_memory_usage, datastore_path, monkeypatch):
+    """The globalSettingsLLMForm validator must block private api_base values
+    when ALLOW_IANA_RESTRICTED_ADDRESSES is not set, and must NOT persist them
+    to the datastore."""
+    monkeypatch.delenv('ALLOW_IANA_RESTRICTED_ADDRESSES', raising=False)
+
+    ds = client.application.config.get('DATASTORE')
+    # Make sure no stale api_base exists from previous tests.
+    ds.data['settings']['application'].pop('llm', None)
+
+    res = client.post(
+        url_for('settings.settings_page'),
+        data={
+            'llm-model':    'gpt-4o',
+            'llm-api_key':  '',
+            'llm-api_base': 'http://127.0.0.1:11434',
+            'application-pager_size': '50',
+            'application-notification_format': 'System default',
+            'requests-time_between_check-days': '0',
+            'requests-time_between_check-hours': '0',
+            'requests-time_between_check-minutes': '5',
+            'requests-time_between_check-seconds': '0',
+            'requests-time_between_check-weeks': '0',
+            'requests-workers': '10',
+            'requests-timeout': '60',
+        },
+        follow_redirects=True,
+    )
+    # Form re-renders with the validation error — page itself returns 200.
+    assert res.status_code == 200
+    body = res.data.decode('utf-8', errors='replace')
+    assert 'ALLOW_IANA_RESTRICTED_ADDRESSES' in body, \
+        "Settings page should surface the SSRF guard's bypass-env-var hint"
+
+    saved = ds.data['settings']['application'].get('llm', {}).get('api_base', '')
+    assert saved != 'http://127.0.0.1:11434', \
+        f"Private api_base must not have been persisted (got {saved!r})"
+
+
+# ---------------------------------------------------------------------------
+# Credential exfiltration — stored api_key must NOT be auto-substituted when
+# the caller points api_base at a different (potentially attacker-controlled)
+# endpoint. GHSA-g36r-fm2p-87xm.
+# ---------------------------------------------------------------------------
+
+def test_llm_models_refuses_to_leak_stored_key_to_different_api_base(
+        client, live_server, measure_memory_usage, datastore_path, monkeypatch):
+    """If the request supplies an api_base that differs from the saved one but
+    omits api_key, the endpoint must refuse — otherwise CSRF can ship the
+    stored Authorization: Bearer <key> to an attacker-controlled URL."""
+    monkeypatch.delenv('ALLOW_IANA_RESTRICTED_ADDRESSES', raising=False)
+    ds = client.application.config.get('DATASTORE')
+    _configure_llm(ds)   # stores CANARY_KEY, leaves api_base unset
+
+    # Patch litellm.get_valid_models so that if the guard ever lets us through
+    # we'd see it called — and we can assert it wasn't.
+    import litellm
+    calls = []
+    monkeypatch.setattr(litellm, 'get_valid_models',
+                        lambda **kwargs: calls.append(kwargs) or [])
+
+    res = client.get(
+        url_for('settings.llm.llm_get_models'),
+        query_string={
+            'provider': 'openai',
+            'api_base': 'https://attacker.example/v1',
+            # api_key intentionally omitted — this is the CSRF case
+        },
+    )
+    assert res.status_code == 400, \
+        "Endpoint should refuse to substitute stored key to a mismatched api_base"
+    body = res.get_json()
+    assert 'api_key' in body['error'], \
+        f"Error should call out that api_key is required: {body['error']!r}"
+    assert calls == [], "litellm must not have been invoked at all"
+
+
+def test_llm_test_refuses_to_leak_stored_key_to_different_api_base(
+        client, live_server, measure_memory_usage, datastore_path, monkeypatch):
+    """Same guard on /settings/llm/test — attacker-supplied api_base + missing
+    api_key must not result in the stored key being sent to that URL."""
+    monkeypatch.delenv('ALLOW_IANA_RESTRICTED_ADDRESSES', raising=False)
+    ds = client.application.config.get('DATASTORE')
+    _configure_llm(ds)   # stores CANARY_KEY, no stored api_base
+
+    calls = []
+    # Patch the completion wrapper so we'd notice if litellm were invoked.
+    import changedetectionio.llm.client as llm_client
+    monkeypatch.setattr(llm_client, 'completion',
+                        lambda **kw: calls.append(kw) or ('', 0, 0, 0))
+
+    res = client.get(
+        url_for('settings.llm.llm_test'),
+        query_string={
+            'model': 'gpt-4o-mini',
+            'api_base': 'https://attacker.example/v1',
+            # api_key intentionally omitted
+        },
+    )
+    assert res.status_code == 400
+    body = res.get_json()
+    assert body['ok'] is False
+    assert 'api_key' in body['error']
+    assert calls == [], "completion() must not have been invoked"
+
+
+def test_llm_models_allows_stored_key_when_api_base_matches_saved(
+        client, live_server, measure_memory_usage, datastore_path, monkeypatch):
+    """Regression: the legit UI flow (test saved config without retyping the key)
+    must still work — i.e. when request api_base matches the stored api_base,
+    the stored key IS substituted."""
+    monkeypatch.delenv('ALLOW_IANA_RESTRICTED_ADDRESSES', raising=False)
+    monkeypatch.setenv('ALLOW_IANA_RESTRICTED_ADDRESSES', 'true')  # so localhost passes SSRF
+    ds = client.application.config.get('DATASTORE')
+    _configure_llm(ds)
+    ds.data['settings']['application']['llm']['api_base'] = 'http://localhost:11434'
+
+    received = []
+    import litellm
+    monkeypatch.setattr(litellm, 'get_valid_models',
+                        lambda **kwargs: (received.append(kwargs), ['llama3.2'])[1])
+
+    res = client.get(
+        url_for('settings.llm.llm_get_models'),
+        query_string={
+            'provider': 'openai_compatible',
+            'api_base': 'http://localhost:11434',  # matches saved
+            # api_key omitted — should fall back to stored CANARY_KEY
+        },
+    )
+    assert res.status_code == 200, res.get_json()
+    assert received and received[0].get('api_key') == CANARY_KEY, \
+        "When api_base matches saved, the stored api_key should be used"
+
+
+# ---------------------------------------------------------------------------
+# CSRF — /clear and /clear-summary-cache must not mutate state on GET
+# (GHSA-g36r-fm2p-87xm). The <img src=...> CSRF vector relies on GET firing the
+# mutation; the production guard is "POST only + Flask-WTF CSRF token". The
+# test config disables WTF_CSRF_ENABLED, so we verify the GET vector by
+# asserting the mutation didn't happen, and verify POST routing by exercising
+# the legit confirm-then-POST flow.
+#
+# NB: the app registers a catch-all '/<path:filename>' static route, which
+# intercepts any GET that isn't claimed by a method-matching rule and returns
+# 404 — so we can't simply assert on status code. The behaviour test below is
+# the actual security property.
+# ---------------------------------------------------------------------------
+
+def test_llm_clear_get_does_not_wipe_config(
+        client, live_server, measure_memory_usage, datastore_path):
+    """The CSRF surface is GET → mutation. After this fix the endpoint is
+    POST-only, so a GET must leave LLM config intact."""
+    ds = client.application.config.get('DATASTORE')
+    _configure_llm(ds)
+    assert ds.data['settings']['application'].get('llm', {}).get('api_key') == CANARY_KEY
+
+    client.get(url_for('settings.llm.llm_clear'))
+
+    # Mutation must not have happened — that's what defeats <img src=...> CSRF.
+    assert ds.data['settings']['application'].get('llm', {}).get('api_key') == CANARY_KEY, \
+        "GET /settings/llm/clear must not wipe LLM config (CSRF guard)"
+
+
+def test_llm_clear_summary_cache_get_does_not_wipe_cache(
+        client, live_server, measure_memory_usage, datastore_path):
+    """Same property for the cache wipe endpoint — GET must not delete the
+    change-summary-*.txt files the endpoint targets. To exercise the actual
+    deletion path we have to create a real watch (so a real data_dir exists)
+    and drop a real change-summary-*.txt inside it. POST should remove it;
+    GET must not."""
+    import os
+    ds = client.application.config.get('DATASTORE')
+    _configure_llm(ds)
+    api_token = _api_token(client)
+
+    # Create a real watch — required to exercise llm_clear_summary_cache's
+    # iteration over datastore.data['watching'].values().
+    test_url = url_for('test_endpoint', _external=True)
+    res = client.post(
+        '/api/v1/watch',
+        data=json.dumps({'url': test_url}),
+        headers={'content-type': 'application/json', 'x-api-key': api_token},
+        follow_redirects=True,
+    )
+    assert res.status_code == 201
+    uuid = res.json.get('uuid')
+
+    watch = ds.data['watching'][uuid]
+    data_dir = watch.data_dir
+    assert data_dir, "Watch must have a data_dir for this test to be meaningful"
+    os.makedirs(data_dir, exist_ok=True)
+
+    summary_file = os.path.join(data_dir, 'change-summary-csrf-canary.txt')
+    with open(summary_file, 'w') as f:
+        f.write('do-not-delete-via-GET')
+
+    # GET must NOT trigger the wipe — this is the CSRF surface that was open
+    # via <img src="/settings/llm/clear-summary-cache">.
+    client.get(url_for('settings.llm.llm_clear_summary_cache'))
+    assert os.path.exists(summary_file), \
+        "GET on /settings/llm/clear-summary-cache must not invoke the cache wipe"
+
+    # Sanity check: POST does remove it — confirms our test actually exercises
+    # the deletion path the GET test is guarding against.
+    client.post(url_for('settings.llm.llm_clear_summary_cache'))
+    assert not os.path.exists(summary_file), \
+        "POST on /settings/llm/clear-summary-cache should remove change-summary-*.txt"
+
+    delete_all_watches(client)
+
+
+def test_llm_clear_via_post_still_works(
+        client, live_server, measure_memory_usage, datastore_path):
+    """Confirm the legit confirm-then-POST flow wipes the provider credentials.
+
+    Post-LLMSettings: /llm/clear strips only the connection fields (model, api_key,
+    api_base, provider_kind, local_token_multiplier). User-set toggles, the global
+    summary prompt, monthly budgets, and system token counters survive. This matches
+    the settings-page "empty model" save semantic and the LLMSettings.CONNECTION_FIELDS
+    grouping — see PYDANTIC_MIGRATION.md.
+    """
+    ds = client.application.config.get('DATASTORE')
+    _configure_llm(ds)
+    assert ds.data['settings']['application'].get('llm', {}).get('api_key') == CANARY_KEY
+
+    res = client.post(url_for('settings.llm.llm_clear'), follow_redirects=True)
+    assert res.status_code == 200
+
+    # The api_key must be gone (this is what the test really cares about).
+    llm = ds.data['settings']['application'].get('llm') or {}
+    assert 'api_key' not in llm, f"api_key should have been wiped, got: {llm!r}"
+    assert 'model' not in llm
+    assert 'api_base' not in llm

changedetectionio/tests/test_llm_change_summary.py

@@ -28,7 +28,11 @@ def _set_response(datastore_path, content):
 
 def _configure_llm(client):
     ds = client.application.config.get('DATASTORE')
-    ds.data['settings']['application']['llm'] = {'model': 'gpt-4o-mini', 'api_key': 'sk-test'}
+    # Merge into the existing llm dict so other test setup (e.g. change_summary_default
+    # set via _set_global_default) survives.
+    existing = ds.data['settings']['application'].get('llm') or {}
+    existing.update({'model': 'gpt-4o-mini', 'api_key': 'sk-test'})
+    ds.data['settings']['application']['llm'] = existing
 
 
 # ---------------------------------------------------------------------------
@@ -238,7 +242,9 @@ def test_llm_summary_ajax_error_displayed_not_silenced(
 # ---------------------------------------------------------------------------
 
 def _set_global_default(ds, prompt):
-    ds.data['settings']['application']['llm_change_summary_default'] = prompt
+    llm = ds.data['settings']['application'].get('llm') or {}
+    llm['change_summary_default'] = prompt
+    ds.data['settings']['application']['llm'] = llm
 
 
 def test_global_default_used_when_watch_and_tag_have_none(
@@ -329,7 +335,7 @@ def test_hardcoded_fallback_when_nothing_set(
     watch['llm_change_summary'] = ''
 
     # Ensure global default is also empty
-    ds.data['settings']['application']['llm_change_summary_default'] = ''
+    _set_global_default(ds, '')
 
     assert get_effective_summary_prompt(watch, ds) == DEFAULT_CHANGE_SUMMARY_PROMPT
 
@@ -391,8 +397,8 @@ def test_llm_summary_ajax_sets_last_viewed(
 def test_global_default_saved_and_loaded_via_settings_form(
         client, live_server, measure_memory_usage, datastore_path):
     """
-    Submitting the settings form persists llm_change_summary_default at
-    settings.application level (not inside the llm credentials dict).
+    Submitting the settings form persists the global default prompt into
+    application.llm.change_summary_default (single nested home for all LLM settings).
     """
     from changedetectionio.tests.util import live_server_setup
     live_server_setup(live_server)
@@ -405,21 +411,20 @@ def test_global_default_saved_and_loaded_via_settings_form(
             'application-empty_pages_are_a_change': '',
             'requests-time_between_check-minutes': 180,
             'application-fetch_backend': 'html_requests',
-            'llm-llm_change_summary_default': 'Saved global prompt.',
+            'llm-change_summary_default': 'Saved global prompt.',
             # Keep existing model so llm block is retained
-            'llm-llm_model': 'gpt-4o-mini',
+            'llm-model': 'gpt-4o-mini',
         },
         follow_redirects=True,
     )
     assert b'Settings updated.' in res.data
 
     ds = client.application.config.get('DATASTORE')
-    stored = ds.data['settings']['application'].get('llm_change_summary_default', '')
-    assert stored == 'Saved global prompt.', f"Got: {stored!r}"
-
-    # Must NOT be buried inside the llm credentials dict
     llm_dict = ds.data['settings']['application'].get('llm', {})
-    assert 'change_summary_default' not in llm_dict
+    assert llm_dict.get('change_summary_default') == 'Saved global prompt.', f"Got: {llm_dict!r}"
+
+    # And the old flat key must not be re-introduced
+    assert 'llm_change_summary_default' not in ds.data['settings']['application']
 
     delete_all_watches(client)
 
@@ -437,10 +442,14 @@ def test_global_default_survives_llm_clear(
     ds = client.application.config.get('DATASTORE')
     _set_global_default(ds, 'Surviving prompt.')
 
-    res = client.get(url_for('settings.llm.llm_clear'), follow_redirects=True)
+    res = client.post(url_for('settings.llm.llm_clear'), follow_redirects=True)
     assert res.status_code == 200
 
-    assert ds.data['settings']['application'].get('llm_change_summary_default') == 'Surviving prompt.'
+    llm_dict = ds.data['settings']['application'].get('llm') or {}
+    assert llm_dict.get('change_summary_default') == 'Surviving prompt.'
+    # The credential fields should be gone
+    assert 'model' not in llm_dict
+    assert 'api_key' not in llm_dict
 
     delete_all_watches(client)
 

changedetectionio/tests/test_llm_token_budget.py

@@ -168,9 +168,9 @@ def test_settings_form_preserves_token_counters(
         url_for('settings.settings_page'),
         data={
             # LLM sub-form fields
-            'llm-llm_model': 'gpt-4o',
-            'llm-llm_api_key': 'sk-different-key',
-            'llm-llm_api_base': '',
+            'llm-model': 'gpt-4o',
+            'llm-api_key': 'sk-different-key',
+            'llm-api_base': '',
             # Minimal required fields to pass form validation
             'application-pager_size': '50',
             'application-notification_format': 'System default',
@@ -196,6 +196,81 @@ def test_settings_form_preserves_token_counters(
     delete_all_watches(client)
 
 
+def test_settings_form_blank_llm_integer_fields_preserve_stored_values(
+        client, live_server, measure_memory_usage, datastore_path):
+    """
+    Empty IntegerField submissions come back as None from WTForms. LLMSettings
+    declares token_budget_month / max_input_chars / max_tokens_per_count_period /
+    local_token_multiplier as strict `int`, so a None passed through to
+    model_validate raises ValidationError and 500s the settings save.
+
+    Regression for settings/__init__.py — the LLM merge must drop None values
+    (treat them like absent keys) so blank IntegerField submissions preserve
+    the stored value instead of crashing the form.
+    """
+    ds = client.application.config.get('DATASTORE')
+    ds.data['settings']['application']['llm'] = {
+        'model': 'gpt-4o',
+        'api_key': 'sk-existing',
+        'token_budget_month': 50000,
+        'max_input_chars': 200000,
+        'max_tokens_per_count_period': 1000,
+        'local_token_multiplier': 3,
+    }
+
+    res = client.post(
+        url_for('settings.settings_page'),
+        data={
+            'llm-model': 'gpt-4o',
+            'llm-api_key': '',
+            'llm-api_base': '',
+            # The bug-trigger: every LLM IntegerField submitted blank
+            'llm-token_budget_month': '',
+            'llm-max_input_chars': '',
+            'llm-max_tokens_per_count_period': '',
+            'llm-local_token_multiplier': '',
+            # Minimal required fields for the rest of the form to validate.
+            # 'System default' is popped from notification_format choices for the
+            # global form, so it must be one of the real codes (e.g. 'html').
+            'application-pager_size': '50',
+            'application-notification_format': 'html',
+            'application-fetch_backend': 'html_requests',
+            'application-rss_diff_length': '5',
+            'application-filter_failure_notification_threshold_attempts': '0',
+            'requests-time_between_check-days': '0',
+            'requests-time_between_check-hours': '0',
+            'requests-time_between_check-minutes': '5',
+            'requests-time_between_check-seconds': '0',
+            'requests-time_between_check-weeks': '0',
+            'requests-jitter_seconds': '0',
+            'requests-workers': '10',
+            'requests-timeout': '60',
+        },
+        follow_redirects=True,
+    )
+    assert res.status_code == 200, \
+        f"Settings save crashed on blank LLM IntegerField submission (got {res.status_code})"
+    # Sanity: the form must have actually validated and reached the LLM save path
+    # — without this the test would trivially pass because the buggy code never ran.
+    assert b'Settings updated.' in res.data, \
+        "Settings form did not validate — the bug-path was never exercised. Check fixture fields."
+    body = res.data.decode('utf-8', errors='replace')
+    assert 'ValidationError' not in body, \
+        "Pydantic ValidationError leaked into the response — blank IntegerField wasn't filtered"
+
+    llm = ds.data['settings']['application'].get('llm') or {}
+    assert llm.get('token_budget_month') == 50000, \
+        f"Blank submission must preserve stored token_budget_month (got {llm.get('token_budget_month')!r})"
+    assert llm.get('max_input_chars') == 200000, \
+        f"Blank submission must preserve stored max_input_chars (got {llm.get('max_input_chars')!r})"
+    assert llm.get('max_tokens_per_count_period') == 1000, \
+        f"Blank submission must preserve stored max_tokens_per_count_period (got {llm.get('max_tokens_per_count_period')!r})"
+    assert llm.get('local_token_multiplier') == 3, \
+        f"Blank submission must preserve stored local_token_multiplier (got {llm.get('local_token_multiplier')!r})"
+
+    delete_all_watches(client)
+
+
 def test_settings_form_cannot_inject_fake_token_counts(
         client, live_server, measure_memory_usage, datastore_path):
     """
@@ -209,9 +284,9 @@ def test_settings_form_cannot_inject_fake_token_counts(
     res = client.post(
         url_for('settings.settings_page'),
         data={
-            'llm-llm_model': 'gpt-4o-mini',
-            'llm-llm_api_key': 'sk-test',
-            'llm-llm_api_base': '',
+            'llm-model': 'gpt-4o-mini',
+            'llm-api_key': 'sk-test',
+            'llm-api_base': '',
             # Attempted injection of token counter fields
             'llm-tokens_this_month': '0',
             'llm-tokens_total_cumulative': '0',
@@ -471,9 +546,9 @@ def test_cost_fields_are_tamper_proof_via_settings_form(
     client.post(
         url_for('settings.settings_page'),
         data={
-            'llm-llm_model': 'gpt-4o',
-            'llm-llm_api_key': 'sk-test',
-            'llm-llm_api_base': '',
+            'llm-model': 'gpt-4o',
+            'llm-api_key': 'sk-test',
+            'llm-api_base': '',
             'llm-cost_usd_this_month': '0',       # injection attempt
             'llm-cost_usd_total_cumulative': '0',  # injection attempt
             'application-pager_size': '50',

changedetectionio/tests/test_notification.py

@@ -634,6 +634,12 @@ def _test_color_notifications(client, notification_body_token, datastore_path):
 def test_html_color_notifications(client, live_server, measure_memory_usage, datastore_path):
     _test_color_notifications(client, '{{diff}}',datastore_path=datastore_path)
     _test_color_notifications(client, '{{diff_full}}',datastore_path=datastore_path)
+    # Regression: the html-output escape pass in handler.py used to convert
+    # FormattableDiff into a plain str, stripping its __call__ and breaking any
+    # {{ diff(...) }} / {{ diff_added(...) }} token on htmlcolor/html notifications
+    # with 'str' object is not callable (see commit 08d30c6 + #3923).
+    # word_diff=false reproduces the exact form the user-reported failure used.
+    _test_color_notifications(client, '{{diff(word_diff=false)}}', datastore_path=datastore_path)
 
 
 def _test_custom_html_in_notification_body_not_escaped(client, datastore_path, content_type=None):

changedetectionio/tests/test_security.py

@@ -760,7 +760,9 @@ def test_ssrf_private_ip_blocked(client, live_server, monkeypatch, measure_memor
 
     f = RequestsFetcher()
 
-    with patch('changedetectionio.content_fetchers.requests.is_private_hostname', return_value=True):
+    # Patch the underlying is_private_hostname in validate_url — the fetcher now goes through
+    # is_url_private_or_parser_confused() (GHSA-rph4-96w6-q594), which calls it transitively.
+    with patch('changedetectionio.validate_url.is_private_hostname', return_value=True):
         with pytest.raises(Exception, match='private/reserved'):
             f._run_sync(
                 url='http://example.com/',
@@ -784,7 +786,7 @@ def test_ssrf_private_ip_blocked(client, live_server, monkeypatch, measure_memor
         return hostname in {'169.254.169.254', '10.0.0.1', '172.16.0.1',
                             '192.168.0.1', '127.0.0.1', '::1'}
 
-    with patch('changedetectionio.content_fetchers.requests.is_private_hostname',
+    with patch('changedetectionio.validate_url.is_private_hostname',
                side_effect=_private_only_for_redirect):
         with patch('requests.Session.request', return_value=mock_redirect):
             with pytest.raises(Exception, match='Redirect blocked'):
@@ -829,6 +831,113 @@ def test_unresolvable_hostname_is_allowed(client, live_server, monkeypatch):
         "Unresolvable hostname watch should appear in the watch overview list"
 
 
+def test_ghsa_rph4_96w6_q594_urlparse_urllib3_parser_differential_ssrf(client, live_server, monkeypatch, measure_memory_usage, datastore_path):
+    """
+    GHSA-rph4-96w6-q594: SSRF via urlparse/urllib3 parser differential.
+
+    A URL like  http://INTERNAL:8888\\@PUBLIC/  is parsed two different ways:
+      - urlparse() treats \\@ as a credential separator     → hostname = PUBLIC
+      - urllib3   treats \\  as a path character           → hostname = INTERNAL
+    The pre-fetch SSRF check used urlparse(), but requests/urllib3 actually connected
+    to INTERNAL. Fix: parser-agnostic gate that (a) blocks any URL containing a
+    backslash and (b) validates every hostname both parsers produce.
+
+    Covers:
+      1. extract_url_hostnames() reveals BOTH hostnames for the payload
+      2. is_url_private_or_parser_confused() blocks backslash payloads outright
+      3. is_safe_valid_url() rejects backslash payloads at add-time
+      4. The /api/v1/watch add endpoint rejects the payload
+      5. The requests fetcher refuses the payload at fetch-time
+       6. The redirect-following loop refuses a backslash payload in Location
+    """
+    from unittest.mock import patch, MagicMock
+    from changedetectionio.validate_url import (
+        extract_url_hostnames,
+        is_safe_valid_url,
+        is_url_private_or_parser_confused,
+    )
+
+    monkeypatch.setenv('ALLOW_IANA_RESTRICTED_ADDRESSES', 'false')
+
+    # The published proof-of-concept payload — backslash splits the two parsers' views.
+    payload = "http://169.254.169.254:8888" + chr(92) + "@httpbin.org/latest/meta-data/"
+
+    # ---------------------------------------------------------------
+    # 1. extract_url_hostnames() returns BOTH parsers' hostnames
+    # ---------------------------------------------------------------
+    hosts = extract_url_hostnames(payload)
+    assert '169.254.169.254' in hosts, \
+        f"urllib3 sees 169.254.169.254 as the connect target; extract_url_hostnames must surface it. Got {hosts!r}"
+    assert 'httpbin.org' in hosts, \
+        f"urlparse sees httpbin.org; extract_url_hostnames must surface it too. Got {hosts!r}"
+
+    # ---------------------------------------------------------------
+    # 2. Parser-agnostic gate blocks the payload
+    # ---------------------------------------------------------------
+    assert is_url_private_or_parser_confused(payload), \
+        "Parser-differential payload must be blocked by the SSRF gate"
+
+    # And a plain backslash anywhere in the URL is enough to block, even without a private IP
+    assert is_url_private_or_parser_confused("http://example.com" + chr(92) + "@evil.com/"), \
+        "Any backslash in a URL must trigger the parser-differential block"
+
+    # Sanity: a regular public URL is not blocked
+    assert not is_url_private_or_parser_confused("http://example.com/path"), \
+        "Plain public URLs must continue to pass the gate"
+
+    # ---------------------------------------------------------------
+    # 3. is_safe_valid_url() rejects backslash payloads at add-time
+    # ---------------------------------------------------------------
+    assert not is_safe_valid_url(payload), \
+        "is_safe_valid_url must reject URLs containing a backslash (parser-differential vector)"
+
+    # ---------------------------------------------------------------
+    # 4. The watch-add API endpoint rejects the payload
+    # ---------------------------------------------------------------
+    api_key = live_server.app.config['DATASTORE'].data['settings']['application'].get('api_access_token')
+    res = client.post(
+        url_for('createwatch'),
+        data='{"url": "%s", "fetch_backend": "html_requests"}' % payload,
+        headers={'x-api-key': api_key, 'Content-Type': 'application/json'},
+    )
+    assert res.status_code >= 400, \
+        f"API must refuse to create a watch for parser-differential URL; got status {res.status_code} body {res.data!r}"
+
+    # ---------------------------------------------------------------
+    # 5. Requests fetcher refuses the payload at fetch-time
+    # ---------------------------------------------------------------
+    from changedetectionio.content_fetchers.requests import fetcher as RequestsFetcher
+
+    f = RequestsFetcher()
+    with pytest.raises(Exception, match='private/reserved|parser-differential'):
+        f._run_sync(
+            url=payload,
+            timeout=5,
+            request_headers={},
+            request_body=None,
+            request_method='GET',
+        )
+
+    # ---------------------------------------------------------------
+    # 6. A 302 Location header pointing at a backslash payload is blocked
+    #    (open-redirect → SSRF via parser differential)
+    # ---------------------------------------------------------------
+    mock_redirect = MagicMock()
+    mock_redirect.is_redirect = True
+    mock_redirect.status_code = 302
+    mock_redirect.headers = {'Location': payload}
+
+    with patch('requests.Session.request', return_value=mock_redirect):
+        with pytest.raises(Exception, match='Redirect blocked'):
+            f._run_sync(
+                url='http://example.com/',
+                timeout=5,
+                request_headers={},
+                request_body=None,
+                request_method='GET',
+            )
+
+
 def test_ghsa_8757_69j2_hx56_backup_restore_history_path_traversal(client, live_server, measure_memory_usage, datastore_path):
     """
     GHSA-8757-69j2-hx56: Crafted backup ZIP with absolute path in history.txt must not

changedetectionio/tests/unit/test_llm_settings.py

@@ -0,0 +1,161 @@
+#!/usr/bin/env python3
+
+# run from dir above changedetectionio/ dir
+# python3 -m unittest changedetectionio.tests.unit.test_llm_settings
+
+import unittest
+
+from pydantic import ValidationError
+
+from changedetectionio.model.LLMSettings import (
+    LLMSettings,
+    LLM_DEFAULT_BUDGET_ACTION,
+    LLM_DEFAULT_LOCAL_TOKEN_MULTIPLIER,
+    LLM_DEFAULT_MAX_INPUT_CHARS,
+    LLM_DEFAULT_MAX_SUMMARY_TOKENS,
+    LLM_DEFAULT_THINKING_BUDGET,
+)
+
+
+class TestLLMSettingsDefaults(unittest.TestCase):
+    def test_empty_dict_yields_default_model(self):
+        s = LLMSettings.model_validate({})
+        self.assertTrue(s.enabled)
+        self.assertFalse(s.debug)
+        self.assertEqual(s.model, '')
+        self.assertEqual(s.api_key, '')
+        self.assertEqual(s.thinking_budget, LLM_DEFAULT_THINKING_BUDGET)
+        self.assertEqual(s.max_summary_tokens, LLM_DEFAULT_MAX_SUMMARY_TOKENS)
+        self.assertEqual(s.local_token_multiplier, LLM_DEFAULT_LOCAL_TOKEN_MULTIPLIER)
+        self.assertEqual(s.max_input_chars, LLM_DEFAULT_MAX_INPUT_CHARS)
+        self.assertEqual(s.budget_action, LLM_DEFAULT_BUDGET_ACTION)
+        self.assertEqual(s.tokens_total_cumulative, 0)
+        self.assertEqual(s.cost_usd_this_month, 0.0)
+
+    def test_default_construct_matches_validate_empty(self):
+        self.assertEqual(LLMSettings().model_dump(), LLMSettings.model_validate({}).model_dump())
+
+
+class TestLLMSettingsValidation(unittest.TestCase):
+    def test_stripped_keys_validate(self):
+        s = LLMSettings.model_validate({'model': 'gpt-4o-mini', 'enabled': False})
+        self.assertEqual(s.model, 'gpt-4o-mini')
+        self.assertFalse(s.enabled)
+
+
+class TestLLMSettingsTypeCoercion(unittest.TestCase):
+    def test_select_field_string_int_coerces_to_int(self):
+        # WTForms SelectField returns the choice key as a string ('500');
+        # Pydantic coerces to int so storage stays typed.
+        s = LLMSettings.model_validate({'thinking_budget': '500', 'max_summary_tokens': '5000'})
+        self.assertEqual(s.thinking_budget, 500)
+        self.assertEqual(s.max_summary_tokens, 5000)
+
+    def test_invalid_int_raises(self):
+        with self.assertRaises(ValidationError):
+            LLMSettings.model_validate({'thinking_budget': 'not_a_number'})
+
+
+class TestLLMSettingsExtraForbid(unittest.TestCase):
+    def test_unknown_key_raises(self):
+        # extra='forbid' is the security gate against CWE-915 mass-assignment.
+        with self.assertRaises(ValidationError) as ctx:
+            LLMSettings.model_validate({'model': 'gpt-4o-mini', 'evil_field': 'pwn'})
+        self.assertIn('evil_field', str(ctx.exception))
+
+    def test_dunder_key_raises(self):
+        with self.assertRaises(ValidationError):
+            LLMSettings.model_validate({'model': 'gpt-4o-mini', '__class__': 'attack'})
+
+    def test_legitimate_unknown_key_also_raises(self):
+        # No "future-tolerant" silent acceptance — new fields must be declared.
+        with self.assertRaises(ValidationError):
+            LLMSettings.model_validate({'maybe_future_counter': 42})
+
+    def test_legacy_prefixed_key_raises(self):
+        # Pre-update_31 storage used flat application.llm_* keys (handled by the
+        # migration). After migration the prefix is gone — and any code path that
+        # still tries to write a prefixed key into the LLM dict must be rejected
+        # so the prefix can never reappear through any side channel.
+        with self.assertRaises(ValidationError):
+            LLMSettings.model_validate({'llm_model': 'gpt-4o-mini'})
+
+
+class TestLLMSettingsDumpShapes(unittest.TestCase):
+    def test_dump_uses_field_names(self):
+        s = LLMSettings.model_validate({'model': 'gpt-4o-mini'})
+        out = s.model_dump()
+        self.assertEqual(out['model'], 'gpt-4o-mini')
+        self.assertNotIn('llm_model', out)
+
+    def test_dump_exclude_connection_drops_provider_fields(self):
+        s = LLMSettings.model_validate({
+            'model': 'gpt-4o-mini', 'api_key': 'sk-test', 'api_base': 'https://example',
+            'provider_kind': 'ollama', 'local_token_multiplier': 5,
+            'enabled': False, 'tokens_this_month': 42,
+        })
+        out = s.model_dump(exclude=set(LLMSettings.CONNECTION_FIELDS))
+        for k in LLMSettings.CONNECTION_FIELDS:
+            self.assertNotIn(k, out, f"connection field {k} should be excluded")
+        # Non-connection fields survive
+        self.assertFalse(out['enabled'])
+        self.assertEqual(out['tokens_this_month'], 42)
+
+
+class TestLLMSettingsFieldGroups(unittest.TestCase):
+    def test_connection_fields_all_declared(self):
+        declared = set(LLMSettings.model_fields)
+        for name in LLMSettings.CONNECTION_FIELDS:
+            self.assertIn(name, declared, f"CONNECTION_FIELDS lists undeclared field: {name}")
+
+    def test_protected_fields_all_declared(self):
+        declared = set(LLMSettings.model_fields)
+        for name in LLMSettings.PROTECTED_FIELDS:
+            self.assertIn(name, declared, f"PROTECTED_FIELDS lists undeclared field: {name}")
+
+    def test_connection_and_protected_disjoint(self):
+        # System-managed counters and user-set provider config must not overlap —
+        # otherwise a "clear credentials" action would also wipe counters.
+        overlap = set(LLMSettings.CONNECTION_FIELDS) & set(LLMSettings.PROTECTED_FIELDS)
+        self.assertEqual(overlap, set(), f"CONNECTION/PROTECTED overlap: {overlap}")
+
+
+class TestLLMSettingsRoundTrip(unittest.TestCase):
+    def test_counter_round_trip_via_dump_load(self):
+        original = LLMSettings.model_validate({
+            'model': 'gpt-4o-mini',
+            'tokens_total_cumulative': 123456,
+            'tokens_this_month': 789,
+            'tokens_month_key': '2026-05',
+            'cost_usd_total_cumulative': 12.34,
+            'cost_usd_this_month': 0.56,
+        })
+        roundtripped = LLMSettings.model_validate(original.model_dump())
+        self.assertEqual(roundtripped.tokens_total_cumulative, 123456)
+        self.assertEqual(roundtripped.tokens_this_month, 789)
+        self.assertEqual(roundtripped.tokens_month_key, '2026-05')
+        self.assertEqual(roundtripped.cost_usd_total_cumulative, 12.34)
+        self.assertEqual(roundtripped.cost_usd_this_month, 0.56)
+
+    def test_form_merge_preserves_counters(self):
+        # The POST handler pattern: validate existing storage, overlay form input
+        # (with PROTECTED_FIELDS stripped), re-validate. Counters in storage must
+        # survive even if the form somehow tried to set them.
+        existing = LLMSettings.model_validate({
+            'model': 'gpt-4o-mini', 'tokens_total_cumulative': 99999,
+        })
+        form_input = {
+            'model': 'claude-3-5-haiku-20251001',
+            'enabled': False,
+        }
+        # Strip protected fields from form input as the route handler does
+        for protected in LLMSettings.PROTECTED_FIELDS:
+            form_input.pop(protected, None)
+        merged = LLMSettings.model_validate({**existing.model_dump(), **form_input})
+        self.assertEqual(merged.model, 'claude-3-5-haiku-20251001')
+        self.assertFalse(merged.enabled)
+        self.assertEqual(merged.tokens_total_cumulative, 99999)
+
+
+if __name__ == '__main__':
+    unittest.main()

changedetectionio/translations/README.md

@@ -247,34 +247,31 @@ dennis-cmd lint --excluderules=W302 changedetectionio/translations/
 
 The `W303` rule ensures that HTML tags in the `msgstr` match the `msgid`. This is crucial for catching broken markup (e.g., missing closing tags).
 
-##### Handling intentional deviations and false positives
+##### Handling intentional deviations
 
-Some W303 warnings are intentional or result from upstream false positives.
+Some W303 warnings are intentional.
 Use the `dennis-ignore: W303` comment in the source files (templates or Python code) within a `TRANSLATORS` comment to suppress these warnings.
 This ensures the ignore instruction is extracted into the `.po` files.
 
 - **CJK italic policy**: When replacing `<i>` with locale-conventional quotation marks, tags will no longer match.
-- **Upstream false positive**: Dennis misinterprets certain HTML tags (e.g., `<title>`) within `msgstr`. See https://github.com/mozilla/dennis/issues/213.
 
 **Examples in Jinja2 templates:**
 
 ```jinja
 {# 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>
-
-{# TRANSLATORS: dennis-ignore: W303 - False positive caused by <title>. https://github.com/mozilla/dennis/issues/213 #}
-<td>{{ _('The page title of the watch, uses <title> if not set, falls back to URL') }}</td>
 ```
 
 **Example in Python source:**
 
 ```python
-# dennis-ignore: W303 - False positive caused by <title>. https://github.com/mozilla/dennis/issues/213
-use_page_title_in_list = BooleanField(_l('Use page <title> in watch overview list'))
+# dennis-ignore: W303 - CJK fonts lack native italics; allow substitution with conventional local styling.
+message = StringField(_l('This is <i>experimental</i> and may change'))
 ```
 
 ---
 
+
 ## CI linter
 
 A GitHub Actions job (`lint-template-i18n`) checks for adjacent `{{ _(...) }}` calls on the same line

changedetectionio/translations/cs/LC_MESSAGES/messages.po

@@ -360,6 +360,12 @@ msgstr "Všechna oznámení ztlumena."
 msgid "All notifications unmuted."
 msgstr "Všechna oznámení odtlumena."
 
+#: changedetectionio/blueprint/settings/llm.py
+msgid ""
+"api_key is required when api_base differs from the saved configuration. Refusing to send the stored API key to a "
+"different endpoint."
+msgstr ""
+
 #: changedetectionio/blueprint/settings/llm.py
 msgid "AI / LLM configuration removed."
 msgstr "AI / LLM konfigurace odstraněna."
@@ -836,6 +842,10 @@ msgstr ""
 msgid "AI Provider"
 msgstr ""
 
+#: changedetectionio/blueprint/settings/templates/settings_llm_tab.html
+msgid "Master switch — when off, all AI lookups are skipped even if a provider is configured below."
+msgstr ""
+
 #: changedetectionio/blueprint/settings/templates/settings_llm_tab.html
 msgid "Third-party data transfer — please read"
 msgstr ""
@@ -1086,6 +1096,12 @@ msgstr ""
 msgid "characters — currently enforcing: %(limit)s"
 msgstr ""
 
+#: changedetectionio/blueprint/settings/templates/settings_llm_tab.html
+msgid ""
+"tokens — skips AI evaluation on a watch once its usage within the current period (monthly) hits this cap (0 = "
+"unlimited)"
+msgstr ""
+
 #: changedetectionio/blueprint/settings/templates/settings_llm_tab.html
 msgid "No AI usage recorded yet."
 msgstr ""
@@ -2943,7 +2959,6 @@ msgstr "Spojit všechny následující položky"
 msgid "Match any of the following"
 msgstr "Přiřaďte kteroukoli z následujících možností"
 
-#. dennis-ignore: W303 - False positive caused by <title>. https://github.com/mozilla/dennis/issues/213
 #: changedetectionio/forms.py
 msgid "Use page <title> in list"
 msgstr "V seznamu použijte stránku <title>"
@@ -3043,7 +3058,6 @@ msgstr "Aktualizace UI v reálném čase"
 msgid "Favicons Enabled"
 msgstr "Povolit favikony"
 
-#. dennis-ignore: W303 - False positive caused by <title>. https://github.com/mozilla/dennis/issues/213
 #: changedetectionio/forms.py
 msgid "Use page <title> in watch overview list"
 msgstr "Použijte stránku <title> v přehledu sledování"
@@ -3156,12 +3170,8 @@ msgstr ""
 msgid "Default AI Change Summary prompt"
 msgstr ""
 
-#: changedetectionio/forms.py
-msgid "Max tokens per check"
-msgstr ""
-
-#: changedetectionio/forms.py
-msgid "Max cumulative tokens (per watch)"
+#: changedetectionio/blueprint/settings/templates/settings_llm_tab.html changedetectionio/forms.py
+msgid "Max tokens per watch per period"
 msgstr ""
 
 #: changedetectionio/forms.py
@@ -3172,6 +3182,10 @@ msgstr "Měsíční rozpočet tokenů"
 msgid "Max input characters"
 msgstr ""
 
+#: changedetectionio/forms.py
+msgid "Enable AI / LLM features"
+msgstr ""
+
 #: changedetectionio/forms.py
 msgid "Replace {{diff}} notification token with AI summary"
 msgstr ""
@@ -3437,7 +3451,6 @@ msgstr ""
 msgid "The UUID of the watch."
 msgstr "UUID monitoru."
 
-#. dennis-ignore: W303 - False positive caused by <title>. https://github.com/mozilla/dennis/issues/213
 #: changedetectionio/templates/_common_fields.html
 msgid "The page title of the watch, uses <title> if not set, falls back to URL"
 msgstr ""
@@ -4165,6 +4178,17 @@ msgstr "Změnit jazyk"
 msgid "Change language"
 msgstr "Změnit jazyk"
 
+#: changedetectionio/validate_url.py
+msgid "API Base URL is not a valid http(s) URL."
+msgstr ""
+
+#: changedetectionio/validate_url.py
+msgid ""
+"API Base URL resolves to a private, loopback, link-local or reserved IP address and was blocked to prevent SSRF. To "
+"allow LLM endpoints on private networks (e.g. a local Ollama server) set the environment variable "
+"ALLOW_IANA_RESTRICTED_ADDRESSES=true and restart."
+msgstr ""
+
 #: changedetectionio/widgets/ternary_boolean.py
 msgid "Yes"
 msgstr "Ano"

changedetectionio/translations/de/LC_MESSAGES/messages.po

@@ -366,6 +366,12 @@ msgstr "Alle Benachrichtigungen stummgeschaltet."
 msgid "All notifications unmuted."
 msgstr "Alle Benachrichtigungen entstummt."
 
+#: changedetectionio/blueprint/settings/llm.py
+msgid ""
+"api_key is required when api_base differs from the saved configuration. Refusing to send the stored API key to a "
+"different endpoint."
+msgstr ""
+
 #: changedetectionio/blueprint/settings/llm.py
 msgid "AI / LLM configuration removed."
 msgstr ""
@@ -852,6 +858,10 @@ msgstr ""
 msgid "AI Provider"
 msgstr ""
 
+#: changedetectionio/blueprint/settings/templates/settings_llm_tab.html
+msgid "Master switch — when off, all AI lookups are skipped even if a provider is configured below."
+msgstr ""
+
 #: changedetectionio/blueprint/settings/templates/settings_llm_tab.html
 msgid "Third-party data transfer — please read"
 msgstr ""
@@ -1102,6 +1112,12 @@ msgstr ""
 msgid "characters — currently enforcing: %(limit)s"
 msgstr ""
 
+#: changedetectionio/blueprint/settings/templates/settings_llm_tab.html
+msgid ""
+"tokens — skips AI evaluation on a watch once its usage within the current period (monthly) hits this cap (0 = "
+"unlimited)"
+msgstr ""
+
 #: changedetectionio/blueprint/settings/templates/settings_llm_tab.html
 msgid "No AI usage recorded yet."
 msgstr ""
@@ -2995,7 +3011,6 @@ msgstr "Passen Sie alle folgenden Punkte an"
 msgid "Match any of the following"
 msgstr "Entspricht einer der folgenden Bedingungen"
 
-#. dennis-ignore: W303 - False positive caused by <title>. https://github.com/mozilla/dennis/issues/213
 #: changedetectionio/forms.py
 msgid "Use page <title> in list"
 msgstr "Verwenden Sie Seite <title> in der Liste"
@@ -3095,7 +3110,6 @@ msgstr "Echtzeit-UI-Updates aktiviert"
 msgid "Favicons Enabled"
 msgstr "Favicons Aktiviert"
 
-#. dennis-ignore: W303 - False positive caused by <title>. https://github.com/mozilla/dennis/issues/213
 #: changedetectionio/forms.py
 msgid "Use page <title> in watch overview list"
 msgstr "Verwenden Sie die Seite <title> in der Übersichtsliste der Beobachtungen"
@@ -3208,12 +3222,8 @@ msgstr ""
 msgid "Default AI Change Summary prompt"
 msgstr ""
 
-#: changedetectionio/forms.py
-msgid "Max tokens per check"
-msgstr ""
-
-#: changedetectionio/forms.py
-msgid "Max cumulative tokens (per watch)"
+#: changedetectionio/blueprint/settings/templates/settings_llm_tab.html changedetectionio/forms.py
+msgid "Max tokens per watch per period"
 msgstr ""
 
 #: changedetectionio/forms.py
@@ -3224,6 +3234,10 @@ msgstr ""
 msgid "Max input characters"
 msgstr ""
 
+#: changedetectionio/forms.py
+msgid "Enable AI / LLM features"
+msgstr ""
+
 #: changedetectionio/forms.py
 msgid "Replace {{diff}} notification token with AI summary"
 msgstr ""
@@ -3491,7 +3505,6 @@ msgstr ""
 msgid "The UUID of the watch."
 msgstr "Die UUID der Überwachung."
 
-#. dennis-ignore: W303 - False positive caused by <title>. https://github.com/mozilla/dennis/issues/213
 #: changedetectionio/templates/_common_fields.html
 msgid "The page title of the watch, uses <title> if not set, falls back to URL"
 msgstr ""
@@ -4221,6 +4234,17 @@ msgstr "Sprache ändern"
 msgid "Change language"
 msgstr "Sprache ändern"
 
+#: changedetectionio/validate_url.py
+msgid "API Base URL is not a valid http(s) URL."
+msgstr ""
+
+#: changedetectionio/validate_url.py
+msgid ""
+"API Base URL resolves to a private, loopback, link-local or reserved IP address and was blocked to prevent SSRF. To "
+"allow LLM endpoints on private networks (e.g. a local Ollama server) set the environment variable "
+"ALLOW_IANA_RESTRICTED_ADDRESSES=true and restart."
+msgstr ""
+
 #: changedetectionio/widgets/ternary_boolean.py
 msgid "Yes"
 msgstr "Ja"

changedetectionio/translations/en_GB/LC_MESSAGES/messages.po

@@ -358,6 +358,12 @@ msgstr ""
 msgid "All notifications unmuted."
 msgstr ""
 
+#: changedetectionio/blueprint/settings/llm.py
+msgid ""
+"api_key is required when api_base differs from the saved configuration. Refusing to send the stored API key to a "
+"different endpoint."
+msgstr ""
+
 #: changedetectionio/blueprint/settings/llm.py
 msgid "AI / LLM configuration removed."
 msgstr ""
@@ -834,6 +840,10 @@ msgstr ""
 msgid "AI Provider"
 msgstr ""
 
+#: changedetectionio/blueprint/settings/templates/settings_llm_tab.html
+msgid "Master switch — when off, all AI lookups are skipped even if a provider is configured below."
+msgstr ""
+
 #: changedetectionio/blueprint/settings/templates/settings_llm_tab.html
 msgid "Third-party data transfer — please read"
 msgstr ""
@@ -1084,6 +1094,12 @@ msgstr ""
 msgid "characters — currently enforcing: %(limit)s"
 msgstr ""
 
+#: changedetectionio/blueprint/settings/templates/settings_llm_tab.html
+msgid ""
+"tokens — skips AI evaluation on a watch once its usage within the current period (monthly) hits this cap (0 = "
+"unlimited)"
+msgstr ""
+
 #: changedetectionio/blueprint/settings/templates/settings_llm_tab.html
 msgid "No AI usage recorded yet."
 msgstr ""
@@ -2937,7 +2953,6 @@ msgstr ""
 msgid "Match any of the following"
 msgstr ""
 
-#. dennis-ignore: W303 - False positive caused by <title>. https://github.com/mozilla/dennis/issues/213
 #: changedetectionio/forms.py
 msgid "Use page <title> in list"
 msgstr ""
@@ -3037,7 +3052,6 @@ msgstr ""
 msgid "Favicons Enabled"
 msgstr ""
 
-#. dennis-ignore: W303 - False positive caused by <title>. https://github.com/mozilla/dennis/issues/213
 #: changedetectionio/forms.py
 msgid "Use page <title> in watch overview list"
 msgstr ""
@@ -3150,12 +3164,8 @@ msgstr ""
 msgid "Default AI Change Summary prompt"
 msgstr ""
 
-#: changedetectionio/forms.py
-msgid "Max tokens per check"
-msgstr ""
-
-#: changedetectionio/forms.py
-msgid "Max cumulative tokens (per watch)"
+#: changedetectionio/blueprint/settings/templates/settings_llm_tab.html changedetectionio/forms.py
+msgid "Max tokens per watch per period"
 msgstr ""
 
 #: changedetectionio/forms.py
@@ -3166,6 +3176,10 @@ msgstr ""
 msgid "Max input characters"
 msgstr ""
 
+#: changedetectionio/forms.py
+msgid "Enable AI / LLM features"
+msgstr ""
+
 #: changedetectionio/forms.py
 msgid "Replace {{diff}} notification token with AI summary"
 msgstr ""
@@ -3431,7 +3445,6 @@ msgstr ""
 msgid "The UUID of the watch."
 msgstr ""
 
-#. dennis-ignore: W303 - False positive caused by <title>. https://github.com/mozilla/dennis/issues/213
 #: changedetectionio/templates/_common_fields.html
 msgid "The page title of the watch, uses <title> if not set, falls back to URL"
 msgstr ""
@@ -4159,6 +4172,17 @@ msgstr ""
 msgid "Change language"
 msgstr ""
 
+#: changedetectionio/validate_url.py
+msgid "API Base URL is not a valid http(s) URL."
+msgstr ""
+
+#: changedetectionio/validate_url.py
+msgid ""
+"API Base URL resolves to a private, loopback, link-local or reserved IP address and was blocked to prevent SSRF. To "
+"allow LLM endpoints on private networks (e.g. a local Ollama server) set the environment variable "
+"ALLOW_IANA_RESTRICTED_ADDRESSES=true and restart."
+msgstr ""
+
 #: changedetectionio/widgets/ternary_boolean.py
 msgid "Yes"
 msgstr ""

changedetectionio/translations/en_US/LC_MESSAGES/messages.po

@@ -358,6 +358,12 @@ msgstr ""
 msgid "All notifications unmuted."
 msgstr ""
 
+#: changedetectionio/blueprint/settings/llm.py
+msgid ""
+"api_key is required when api_base differs from the saved configuration. Refusing to send the stored API key to a "
+"different endpoint."
+msgstr ""
+
 #: changedetectionio/blueprint/settings/llm.py
 msgid "AI / LLM configuration removed."
 msgstr ""
@@ -834,6 +840,10 @@ msgstr ""
 msgid "AI Provider"
 msgstr ""
 
+#: changedetectionio/blueprint/settings/templates/settings_llm_tab.html
+msgid "Master switch — when off, all AI lookups are skipped even if a provider is configured below."
+msgstr ""
+
 #: changedetectionio/blueprint/settings/templates/settings_llm_tab.html
 msgid "Third-party data transfer — please read"
 msgstr ""
@@ -1084,6 +1094,12 @@ msgstr ""
 msgid "characters — currently enforcing: %(limit)s"
 msgstr ""
 
+#: changedetectionio/blueprint/settings/templates/settings_llm_tab.html
+msgid ""
+"tokens — skips AI evaluation on a watch once its usage within the current period (monthly) hits this cap (0 = "
+"unlimited)"
+msgstr ""
+
 #: changedetectionio/blueprint/settings/templates/settings_llm_tab.html
 msgid "No AI usage recorded yet."
 msgstr ""
@@ -2937,7 +2953,6 @@ msgstr ""
 msgid "Match any of the following"
 msgstr ""
 
-#. dennis-ignore: W303 - False positive caused by <title>. https://github.com/mozilla/dennis/issues/213
 #: changedetectionio/forms.py
 msgid "Use page <title> in list"
 msgstr ""
@@ -3037,7 +3052,6 @@ msgstr ""
 msgid "Favicons Enabled"
 msgstr ""
 
-#. dennis-ignore: W303 - False positive caused by <title>. https://github.com/mozilla/dennis/issues/213
 #: changedetectionio/forms.py
 msgid "Use page <title> in watch overview list"
 msgstr ""
@@ -3150,12 +3164,8 @@ msgstr ""
 msgid "Default AI Change Summary prompt"
 msgstr ""
 
-#: changedetectionio/forms.py
-msgid "Max tokens per check"
-msgstr ""
-
-#: changedetectionio/forms.py
-msgid "Max cumulative tokens (per watch)"
+#: changedetectionio/blueprint/settings/templates/settings_llm_tab.html changedetectionio/forms.py
+msgid "Max tokens per watch per period"
 msgstr ""
 
 #: changedetectionio/forms.py
@@ -3166,6 +3176,10 @@ msgstr ""
 msgid "Max input characters"
 msgstr ""
 
+#: changedetectionio/forms.py
+msgid "Enable AI / LLM features"
+msgstr ""
+
 #: changedetectionio/forms.py
 msgid "Replace {{diff}} notification token with AI summary"
 msgstr ""
@@ -3431,7 +3445,6 @@ msgstr ""
 msgid "The UUID of the watch."
 msgstr ""
 
-#. dennis-ignore: W303 - False positive caused by <title>. https://github.com/mozilla/dennis/issues/213
 #: changedetectionio/templates/_common_fields.html
 msgid "The page title of the watch, uses <title> if not set, falls back to URL"
 msgstr ""
@@ -4159,6 +4172,17 @@ msgstr ""
 msgid "Change language"
 msgstr ""
 
+#: changedetectionio/validate_url.py
+msgid "API Base URL is not a valid http(s) URL."
+msgstr ""
+
+#: changedetectionio/validate_url.py
+msgid ""
+"API Base URL resolves to a private, loopback, link-local or reserved IP address and was blocked to prevent SSRF. To "
+"allow LLM endpoints on private networks (e.g. a local Ollama server) set the environment variable "
+"ALLOW_IANA_RESTRICTED_ADDRESSES=true and restart."
+msgstr ""
+
 #: changedetectionio/widgets/ternary_boolean.py
 msgid "Yes"
 msgstr ""

changedetectionio/translations/es/LC_MESSAGES/messages.po

@@ -366,6 +366,12 @@ msgstr "Todas las notificaciones silenciadas."
 msgid "All notifications unmuted."
 msgstr "Todas las notificaciones activadas."
 
+#: changedetectionio/blueprint/settings/llm.py
+msgid ""
+"api_key is required when api_base differs from the saved configuration. Refusing to send the stored API key to a "
+"different endpoint."
+msgstr ""
+
 #: changedetectionio/blueprint/settings/llm.py
 msgid "AI / LLM configuration removed."
 msgstr ""
@@ -872,6 +878,10 @@ msgstr ""
 msgid "AI Provider"
 msgstr ""
 
+#: changedetectionio/blueprint/settings/templates/settings_llm_tab.html
+msgid "Master switch — when off, all AI lookups are skipped even if a provider is configured below."
+msgstr ""
+
 #: changedetectionio/blueprint/settings/templates/settings_llm_tab.html
 msgid "Third-party data transfer — please read"
 msgstr ""
@@ -1122,6 +1132,12 @@ msgstr ""
 msgid "characters — currently enforcing: %(limit)s"
 msgstr ""
 
+#: changedetectionio/blueprint/settings/templates/settings_llm_tab.html
+msgid ""
+"tokens — skips AI evaluation on a watch once its usage within the current period (monthly) hits this cap (0 = "
+"unlimited)"
+msgstr ""
+
 #: changedetectionio/blueprint/settings/templates/settings_llm_tab.html
 msgid "No AI usage recorded yet."
 msgstr ""
@@ -2308,11 +2324,11 @@ msgstr "Último Comprobado"
 
 #: changedetectionio/blueprint/watchlist/templates/watch-overview.html
 msgid "Changed"
-msgstr "Cambiadp"
+msgstr "Cambiado"
 
 #: changedetectionio/blueprint/watchlist/templates/watch-overview.html
 msgid "Last Changed"
-msgstr "Último Cambiadp"
+msgstr "Último Cambiado"
 
 #: changedetectionio/blueprint/watchlist/templates/watch-overview.html
 msgid "No web page change detection watches configured, please add a URL in the box above, or"
@@ -3010,7 +3026,6 @@ msgstr "Coincide con todo lo siguiente"
 msgid "Match any of the following"
 msgstr "Coincide con cualquiera de los siguientes"
 
-#. dennis-ignore: W303 - False positive caused by <title>. https://github.com/mozilla/dennis/issues/213
 #: changedetectionio/forms.py
 msgid "Use page <title> in list"
 msgstr "Usar página <title> en la lista"
@@ -3110,7 +3125,6 @@ msgstr "Actualizaciones de UI en tiempo real habilitadas"
 msgid "Favicons Enabled"
 msgstr "Favicones habilitados"
 
-#. dennis-ignore: W303 - False positive caused by <title>. https://github.com/mozilla/dennis/issues/213
 #: changedetectionio/forms.py
 msgid "Use page <title> in watch overview list"
 msgstr "Usar <title> de la página en la lista general de monitores"
@@ -3223,12 +3237,8 @@ msgstr ""
 msgid "Default AI Change Summary prompt"
 msgstr ""
 
-#: changedetectionio/forms.py
-msgid "Max tokens per check"
-msgstr ""
-
-#: changedetectionio/forms.py
-msgid "Max cumulative tokens (per watch)"
+#: changedetectionio/blueprint/settings/templates/settings_llm_tab.html changedetectionio/forms.py
+msgid "Max tokens per watch per period"
 msgstr ""
 
 #: changedetectionio/forms.py
@@ -3239,6 +3249,10 @@ msgstr ""
 msgid "Max input characters"
 msgstr ""
 
+#: changedetectionio/forms.py
+msgid "Enable AI / LLM features"
+msgstr ""
+
 #: changedetectionio/forms.py
 msgid "Replace {{diff}} notification token with AI summary"
 msgstr ""
@@ -3504,7 +3518,6 @@ msgstr "La URL que se está viendo."
 msgid "The UUID of the watch."
 msgstr "El UUID del monitor."
 
-#. dennis-ignore: W303 - False positive caused by <title>. https://github.com/mozilla/dennis/issues/213
 #: changedetectionio/templates/_common_fields.html
 msgid "The page title of the watch, uses <title> if not set, falls back to URL"
 msgstr "El título de la página del monitor, utiliza <title> si no se establece, vuelve a la URL"
@@ -4245,6 +4258,17 @@ msgstr "Cambiar idioma"
 msgid "Change language"
 msgstr "Cambiar idioma"
 
+#: changedetectionio/validate_url.py
+msgid "API Base URL is not a valid http(s) URL."
+msgstr ""
+
+#: changedetectionio/validate_url.py
+msgid ""
+"API Base URL resolves to a private, loopback, link-local or reserved IP address and was blocked to prevent SSRF. To "
+"allow LLM endpoints on private networks (e.g. a local Ollama server) set the environment variable "
+"ALLOW_IANA_RESTRICTED_ADDRESSES=true and restart."
+msgstr ""
+
 #: changedetectionio/widgets/ternary_boolean.py
 msgid "Yes"
 msgstr "Sí"

changedetectionio/translations/fr/LC_MESSAGES/messages.po

@@ -362,6 +362,12 @@ msgstr "Toutes les notifications sont désactivées."
 msgid "All notifications unmuted."
 msgstr "Toutes les notifications sont activées."
 
+#: changedetectionio/blueprint/settings/llm.py
+msgid ""
+"api_key is required when api_base differs from the saved configuration. Refusing to send the stored API key to a "
+"different endpoint."
+msgstr ""
+
 #: changedetectionio/blueprint/settings/llm.py
 msgid "AI / LLM configuration removed."
 msgstr ""
@@ -840,6 +846,10 @@ msgstr ""
 msgid "AI Provider"
 msgstr ""
 
+#: changedetectionio/blueprint/settings/templates/settings_llm_tab.html
+msgid "Master switch — when off, all AI lookups are skipped even if a provider is configured below."
+msgstr ""
+
 #: changedetectionio/blueprint/settings/templates/settings_llm_tab.html
 msgid "Third-party data transfer — please read"
 msgstr ""
@@ -1090,6 +1100,12 @@ msgstr ""
 msgid "characters — currently enforcing: %(limit)s"
 msgstr ""
 
+#: changedetectionio/blueprint/settings/templates/settings_llm_tab.html
+msgid ""
+"tokens — skips AI evaluation on a watch once its usage within the current period (monthly) hits this cap (0 = "
+"unlimited)"
+msgstr ""
+
 #: changedetectionio/blueprint/settings/templates/settings_llm_tab.html
 msgid "No AI usage recorded yet."
 msgstr ""
@@ -2950,7 +2966,6 @@ msgstr "Faites correspondre tous les éléments suivants"
 msgid "Match any of the following"
 msgstr "Faites correspondre l'un des éléments suivants"
 
-#. dennis-ignore: W303 - False positive caused by <title>. https://github.com/mozilla/dennis/issues/213
 #: changedetectionio/forms.py
 msgid "Use page <title> in list"
 msgstr "Utiliser la page <title> dans la liste"
@@ -3050,7 +3065,6 @@ msgstr "Mises à jour en temps réel hors ligne"
 msgid "Favicons Enabled"
 msgstr "Favicons Activés"
 
-#. dennis-ignore: W303 - False positive caused by <title>. https://github.com/mozilla/dennis/issues/213
 #: changedetectionio/forms.py
 msgid "Use page <title> in watch overview list"
 msgstr "Utiliser la page <title> dans la liste de présentation des moniteurs"
@@ -3163,12 +3177,8 @@ msgstr ""
 msgid "Default AI Change Summary prompt"
 msgstr ""
 
-#: changedetectionio/forms.py
-msgid "Max tokens per check"
-msgstr ""
-
-#: changedetectionio/forms.py
-msgid "Max cumulative tokens (per watch)"
+#: changedetectionio/blueprint/settings/templates/settings_llm_tab.html changedetectionio/forms.py
+msgid "Max tokens per watch per period"
 msgstr ""
 
 #: changedetectionio/forms.py
@@ -3179,6 +3189,10 @@ msgstr ""
 msgid "Max input characters"
 msgstr ""
 
+#: changedetectionio/forms.py
+msgid "Enable AI / LLM features"
+msgstr ""
+
 #: changedetectionio/forms.py
 msgid "Replace {{diff}} notification token with AI summary"
 msgstr ""
@@ -3444,7 +3458,6 @@ msgstr ""
 msgid "The UUID of the watch."
 msgstr "L'UUID du moniteur."
 
-#. dennis-ignore: W303 - False positive caused by <title>. https://github.com/mozilla/dennis/issues/213
 #: changedetectionio/templates/_common_fields.html
 msgid "The page title of the watch, uses <title> if not set, falls back to URL"
 msgstr ""
@@ -4174,6 +4187,17 @@ msgstr "Changer de langue"
 msgid "Change language"
 msgstr "Changer de langue"
 
+#: changedetectionio/validate_url.py
+msgid "API Base URL is not a valid http(s) URL."
+msgstr ""
+
+#: changedetectionio/validate_url.py
+msgid ""
+"API Base URL resolves to a private, loopback, link-local or reserved IP address and was blocked to prevent SSRF. To "
+"allow LLM endpoints on private networks (e.g. a local Ollama server) set the environment variable "
+"ALLOW_IANA_RESTRICTED_ADDRESSES=true and restart."
+msgstr ""
+
 #: changedetectionio/widgets/ternary_boolean.py
 msgid "Yes"
 msgstr "Oui"

changedetectionio/translations/it/LC_MESSAGES/messages.po

@@ -360,6 +360,12 @@ msgstr "Tutte le notifiche disattivate."
 msgid "All notifications unmuted."
 msgstr "Tutte le notifiche attivate."
 
+#: changedetectionio/blueprint/settings/llm.py
+msgid ""
+"api_key is required when api_base differs from the saved configuration. Refusing to send the stored API key to a "
+"different endpoint."
+msgstr ""
+
 #: changedetectionio/blueprint/settings/llm.py
 msgid "AI / LLM configuration removed."
 msgstr ""
@@ -836,6 +842,10 @@ msgstr ""
 msgid "AI Provider"
 msgstr ""
 
+#: changedetectionio/blueprint/settings/templates/settings_llm_tab.html
+msgid "Master switch — when off, all AI lookups are skipped even if a provider is configured below."
+msgstr ""
+
 #: changedetectionio/blueprint/settings/templates/settings_llm_tab.html
 msgid "Third-party data transfer — please read"
 msgstr ""
@@ -1086,6 +1096,12 @@ msgstr ""
 msgid "characters — currently enforcing: %(limit)s"
 msgstr ""
 
+#: changedetectionio/blueprint/settings/templates/settings_llm_tab.html
+msgid ""
+"tokens — skips AI evaluation on a watch once its usage within the current period (monthly) hits this cap (0 = "
+"unlimited)"
+msgstr ""
+
 #: changedetectionio/blueprint/settings/templates/settings_llm_tab.html
 msgid "No AI usage recorded yet."
 msgstr ""
@@ -2939,7 +2955,6 @@ msgstr "Corrisponde a tutti i seguenti"
 msgid "Match any of the following"
 msgstr "Corrisponde a uno qualsiasi dei seguenti"
 
-#. dennis-ignore: W303 - False positive caused by <title>. https://github.com/mozilla/dennis/issues/213
 #: changedetectionio/forms.py
 msgid "Use page <title> in list"
 msgstr "Usa <title> pagina nell'elenco"
@@ -3039,7 +3054,6 @@ msgstr "Aggiornamenti UI in tempo reale attivi"
 msgid "Favicons Enabled"
 msgstr "Favicon attive"
 
-#. dennis-ignore: W303 - False positive caused by <title>. https://github.com/mozilla/dennis/issues/213
 #: changedetectionio/forms.py
 msgid "Use page <title> in watch overview list"
 msgstr "Usa <title> pagina nell'elenco osservati"
@@ -3152,12 +3166,8 @@ msgstr ""
 msgid "Default AI Change Summary prompt"
 msgstr ""
 
-#: changedetectionio/forms.py
-msgid "Max tokens per check"
-msgstr ""
-
-#: changedetectionio/forms.py
-msgid "Max cumulative tokens (per watch)"
+#: changedetectionio/blueprint/settings/templates/settings_llm_tab.html changedetectionio/forms.py
+msgid "Max tokens per watch per period"
 msgstr ""
 
 #: changedetectionio/forms.py
@@ -3168,6 +3178,10 @@ msgstr ""
 msgid "Max input characters"
 msgstr ""
 
+#: changedetectionio/forms.py
+msgid "Enable AI / LLM features"
+msgstr ""
+
 #: changedetectionio/forms.py
 msgid "Replace {{diff}} notification token with AI summary"
 msgstr ""
@@ -3433,7 +3447,6 @@ msgstr ""
 msgid "The UUID of the watch."
 msgstr "L'UUID del monitor."
 
-#. dennis-ignore: W303 - False positive caused by <title>. https://github.com/mozilla/dennis/issues/213
 #: changedetectionio/templates/_common_fields.html
 msgid "The page title of the watch, uses <title> if not set, falls back to URL"
 msgstr ""
@@ -4161,6 +4174,17 @@ msgstr "Cambia Lingua"
 msgid "Change language"
 msgstr "Cambia lingua"
 
+#: changedetectionio/validate_url.py
+msgid "API Base URL is not a valid http(s) URL."
+msgstr ""
+
+#: changedetectionio/validate_url.py
+msgid ""
+"API Base URL resolves to a private, loopback, link-local or reserved IP address and was blocked to prevent SSRF. To "
+"allow LLM endpoints on private networks (e.g. a local Ollama server) set the environment variable "
+"ALLOW_IANA_RESTRICTED_ADDRESSES=true and restart."
+msgstr ""
+
 #: changedetectionio/widgets/ternary_boolean.py
 msgid "Yes"
 msgstr "Sì"

changedetectionio/translations/ja/LC_MESSAGES/messages.po

@@ -362,6 +362,12 @@ msgstr "すべての通知をミュートしました。"
 msgid "All notifications unmuted."
 msgstr "すべての通知のミュートを解除しました。"
 
+#: changedetectionio/blueprint/settings/llm.py
+msgid ""
+"api_key is required when api_base differs from the saved configuration. Refusing to send the stored API key to a "
+"different endpoint."
+msgstr ""
+
 #: changedetectionio/blueprint/settings/llm.py
 msgid "AI / LLM configuration removed."
 msgstr ""
@@ -841,6 +847,10 @@ msgstr ""
 msgid "AI Provider"
 msgstr ""
 
+#: changedetectionio/blueprint/settings/templates/settings_llm_tab.html
+msgid "Master switch — when off, all AI lookups are skipped even if a provider is configured below."
+msgstr ""
+
 #: changedetectionio/blueprint/settings/templates/settings_llm_tab.html
 msgid "Third-party data transfer — please read"
 msgstr ""
@@ -1091,6 +1101,12 @@ msgstr ""
 msgid "characters — currently enforcing: %(limit)s"
 msgstr ""
 
+#: changedetectionio/blueprint/settings/templates/settings_llm_tab.html
+msgid ""
+"tokens — skips AI evaluation on a watch once its usage within the current period (monthly) hits this cap (0 = "
+"unlimited)"
+msgstr ""
+
 #: changedetectionio/blueprint/settings/templates/settings_llm_tab.html
 msgid "No AI usage recorded yet."
 msgstr ""
@@ -2956,7 +2972,6 @@ msgstr "以下のすべてに一致"
 msgid "Match any of the following"
 msgstr "以下のいずれかに一致"
 
-#. dennis-ignore: W303 - False positive caused by <title>. https://github.com/mozilla/dennis/issues/213
 #: changedetectionio/forms.py
 msgid "Use page <title> in list"
 msgstr "リストでページの <title> を使用"
@@ -3056,7 +3071,6 @@ msgstr "リアルタイムUI更新を有効化"
 msgid "Favicons Enabled"
 msgstr "ファビコンを有効化"
 
-#. dennis-ignore: W303 - False positive caused by <title>. https://github.com/mozilla/dennis/issues/213
 #: changedetectionio/forms.py
 msgid "Use page <title> in watch overview list"
 msgstr "ウォッチ一覧リストでページの <title> を使用"
@@ -3169,12 +3183,8 @@ msgstr ""
 msgid "Default AI Change Summary prompt"
 msgstr ""
 
-#: changedetectionio/forms.py
-msgid "Max tokens per check"
-msgstr ""
-
-#: changedetectionio/forms.py
-msgid "Max cumulative tokens (per watch)"
+#: changedetectionio/blueprint/settings/templates/settings_llm_tab.html changedetectionio/forms.py
+msgid "Max tokens per watch per period"
 msgstr ""
 
 #: changedetectionio/forms.py
@@ -3185,6 +3195,10 @@ msgstr ""
 msgid "Max input characters"
 msgstr ""
 
+#: changedetectionio/forms.py
+msgid "Enable AI / LLM features"
+msgstr ""
+
 #: changedetectionio/forms.py
 msgid "Replace {{diff}} notification token with AI summary"
 msgstr ""
@@ -3450,7 +3464,6 @@ msgstr "監視中のURL。"
 msgid "The UUID of the watch."
 msgstr "ウォッチのUUID。"
 
-#. dennis-ignore: W303 - False positive caused by <title>. https://github.com/mozilla/dennis/issues/213
 #: changedetectionio/templates/_common_fields.html
 msgid "The page title of the watch, uses <title> if not set, falls back to URL"
 msgstr "ウォッチのページタイトル。設定されていない場合は <title> を使用し、それもなければURLにフォールバックします。"
@@ -4202,6 +4215,17 @@ msgstr "言語の変更"
 msgid "Change language"
 msgstr "言語を変更"
 
+#: changedetectionio/validate_url.py
+msgid "API Base URL is not a valid http(s) URL."
+msgstr ""
+
+#: changedetectionio/validate_url.py
+msgid ""
+"API Base URL resolves to a private, loopback, link-local or reserved IP address and was blocked to prevent SSRF. To "
+"allow LLM endpoints on private networks (e.g. a local Ollama server) set the environment variable "
+"ALLOW_IANA_RESTRICTED_ADDRESSES=true and restart."
+msgstr ""
+
 #: changedetectionio/widgets/ternary_boolean.py
 msgid "Yes"
 msgstr "はい"

changedetectionio/translations/ko/LC_MESSAGES/messages.po

@@ -360,6 +360,12 @@ msgstr "모든 알림이 음소거되었습니다."
 msgid "All notifications unmuted."
 msgstr "모든 알림의 음소거가 해제되었습니다."
 
+#: changedetectionio/blueprint/settings/llm.py
+msgid ""
+"api_key is required when api_base differs from the saved configuration. Refusing to send the stored API key to a "
+"different endpoint."
+msgstr ""
+
 #: changedetectionio/blueprint/settings/llm.py
 msgid "AI / LLM configuration removed."
 msgstr "AI / LLM 설정이 제거되었습니다."
@@ -838,6 +844,10 @@ msgstr "AI 프로바이더 설정"
 msgid "AI Provider"
 msgstr "AI 프로바이더"
 
+#: changedetectionio/blueprint/settings/templates/settings_llm_tab.html
+msgid "Master switch — when off, all AI lookups are skipped even if a provider is configured below."
+msgstr ""
+
 #: changedetectionio/blueprint/settings/templates/settings_llm_tab.html
 msgid "Third-party data transfer — please read"
 msgstr "제3자 데이터 전송 - 읽어 주세요"
@@ -1092,6 +1102,12 @@ msgstr "(<code>LLM_MAX_INPUT_CHARS</code>로 설정됨)"
 msgid "characters — currently enforcing: %(limit)s"
 msgstr "문자 - 현재 적용 중: %(limit)s"
 
+#: changedetectionio/blueprint/settings/templates/settings_llm_tab.html
+msgid ""
+"tokens — skips AI evaluation on a watch once its usage within the current period (monthly) hits this cap (0 = "
+"unlimited)"
+msgstr ""
+
 #: changedetectionio/blueprint/settings/templates/settings_llm_tab.html
 msgid "No AI usage recorded yet."
 msgstr "아직 기록된 AI 사용량이 없습니다."
@@ -2947,7 +2963,6 @@ msgstr "다음 모두와 일치"
 msgid "Match any of the following"
 msgstr "다음 중 하나와 일치"
 
-#. dennis-ignore: W303 - False positive caused by <title>. https://github.com/mozilla/dennis/issues/213
 #: changedetectionio/forms.py
 msgid "Use page <title> in list"
 msgstr "목록에 페이지 <title> 사용"
@@ -3047,7 +3062,6 @@ msgstr "실시간 UI 업데이트 활성화"
 msgid "Favicons Enabled"
 msgstr "파비콘 활성화"
 
-#. dennis-ignore: W303 - False positive caused by <title>. https://github.com/mozilla/dennis/issues/213
 #: changedetectionio/forms.py
 msgid "Use page <title> in watch overview list"
 msgstr "모니터링 목록에 페이지 <title> 사용"
@@ -3160,13 +3174,9 @@ msgstr ""
 msgid "Default AI Change Summary prompt"
 msgstr "기본 AI 변경 요약 프롬프트"
 
-#: changedetectionio/forms.py
-msgid "Max tokens per check"
-msgstr "확인당 최대 토큰 수"
-
-#: changedetectionio/forms.py
-msgid "Max cumulative tokens (per watch)"
-msgstr "최대 누적 토큰 수 (모니터링별)"
+#: changedetectionio/blueprint/settings/templates/settings_llm_tab.html changedetectionio/forms.py
+msgid "Max tokens per watch per period"
+msgstr ""
 
 #: changedetectionio/forms.py
 msgid "Monthly token budget"
@@ -3176,6 +3186,10 @@ msgstr "월간 토큰 예산"
 msgid "Max input characters"
 msgstr "최대 입력 문자 수"
 
+#: changedetectionio/forms.py
+msgid "Enable AI / LLM features"
+msgstr ""
+
 #: changedetectionio/forms.py
 msgid "Replace {{diff}} notification token with AI summary"
 msgstr "{{diff}} 알림 토큰을 AI 요약으로 대체"
@@ -3441,7 +3455,6 @@ msgstr "모니터링 중인 URL입니다."
 msgid "The UUID of the watch."
 msgstr "모니터링 UUID입니다."
 
-#. dennis-ignore: W303 - False positive caused by <title>. https://github.com/mozilla/dennis/issues/213
 #: changedetectionio/templates/_common_fields.html
 msgid "The page title of the watch, uses <title> if not set, falls back to URL"
 msgstr "모니터링의 페이지 제목입니다. 설정되지 않았으면 <title> 을 사용하고, 없으면 URL을 사용합니다."
@@ -4179,6 +4192,17 @@ msgstr "언어 변경"
 msgid "Change language"
 msgstr "언어 변경"
 
+#: changedetectionio/validate_url.py
+msgid "API Base URL is not a valid http(s) URL."
+msgstr ""
+
+#: changedetectionio/validate_url.py
+msgid ""
+"API Base URL resolves to a private, loopback, link-local or reserved IP address and was blocked to prevent SSRF. To "
+"allow LLM endpoints on private networks (e.g. a local Ollama server) set the environment variable "
+"ALLOW_IANA_RESTRICTED_ADDRESSES=true and restart."
+msgstr ""
+
 #: changedetectionio/widgets/ternary_boolean.py
 msgid "Yes"
 msgstr "예"

changedetectionio/translations/messages.pot

@@ -6,9 +6,9 @@
 #, fuzzy
 msgid ""
 msgstr ""
-"Project-Id-Version: changedetection.io 0.55.3\n"
+"Project-Id-Version: changedetection.io 0.55.6\n"
 "Report-Msgid-Bugs-To: EMAIL@ADDRESS\n"
-"POT-Creation-Date: 2026-05-15 18:31+0200\n"
+"POT-Creation-Date: 2026-05-25 17:59+0200\n"
 "PO-Revision-Date: YEAR-MO-DA HO:MI+ZONE\n"
 "Last-Translator: FULL NAME <EMAIL@ADDRESS>\n"
 "Language-Team: LANGUAGE <LL@li.org>\n"
@@ -357,6 +357,12 @@ msgstr ""
 msgid "All notifications unmuted."
 msgstr ""
 
+#: changedetectionio/blueprint/settings/llm.py
+msgid ""
+"api_key is required when api_base differs from the saved configuration. Refusing to send the stored API key to a "
+"different endpoint."
+msgstr ""
+
 #: changedetectionio/blueprint/settings/llm.py
 msgid "AI / LLM configuration removed."
 msgstr ""
@@ -833,6 +839,10 @@ msgstr ""
 msgid "AI Provider"
 msgstr ""
 
+#: changedetectionio/blueprint/settings/templates/settings_llm_tab.html
+msgid "Master switch — when off, all AI lookups are skipped even if a provider is configured below."
+msgstr ""
+
 #: changedetectionio/blueprint/settings/templates/settings_llm_tab.html
 msgid "Third-party data transfer — please read"
 msgstr ""
@@ -1083,6 +1093,12 @@ msgstr ""
 msgid "characters — currently enforcing: %(limit)s"
 msgstr ""
 
+#: changedetectionio/blueprint/settings/templates/settings_llm_tab.html
+msgid ""
+"tokens — skips AI evaluation on a watch once its usage within the current period (monthly) hits this cap (0 = "
+"unlimited)"
+msgstr ""
+
 #: changedetectionio/blueprint/settings/templates/settings_llm_tab.html
 msgid "No AI usage recorded yet."
 msgstr ""
@@ -2936,7 +2952,6 @@ msgstr ""
 msgid "Match any of the following"
 msgstr ""
 
-#. dennis-ignore: W303 - False positive caused by <title>. https://github.com/mozilla/dennis/issues/213
 #: changedetectionio/forms.py
 msgid "Use page <title> in list"
 msgstr ""
@@ -3036,7 +3051,6 @@ msgstr ""
 msgid "Favicons Enabled"
 msgstr ""
 
-#. dennis-ignore: W303 - False positive caused by <title>. https://github.com/mozilla/dennis/issues/213
 #: changedetectionio/forms.py
 msgid "Use page <title> in watch overview list"
 msgstr ""
@@ -3149,12 +3163,8 @@ msgstr ""
 msgid "Default AI Change Summary prompt"
 msgstr ""
 
-#: changedetectionio/forms.py
-msgid "Max tokens per check"
-msgstr ""
-
-#: changedetectionio/forms.py
-msgid "Max cumulative tokens (per watch)"
+#: changedetectionio/blueprint/settings/templates/settings_llm_tab.html changedetectionio/forms.py
+msgid "Max tokens per watch per period"
 msgstr ""
 
 #: changedetectionio/forms.py
@@ -3165,6 +3175,10 @@ msgstr ""
 msgid "Max input characters"
 msgstr ""
 
+#: changedetectionio/forms.py
+msgid "Enable AI / LLM features"
+msgstr ""
+
 #: changedetectionio/forms.py
 msgid "Replace {{diff}} notification token with AI summary"
 msgstr ""
@@ -3430,7 +3444,6 @@ msgstr ""
 msgid "The UUID of the watch."
 msgstr ""
 
-#. dennis-ignore: W303 - False positive caused by <title>. https://github.com/mozilla/dennis/issues/213
 #: changedetectionio/templates/_common_fields.html
 msgid "The page title of the watch, uses <title> if not set, falls back to URL"
 msgstr ""
@@ -4158,6 +4171,17 @@ msgstr ""
 msgid "Change language"
 msgstr ""
 
+#: changedetectionio/validate_url.py
+msgid "API Base URL is not a valid http(s) URL."
+msgstr ""
+
+#: changedetectionio/validate_url.py
+msgid ""
+"API Base URL resolves to a private, loopback, link-local or reserved IP address and was blocked to prevent SSRF. To "
+"allow LLM endpoints on private networks (e.g. a local Ollama server) set the environment variable "
+"ALLOW_IANA_RESTRICTED_ADDRESSES=true and restart."
+msgstr ""
+
 #: changedetectionio/widgets/ternary_boolean.py
 msgid "Yes"
 msgstr ""

changedetectionio/translations/pt_BR/LC_MESSAGES/messages.po

@@ -363,6 +363,12 @@ msgstr "Todas as notificações silenciadas."
 msgid "All notifications unmuted."
 msgstr "Todas as notificações reativadas."
 
+#: changedetectionio/blueprint/settings/llm.py
+msgid ""
+"api_key is required when api_base differs from the saved configuration. Refusing to send the stored API key to a "
+"different endpoint."
+msgstr ""
+
 #: changedetectionio/blueprint/settings/llm.py
 msgid "AI / LLM configuration removed."
 msgstr ""
@@ -859,6 +865,10 @@ msgstr ""
 msgid "AI Provider"
 msgstr ""
 
+#: changedetectionio/blueprint/settings/templates/settings_llm_tab.html
+msgid "Master switch — when off, all AI lookups are skipped even if a provider is configured below."
+msgstr ""
+
 #: changedetectionio/blueprint/settings/templates/settings_llm_tab.html
 msgid "Third-party data transfer — please read"
 msgstr ""
@@ -1109,6 +1119,12 @@ msgstr ""
 msgid "characters — currently enforcing: %(limit)s"
 msgstr ""
 
+#: changedetectionio/blueprint/settings/templates/settings_llm_tab.html
+msgid ""
+"tokens — skips AI evaluation on a watch once its usage within the current period (monthly) hits this cap (0 = "
+"unlimited)"
+msgstr ""
+
 #: changedetectionio/blueprint/settings/templates/settings_llm_tab.html
 msgid "No AI usage recorded yet."
 msgstr ""
@@ -2987,7 +3003,6 @@ msgstr "Corresponder a TODOS os seguintes"
 msgid "Match any of the following"
 msgstr "Corresponder a QUALQUER um dos seguintes"
 
-#. dennis-ignore: W303 - False positive caused by <title>. https://github.com/mozilla/dennis/issues/213
 #: changedetectionio/forms.py
 msgid "Use page <title> in list"
 msgstr "Usar <title> da página na lista"
@@ -3087,7 +3102,6 @@ msgstr "Atualizações de Interface em Tempo Real Ativadas"
 msgid "Favicons Enabled"
 msgstr "Favicons Ativados"
 
-#. dennis-ignore: W303 - False positive caused by <title>. https://github.com/mozilla/dennis/issues/213
 #: changedetectionio/forms.py
 msgid "Use page <title> in watch overview list"
 msgstr "Usar <title> da página na lista de visão geral"
@@ -3200,12 +3214,8 @@ msgstr ""
 msgid "Default AI Change Summary prompt"
 msgstr ""
 
-#: changedetectionio/forms.py
-msgid "Max tokens per check"
-msgstr ""
-
-#: changedetectionio/forms.py
-msgid "Max cumulative tokens (per watch)"
+#: changedetectionio/blueprint/settings/templates/settings_llm_tab.html changedetectionio/forms.py
+msgid "Max tokens per watch per period"
 msgstr ""
 
 #: changedetectionio/forms.py
@@ -3216,6 +3226,10 @@ msgstr ""
 msgid "Max input characters"
 msgstr ""
 
+#: changedetectionio/forms.py
+msgid "Enable AI / LLM features"
+msgstr ""
+
 #: changedetectionio/forms.py
 msgid "Replace {{diff}} notification token with AI summary"
 msgstr ""
@@ -3481,7 +3495,6 @@ msgstr "A URL que está sendo monitorada."
 msgid "The UUID of the watch."
 msgstr "O UUID do monitoramento."
 
-#. dennis-ignore: W303 - False positive caused by <title>. https://github.com/mozilla/dennis/issues/213
 #: changedetectionio/templates/_common_fields.html
 msgid "The page title of the watch, uses <title> if not set, falls back to URL"
 msgstr "O título da página do monitoramento, usa <title> se não definido, ou a URL"
@@ -4217,6 +4230,17 @@ msgstr "Mudar Idioma"
 msgid "Change language"
 msgstr "Mudar idioma"
 
+#: changedetectionio/validate_url.py
+msgid "API Base URL is not a valid http(s) URL."
+msgstr ""
+
+#: changedetectionio/validate_url.py
+msgid ""
+"API Base URL resolves to a private, loopback, link-local or reserved IP address and was blocked to prevent SSRF. To "
+"allow LLM endpoints on private networks (e.g. a local Ollama server) set the environment variable "
+"ALLOW_IANA_RESTRICTED_ADDRESSES=true and restart."
+msgstr ""
+
 #: changedetectionio/widgets/ternary_boolean.py
 msgid "Yes"
 msgstr "Sim"

changedetectionio/translations/tr/LC_MESSAGES/messages.po

@@ -367,6 +367,12 @@ msgstr "Tüm bildirimler sessize alındı."
 msgid "All notifications unmuted."
 msgstr "Tüm bildirimlerin sesi açıldı."
 
+#: changedetectionio/blueprint/settings/llm.py
+msgid ""
+"api_key is required when api_base differs from the saved configuration. Refusing to send the stored API key to a "
+"different endpoint."
+msgstr ""
+
 #: changedetectionio/blueprint/settings/llm.py
 msgid "AI / LLM configuration removed."
 msgstr ""
@@ -869,6 +875,10 @@ msgstr ""
 msgid "AI Provider"
 msgstr ""
 
+#: changedetectionio/blueprint/settings/templates/settings_llm_tab.html
+msgid "Master switch — when off, all AI lookups are skipped even if a provider is configured below."
+msgstr ""
+
 #: changedetectionio/blueprint/settings/templates/settings_llm_tab.html
 msgid "Third-party data transfer — please read"
 msgstr ""
@@ -1119,6 +1129,12 @@ msgstr ""
 msgid "characters — currently enforcing: %(limit)s"
 msgstr ""
 
+#: changedetectionio/blueprint/settings/templates/settings_llm_tab.html
+msgid ""
+"tokens — skips AI evaluation on a watch once its usage within the current period (monthly) hits this cap (0 = "
+"unlimited)"
+msgstr ""
+
 #: changedetectionio/blueprint/settings/templates/settings_llm_tab.html
 msgid "No AI usage recorded yet."
 msgstr ""
@@ -2990,7 +3006,6 @@ msgstr "Aşağıdakilerin tümünü eşleştir"
 msgid "Match any of the following"
 msgstr "Aşağıdakilerden herhangi birini eşleştir"
 
-#. dennis-ignore: W303 - False positive caused by <title>. https://github.com/mozilla/dennis/issues/213
 #: changedetectionio/forms.py
 msgid "Use page <title> in list"
 msgstr "Listede sayfa <title>'ını kullan"
@@ -3090,7 +3105,6 @@ msgstr "Gerçek Zamanlı Arayüz Güncellemeleri Etkin"
 msgid "Favicons Enabled"
 msgstr "Favicon'lar Etkin"
 
-#. dennis-ignore: W303 - False positive caused by <title>. https://github.com/mozilla/dennis/issues/213
 #: changedetectionio/forms.py
 msgid "Use page <title> in watch overview list"
 msgstr "İzleyici genel bakış listesinde sayfa <title>'ını kullan"
@@ -3203,12 +3217,8 @@ msgstr ""
 msgid "Default AI Change Summary prompt"
 msgstr ""
 
-#: changedetectionio/forms.py
-msgid "Max tokens per check"
-msgstr ""
-
-#: changedetectionio/forms.py
-msgid "Max cumulative tokens (per watch)"
+#: changedetectionio/blueprint/settings/templates/settings_llm_tab.html changedetectionio/forms.py
+msgid "Max tokens per watch per period"
 msgstr ""
 
 #: changedetectionio/forms.py
@@ -3219,6 +3229,10 @@ msgstr ""
 msgid "Max input characters"
 msgstr ""
 
+#: changedetectionio/forms.py
+msgid "Enable AI / LLM features"
+msgstr ""
+
 #: changedetectionio/forms.py
 msgid "Replace {{diff}} notification token with AI summary"
 msgstr ""
@@ -3484,7 +3498,6 @@ msgstr "İzlenen URL."
 msgid "The UUID of the watch."
 msgstr "İzleyicinin UUID'si."
 
-#. dennis-ignore: W303 - False positive caused by <title>. https://github.com/mozilla/dennis/issues/213
 #: changedetectionio/templates/_common_fields.html
 msgid "The page title of the watch, uses <title> if not set, falls back to URL"
 msgstr "İzleyicinin sayfa başlığı, ayarlanmamışsa <title> kullanır, URL'ye geri döner"
@@ -4220,6 +4233,17 @@ msgstr "Dili Değiştir"
 msgid "Change language"
 msgstr "Dili değiştir"
 
+#: changedetectionio/validate_url.py
+msgid "API Base URL is not a valid http(s) URL."
+msgstr ""
+
+#: changedetectionio/validate_url.py
+msgid ""
+"API Base URL resolves to a private, loopback, link-local or reserved IP address and was blocked to prevent SSRF. To "
+"allow LLM endpoints on private networks (e.g. a local Ollama server) set the environment variable "
+"ALLOW_IANA_RESTRICTED_ADDRESSES=true and restart."
+msgstr ""
+
 #: changedetectionio/widgets/ternary_boolean.py
 msgid "Yes"
 msgstr "Evet"

changedetectionio/translations/uk/LC_MESSAGES/messages.po

@@ -361,6 +361,12 @@ msgstr "Усі сповіщення вимкнено."
 msgid "All notifications unmuted."
 msgstr "Усі сповіщення увімкнено."
 
+#: changedetectionio/blueprint/settings/llm.py
+msgid ""
+"api_key is required when api_base differs from the saved configuration. Refusing to send the stored API key to a "
+"different endpoint."
+msgstr ""
+
 #: changedetectionio/blueprint/settings/llm.py
 msgid "AI / LLM configuration removed."
 msgstr ""
@@ -849,6 +855,10 @@ msgstr ""
 msgid "AI Provider"
 msgstr ""
 
+#: changedetectionio/blueprint/settings/templates/settings_llm_tab.html
+msgid "Master switch — when off, all AI lookups are skipped even if a provider is configured below."
+msgstr ""
+
 #: changedetectionio/blueprint/settings/templates/settings_llm_tab.html
 msgid "Third-party data transfer — please read"
 msgstr ""
@@ -1099,6 +1109,12 @@ msgstr ""
 msgid "characters — currently enforcing: %(limit)s"
 msgstr ""
 
+#: changedetectionio/blueprint/settings/templates/settings_llm_tab.html
+msgid ""
+"tokens — skips AI evaluation on a watch once its usage within the current period (monthly) hits this cap (0 = "
+"unlimited)"
+msgstr ""
+
 #: changedetectionio/blueprint/settings/templates/settings_llm_tab.html
 msgid "No AI usage recorded yet."
 msgstr ""
@@ -2969,7 +2985,6 @@ msgstr "Збіг усіх наступних умов"
 msgid "Match any of the following"
 msgstr "Збіг будь-якої з наступних умов"
 
-#. dennis-ignore: W303 - False positive caused by <title>. https://github.com/mozilla/dennis/issues/213
 #: changedetectionio/forms.py
 msgid "Use page <title> in list"
 msgstr "Використовувати <title> сторінки у списку"
@@ -3069,7 +3084,6 @@ msgstr "Оновлення UI в реальному часі увімкнено"
 msgid "Favicons Enabled"
 msgstr "Фавіконки увімкнено"
 
-#. dennis-ignore: W303 - False positive caused by <title>. https://github.com/mozilla/dennis/issues/213
 #: changedetectionio/forms.py
 msgid "Use page <title> in watch overview list"
 msgstr "Використовувати <title> сторінки у списку огляду завдань"
@@ -3182,12 +3196,8 @@ msgstr ""
 msgid "Default AI Change Summary prompt"
 msgstr ""
 
-#: changedetectionio/forms.py
-msgid "Max tokens per check"
-msgstr ""
-
-#: changedetectionio/forms.py
-msgid "Max cumulative tokens (per watch)"
+#: changedetectionio/blueprint/settings/templates/settings_llm_tab.html changedetectionio/forms.py
+msgid "Max tokens per watch per period"
 msgstr ""
 
 #: changedetectionio/forms.py
@@ -3198,6 +3208,10 @@ msgstr ""
 msgid "Max input characters"
 msgstr ""
 
+#: changedetectionio/forms.py
+msgid "Enable AI / LLM features"
+msgstr ""
+
 #: changedetectionio/forms.py
 msgid "Replace {{diff}} notification token with AI summary"
 msgstr ""
@@ -3463,7 +3477,6 @@ msgstr "URL, за яким ведеться спостереження."
 msgid "The UUID of the watch."
 msgstr "UUID завдання."
 
-#. dennis-ignore: W303 - False positive caused by <title>. https://github.com/mozilla/dennis/issues/213
 #: changedetectionio/templates/_common_fields.html
 msgid "The page title of the watch, uses <title> if not set, falls back to URL"
 msgstr "Заголовок сторінки завдання, використовує <title>, якщо не задано - URL"
@@ -4197,6 +4210,17 @@ msgstr "Змінити мову"
 msgid "Change language"
 msgstr "Змінити мову"
 
+#: changedetectionio/validate_url.py
+msgid "API Base URL is not a valid http(s) URL."
+msgstr ""
+
+#: changedetectionio/validate_url.py
+msgid ""
+"API Base URL resolves to a private, loopback, link-local or reserved IP address and was blocked to prevent SSRF. To "
+"allow LLM endpoints on private networks (e.g. a local Ollama server) set the environment variable "
+"ALLOW_IANA_RESTRICTED_ADDRESSES=true and restart."
+msgstr ""
+
 #: changedetectionio/widgets/ternary_boolean.py
 msgid "Yes"
 msgstr "Так"

changedetectionio/translations/zh/LC_MESSAGES/messages.po

@@ -362,6 +362,12 @@ msgstr "所有通知已静音。"
 msgid "All notifications unmuted."
 msgstr "所有通知已取消静音。"
 
+#: changedetectionio/blueprint/settings/llm.py
+msgid ""
+"api_key is required when api_base differs from the saved configuration. Refusing to send the stored API key to a "
+"different endpoint."
+msgstr ""
+
 #: changedetectionio/blueprint/settings/llm.py
 msgid "AI / LLM configuration removed."
 msgstr ""
@@ -838,6 +844,10 @@ msgstr ""
 msgid "AI Provider"
 msgstr ""
 
+#: changedetectionio/blueprint/settings/templates/settings_llm_tab.html
+msgid "Master switch — when off, all AI lookups are skipped even if a provider is configured below."
+msgstr ""
+
 #: changedetectionio/blueprint/settings/templates/settings_llm_tab.html
 msgid "Third-party data transfer — please read"
 msgstr ""
@@ -1088,6 +1098,12 @@ msgstr ""
 msgid "characters — currently enforcing: %(limit)s"
 msgstr ""
 
+#: changedetectionio/blueprint/settings/templates/settings_llm_tab.html
+msgid ""
+"tokens — skips AI evaluation on a watch once its usage within the current period (monthly) hits this cap (0 = "
+"unlimited)"
+msgstr ""
+
 #: changedetectionio/blueprint/settings/templates/settings_llm_tab.html
 msgid "No AI usage recorded yet."
 msgstr ""
@@ -2942,7 +2958,6 @@ msgstr "匹配以下全部"
 msgid "Match any of the following"
 msgstr "匹配以下任意"
 
-#. dennis-ignore: W303 - False positive caused by <title>. https://github.com/mozilla/dennis/issues/213
 #: changedetectionio/forms.py
 msgid "Use page <title> in list"
 msgstr "列表中使用页面 <title>"
@@ -3042,7 +3057,6 @@ msgstr "启用实时界面更新"
 msgid "Favicons Enabled"
 msgstr "启用站点图标"
 
-#. dennis-ignore: W303 - False positive caused by <title>. https://github.com/mozilla/dennis/issues/213
 #: changedetectionio/forms.py
 msgid "Use page <title> in watch overview list"
 msgstr "在监控概览列表中使用页面 <title>"
@@ -3155,12 +3169,8 @@ msgstr ""
 msgid "Default AI Change Summary prompt"
 msgstr ""
 
-#: changedetectionio/forms.py
-msgid "Max tokens per check"
-msgstr ""
-
-#: changedetectionio/forms.py
-msgid "Max cumulative tokens (per watch)"
+#: changedetectionio/blueprint/settings/templates/settings_llm_tab.html changedetectionio/forms.py
+msgid "Max tokens per watch per period"
 msgstr ""
 
 #: changedetectionio/forms.py
@@ -3171,6 +3181,10 @@ msgstr ""
 msgid "Max input characters"
 msgstr ""
 
+#: changedetectionio/forms.py
+msgid "Enable AI / LLM features"
+msgstr ""
+
 #: changedetectionio/forms.py
 msgid "Replace {{diff}} notification token with AI summary"
 msgstr ""
@@ -3436,7 +3450,6 @@ msgstr "被监控的 URL。"
 msgid "The UUID of the watch."
 msgstr "监视器的UUID。"
 
-#. dennis-ignore: W303 - False positive caused by <title>. https://github.com/mozilla/dennis/issues/213
 #: changedetectionio/templates/_common_fields.html
 msgid "The page title of the watch, uses <title> if not set, falls back to URL"
 msgstr "监控项的页面标题，未设置时使用 <title>，否则回退为 URL"
@@ -4165,6 +4178,17 @@ msgstr "切换语言"
 msgid "Change language"
 msgstr "切换语言"
 
+#: changedetectionio/validate_url.py
+msgid "API Base URL is not a valid http(s) URL."
+msgstr ""
+
+#: changedetectionio/validate_url.py
+msgid ""
+"API Base URL resolves to a private, loopback, link-local or reserved IP address and was blocked to prevent SSRF. To "
+"allow LLM endpoints on private networks (e.g. a local Ollama server) set the environment variable "
+"ALLOW_IANA_RESTRICTED_ADDRESSES=true and restart."
+msgstr ""
+
 #: changedetectionio/widgets/ternary_boolean.py
 msgid "Yes"
 msgstr "是"

changedetectionio/translations/zh_Hant_TW/LC_MESSAGES/messages.po

@@ -361,6 +361,12 @@ msgstr "所有通知已靜音。"
 msgid "All notifications unmuted."
 msgstr "所有通知已取消靜音。"
 
+#: changedetectionio/blueprint/settings/llm.py
+msgid ""
+"api_key is required when api_base differs from the saved configuration. Refusing to send the stored API key to a "
+"different endpoint."
+msgstr ""
+
 #: changedetectionio/blueprint/settings/llm.py
 msgid "AI / LLM configuration removed."
 msgstr ""
@@ -837,6 +843,10 @@ msgstr ""
 msgid "AI Provider"
 msgstr ""
 
+#: changedetectionio/blueprint/settings/templates/settings_llm_tab.html
+msgid "Master switch — when off, all AI lookups are skipped even if a provider is configured below."
+msgstr ""
+
 #: changedetectionio/blueprint/settings/templates/settings_llm_tab.html
 msgid "Third-party data transfer — please read"
 msgstr ""
@@ -1087,6 +1097,12 @@ msgstr ""
 msgid "characters — currently enforcing: %(limit)s"
 msgstr ""
 
+#: changedetectionio/blueprint/settings/templates/settings_llm_tab.html
+msgid ""
+"tokens — skips AI evaluation on a watch once its usage within the current period (monthly) hits this cap (0 = "
+"unlimited)"
+msgstr ""
+
 #: changedetectionio/blueprint/settings/templates/settings_llm_tab.html
 msgid "No AI usage recorded yet."
 msgstr ""
@@ -2941,7 +2957,6 @@ msgstr "符合以下所有條件"
 msgid "Match any of the following"
 msgstr "符合以下任一條件"
 
-#. dennis-ignore: W303 - False positive caused by <title>. https://github.com/mozilla/dennis/issues/213
 #: changedetectionio/forms.py
 msgid "Use page <title> in list"
 msgstr "在列表中使用頁面 <title>"
@@ -3041,7 +3056,6 @@ msgstr "已啟用即時 UI 更新"
 msgid "Favicons Enabled"
 msgstr "啟用網站圖示 (Favicons)"
 
-#. dennis-ignore: W303 - False positive caused by <title>. https://github.com/mozilla/dennis/issues/213
 #: changedetectionio/forms.py
 msgid "Use page <title> in watch overview list"
 msgstr "在監測概覽列表中使用頁面 <title>"
@@ -3154,12 +3168,8 @@ msgstr ""
 msgid "Default AI Change Summary prompt"
 msgstr ""
 
-#: changedetectionio/forms.py
-msgid "Max tokens per check"
-msgstr ""
-
-#: changedetectionio/forms.py
-msgid "Max cumulative tokens (per watch)"
+#: changedetectionio/blueprint/settings/templates/settings_llm_tab.html changedetectionio/forms.py
+msgid "Max tokens per watch per period"
 msgstr ""
 
 #: changedetectionio/forms.py
@@ -3170,6 +3180,10 @@ msgstr ""
 msgid "Max input characters"
 msgstr ""
 
+#: changedetectionio/forms.py
+msgid "Enable AI / LLM features"
+msgstr ""
+
 #: changedetectionio/forms.py
 msgid "Replace {{diff}} notification token with AI summary"
 msgstr ""
@@ -3435,7 +3449,6 @@ msgstr ""
 msgid "The UUID of the watch."
 msgstr "監測任務的 UUID。"
 
-#. dennis-ignore: W303 - False positive caused by <title>. https://github.com/mozilla/dennis/issues/213
 #: changedetectionio/templates/_common_fields.html
 msgid "The page title of the watch, uses <title> if not set, falls back to URL"
 msgstr ""
@@ -4163,6 +4176,17 @@ msgstr "更改語言"
 msgid "Change language"
 msgstr "更改語言"
 
+#: changedetectionio/validate_url.py
+msgid "API Base URL is not a valid http(s) URL."
+msgstr ""
+
+#: changedetectionio/validate_url.py
+msgid ""
+"API Base URL resolves to a private, loopback, link-local or reserved IP address and was blocked to prevent SSRF. To "
+"allow LLM endpoints on private networks (e.g. a local Ollama server) set the environment variable "
+"ALLOW_IANA_RESTRICTED_ADDRESSES=true and restart."
+msgstr ""
+
 #: changedetectionio/widgets/ternary_boolean.py
 msgid "Yes"
 msgstr "是"

changedetectionio/validate_url.py

@@ -80,6 +80,91 @@ def is_private_hostname(hostname):
     return False
 
 
+def extract_url_hostnames(url):
+    """Return every hostname this URL could resolve to under different URL parsers.
+
+    Why: urllib's urlparse() and urllib3's parse_url() disagree on URLs containing
+    a backslash (e.g. http://INTERNAL:8888\\@PUBLIC/ — urlparse extracts PUBLIC, but
+    urllib3/requests will actually connect to INTERNAL). Any SSRF check that trusts
+    only one parser can be bypassed by the other. Callers should reject the fetch
+    if ANY hostname returned here is private/reserved.
+
+    See GHSA-rph4-96w6-q594.
+    """
+    hostnames = set()
+    try:
+        h = urlparse(url).hostname
+        if h:
+            hostnames.add(h)
+    except Exception:
+        pass
+    try:
+        from urllib3.util.url import parse_url as _u3_parse_url
+        u3 = _u3_parse_url(url)
+        if u3.host:
+            # urllib3 keeps IPv6 brackets in `.host`; strip them so socket.getaddrinfo() accepts the literal.
+            hostnames.add(u3.host.strip('[]'))
+    except Exception:
+        pass
+    return hostnames
+
+
+def is_url_private_or_parser_confused(url):
+    """SSRF gate that defends against urlparse/urllib3 parser-differential attacks.
+
+    Returns True (block the fetch) when:
+      * the URL contains a backslash — no legitimate URL needs one, and it is the
+        established vector for the parser-differential bypass (GHSA-rph4-96w6-q594), OR
+      * any hostname produced by urlparse OR urllib3 resolves to a private/reserved IP.
+    """
+    if '\\' in url:
+        logger.warning(f"URL '{url}' contains a backslash — rejected to prevent urlparse/urllib3 parser-differential SSRF.")
+        return True
+    for hostname in extract_url_hostnames(url):
+        if is_private_hostname(hostname):
+            return True
+    return False
+
+
+def is_llm_api_base_safe(api_base):
+    """SSRF guard for the LLM `api_base` setting (GHSA-jrxm-qjfh-g54f).
+
+    Returns (ok: bool, reason: str). Empty/None api_base is allowed (cloud providers
+    don't need it). When ALLOW_IANA_RESTRICTED_ADDRESSES=true the check is bypassed
+    so operators can intentionally point at local Ollama / vLLM / LM Studio.
+
+    Call this from EVERY write path that accepts `llm.api_base` from the user —
+    form validation, AJAX endpoints, and any future REST/import endpoint. The
+    existing call sites are forms.py (validateLLMApiBaseSafe) and
+    blueprint/settings/llm.py (both /models and /test).
+    """
+    import os
+    from changedetectionio.strtobool import strtobool
+    from flask_babel import gettext
+
+    if not api_base or not api_base.strip():
+        return True, ''
+
+    if strtobool(os.getenv('ALLOW_IANA_RESTRICTED_ADDRESSES', 'false')):
+        return True, ''
+
+    api_base = api_base.strip()
+
+    if not is_safe_valid_url(api_base):
+        return False, gettext("API Base URL is not a valid http(s) URL.")
+
+    hostname = urlparse(api_base).hostname
+    if hostname and is_private_hostname(hostname):
+        return False, gettext(
+            "API Base URL resolves to a private, loopback, link-local or reserved "
+            "IP address and was blocked to prevent SSRF. To allow LLM endpoints on private networks "
+            "(e.g. a local Ollama server) set the environment variable "
+            "ALLOW_IANA_RESTRICTED_ADDRESSES=true and restart."
+        )
+
+    return True, ''
+
+
 def is_safe_valid_url(test_url):
     from changedetectionio import strtobool
     from changedetectionio.jinja2_custom import render as jinja_render
@@ -139,6 +224,13 @@ def is_safe_valid_url(test_url):
         logger.warning(f'URL "{test_url}" contains suspicious characters')
         return False
 
+    # Reject backslashes — urllib's urlparse and urllib3's parse_url disagree on URLs containing
+    # a backslash (e.g. http://INTERNAL:8888\@PUBLIC/), which is the documented SSRF bypass in
+    # GHSA-rph4-96w6-q594. A backslash has no legitimate use in an HTTP URL, so block at add-time.
+    if '\\' in test_url:
+        logger.warning(f'URL "{test_url}" contains a backslash — rejected (parser-differential SSRF vector).')
+        return False
+
     # Normalize URL encoding - handle both encoded and unencoded query parameters
     test_url = normalize_url_encoding(test_url)
 

changedetectionio/worker.py

@@ -432,9 +432,15 @@ async def async_update_worker(worker_id, q, notification_q, app, datastore, exec
                         update_obj['_llm_result'] = None
                         update_obj['_llm_intent'] = ''
                         update_obj['_llm_change_summary'] = ''
-                        # skip_check: when budget exceeded, don't run LLM or the check
-                        _llm_budget_action = datastore.data['settings']['application'].get('llm_budget_action', 'skip_llm')
-                        if _llm_budget_action == 'skip_check':
+                        # skip_check: when budget exceeded, don't run LLM or the check.
+                        # Also gated on llm_enabled — a disabled LLM can't be spending tokens,
+                        # so the budget enforcement shouldn't suppress changes when the user
+                        # has explicitly switched LLM off.
+                        from changedetectionio.llm.evaluator import is_llm_features_disabled as _is_llm_features_disabled, get_llm_settings as _get_llm_settings
+                        _llm_settings = _get_llm_settings(datastore)
+                        _llm_master_enabled = _llm_settings.enabled and not _is_llm_features_disabled()
+                        _llm_budget_action = _llm_settings.budget_action
+                        if _llm_master_enabled and _llm_budget_action == 'skip_check':
                             from changedetectionio.llm.evaluator import is_global_token_budget_exceeded
                             if is_global_token_budget_exceeded(datastore):
                                 logger.info(f"LLM monthly budget exceeded — skipping check for {uuid} (budget_action=skip_check)")
@@ -444,9 +450,14 @@ async def async_update_worker(worker_id, q, notification_q, app, datastore, exec
                             try:
                                 from changedetectionio.llm.evaluator import (
                                     evaluate_change, resolve_intent, resolve_llm_field,
-                                    summarise_change, get_llm_config,
+                                    summarise_change, _runtime_llm_config,
                                 )
-                                _llm_cfg = get_llm_config(datastore)
+                                # _runtime_llm_config returns None (and logs a debug skip
+                                # message) when the master 'llm_enabled' toggle is off, so
+                                # the whole block — diff computation, status minitext, and
+                                # the two executor dispatches — is skipped, not just the
+                                # inner LLM lookups.
+                                _llm_cfg = _runtime_llm_config(datastore)
                                 if _llm_cfg:
                                     # Compute unified diff once — used by both intent and summary
                                     _watch_dates = list(watch.history.keys())
@@ -538,8 +549,10 @@ async def async_update_worker(worker_id, q, notification_q, app, datastore, exec
                                         get_effective_summary_prompt, build_summary_cache_prompt,
                                     )
                                     _llm_to_version = list(watch.history.keys())[-1]
-                                    _llm_max_summary_tokens = datastore.data['settings']['application'].get('llm_max_summary_tokens', 3000)
-                                    _llm_model = (datastore.data['settings']['application'].get('llm') or {}).get('model', '')
+                                    from changedetectionio.llm.evaluator import get_llm_settings as _get_llm_settings_inner
+                                    _ls = _get_llm_settings_inner(datastore)
+                                    _llm_max_summary_tokens = _ls.max_summary_tokens
+                                    _llm_model = _ls.model
                                     _llm_cache_prompt = build_summary_cache_prompt(
                                         effective_prompt=get_effective_summary_prompt(watch, datastore),
                                         max_summary_tokens=_llm_max_summary_tokens,

docs/api-spec.yaml

@@ -621,6 +621,14 @@ components:
           type: [integer, 'null']
           readOnly: true
           description: Total tokens consumed by the AI across all checks for this watch.
+        llm_tokens_this_period:
+          type: [integer, 'null']
+          readOnly: true
+          description: Tokens consumed by the AI on this watch within the current rollover period (currently month). Used to enforce max_tokens_per_count_period.
+        llm_tokens_period_key:
+          type: [string, 'null']
+          readOnly: true
+          description: Identifier of the current rollover period (e.g. "2026-05"). Set automatically; resets llm_tokens_this_period when the period changes.
 
     DaySchedule:
       type: object

requirements.txt

@@ -148,6 +148,9 @@ pluggy ~= 1.6
 
 # LLM intent-based change evaluation (multi-provider via litellm)
 litellm>=1.40.0,<1.83.1  # 1.83.1–1.83.14 exact-pin jsonschema==4.23.0, conflicting with openapi-spec-validator's >=4.24.0 floor; re-evaluate when litellm fixes this
+# Used today for LLMSettings (model/LLMSettings.py); transitively pulled by litellm but pinned explicitly
+# so the validation/typing layer doesn't disappear if litellm drops it.
+pydantic>=2.0,<3.0
 # BM25 relevance trimming for large snapshots (pure Python, no ML)
 rank-bm25>=0.2.2
 
@@ -156,7 +159,7 @@ psutil==7.2.2
 
 ruff >= 0.11.2
 pre_commit >= 4.2.0
-dennis >= 1.2.0
+dennis >= 1.3.0
 
 # For events between checking and socketio updates
 blinker