Notifications - Fix 'str' object is not callable when {{ diff(...) }} callable tokens are used with HTML/htmlcolor output

2026-06-01 14:31:14 +00:00 · 2026-05-19 17:48:21 +02:00
51 changed files with 448 additions and 1402 deletions
@@ -2,7 +2,7 @@

 # Read more https://github.com/dgtlmoon/changedetection.io/wiki
 # Semver means never use .01, or 00. Should be .1.
-__version__ = '0.55.5'
+__version__ = '0.55.4'

 from changedetectionio.strtobool import strtobool
 from json.decoder import JSONDecodeError
@@ -10,15 +10,12 @@ from flask_babel import gettext

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


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

    @settings_blueprint.route("", methods=['GET', "POST"])
    @login_optionally_required
@@ -33,12 +30,24 @@ def construct_blueprint(datastore: ChangeDetectionStore):

        default = deepcopy(datastore.data['settings'])

-        # api_key is intentionally blanked on GET — PasswordField never re-renders
-        # its value, and a blank submission preserves the stored key.
-        default['llm'] = LLMSettings.model_validate(
-            datastore.data['settings']['application'].get('llm') or {}
-        ).model_dump()
-        default['llm']['api_key'] = ''
+        # 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),
+        }

        if datastore.proxy_list is not None:
            available_proxies = list(datastore.proxy_list.keys())
@@ -89,37 +98,79 @@ def construct_blueprint(datastore: ChangeDetectionStore):

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

-                # LLM config lives under settings.application.llm.* (post update_31).
-                # Hydrate the stored dict into LLMSettings, then merge form input over it.
-                # WTForms field names match LLMSettings field names exactly, so both sides
-                # of the merge use the same key shape.
-                existing_llm = LLMSettings.model_validate(
-                    datastore.data['settings']['application'].get('llm') or {}
+                # 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_form_input = dict(form.data.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

-                # 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)
+                # 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

-                # 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)
+                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)

                # Handle dynamic worker count adjustment
                old_worker_count = datastore.data['settings']['requests'].get('workers', 1)
@@ -193,7 +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, get_llm_settings
+            from changedetectionio.llm.evaluator import apply_local_token_multiplier
            text, total_tokens, input_tokens, output_tokens = completion(
                model=model,
                messages=[{'role': 'user', 'content':
@@ -201,7 +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=get_llm_settings(datastore).debug,
+                debug=bool(datastore.data['settings']['application'].get('llm_debug', False)),
            )
            reply = text.strip()
            if not reply:
@@ -232,17 +232,8 @@ def construct_llm_blueprint(datastore: ChangeDetectionStore):
    @llm_blueprint.route("/clear", methods=['POST'])
    @login_optionally_required
    def llm_clear():
-        from changedetectionio.model.LLMSettings import LLMSettings
        logger.debug("LLM configuration cleared by user")
-        # Read existing config, write back a dict that omits the connection fields —
-        # so the saved dict no longer has model/api_key/api_base/etc.
-        # Toggles, prompts, budgets and counters survive.
-        settings = LLMSettings.model_validate(
-            datastore.data['settings']['application'].get('llm') or {}
-        )
-        datastore.data['settings']['application']['llm'] = settings.model_dump(
-            exclude=set(LLMSettings.CONNECTION_FIELDS)
-        )
+        datastore.data['settings']['application'].pop('llm', None)
        datastore.commit()
        flash(gettext("AI / LLM configuration removed."), 'notice')
        return redirect(url_for('settings.settings_page') + '#ai')
@@ -34,9 +34,7 @@
            <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>
@@ -396,9 +394,7 @@ 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>
@@ -69,17 +69,6 @@
  {% 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>
@@ -125,22 +114,22 @@
  </div>

  <div class="pure-control-group">
-    {{ render_field(form.llm.form.api_key) }}
+    {{ render_field(form.llm.form.llm_api_key) }}
    <span class="pure-form-message-inline" id="llm-key-hint"></span>
  </div>
  <div class="pure-control-group" id="llm-base-group" style="display:none">
-    {{ render_field(form.llm.form.api_base) }}
+    {{ render_field(form.llm.form.llm_api_base) }}
    <span class="pure-form-message-inline">{{ _('Only needed for Ollama or custom/self-hosted endpoints. Leave blank for cloud providers.') }}</span>
  </div>

  {# Hidden field carrying the dropdown selection so the backend knows when to apply
     reasoning-friendly token caps (Ollama and OpenAI-compatible endpoints, which commonly
     serve reasoning models that need headroom for chain-of-thought to complete). #}
-  {{ form.llm.form.provider_kind() }}
+  {{ form.llm.form.llm_provider_kind() }}

  <div class="pure-control-group" id="llm-local-advanced-group" style="display:none">
-    <label for="{{ form.llm.form.local_token_multiplier.id }}">{{ form.llm.form.local_token_multiplier.label.text }}</label>
-    {{ form.llm.form.local_token_multiplier() }}
+    <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() }}
    <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>
@@ -163,7 +152,7 @@
  </div>

  <div class="pure-control-group">
-    {{ render_field(form.llm.form.model,
+    {{ render_field(form.llm.form.llm_model,
                    placeholder=_("Enter API key and click 'Load available models'")) }}
  </div>

@@ -223,9 +212,9 @@

  <div class="pure-control-group">
    <label></label>
-    {{ form.llm.form.debug() }}
-    <label for="{{ form.llm.form.debug.id }}" style="display:inline; font-weight:normal;">
-      {{ form.llm.form.debug.label.text }}
+    {{ 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 }}
    </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.') }}
@@ -243,10 +232,10 @@
  <p class="stab-section-title">{{ _('Default AI Change Summary') }}</p>

  <div class="pure-control-group">
-    {{ render_field(form.llm.form.change_summary_default) }}
+    {{ render_field(form.llm.form.llm_change_summary_default) }}
    <span class="pure-form-message-inline">
      {{ _('Used for all watches unless overridden by the watch or its tag/group.') }}
-      &nbsp;<a href="#" class="pure-button button-small" onclick="var t=document.getElementById('llm-change_summary_default'); if(!t.value && t.placeholder) t.value=t.placeholder; return false;">{{ _('Modify default prompt') }}</a>
+      &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>
    </span>
  </div>

@@ -259,9 +248,9 @@
  {% if llm_config and llm_config.get('model') %}
  <div class="pure-control-group">
    <label></label>
-    {{ form.llm.form.override_diff_with_summary() }}
-    <label for="{{ form.llm.form.override_diff_with_summary.id }}" style="display:inline; font-weight:normal;">
-      {{ form.llm.form.override_diff_with_summary.label.text }}
+    {{ 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 }}
    </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.',
@@ -271,9 +260,9 @@

  <div class="pure-control-group">
    <label></label>
-    {{ form.llm.form.restock_use_fallback_extract() }}
-    <label for="{{ form.llm.form.restock_use_fallback_extract.id }}" style="display:inline; font-weight:normal;">
-      {{ form.llm.form.restock_use_fallback_extract.label.text }}
+    {{ 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 }}
    </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.') }}
@@ -281,21 +270,21 @@
  </div>

  <div class="pure-control-group">
-    <label for="{{ form.llm.form.thinking_budget.id }}">{{ form.llm.form.thinking_budget.label.text }}</label>
-    {{ form.llm.form.thinking_budget() }}
+    <label for="{{ form.llm.form.llm_thinking_budget.id }}">{{ form.llm.form.llm_thinking_budget.label.text }}</label>
+    {{ form.llm.form.llm_thinking_budget() }}
    <span class="pure-form-message-inline">{{ _('For Gemini 2.5+ models only. Thinking tokens improve reasoning quality but count against the output budget. Set to Off if summaries are being cut short.') }}</span>
  </div>

  <div class="pure-control-group">
-    <label for="{{ form.llm.form.max_summary_tokens.id }}">{{ form.llm.form.max_summary_tokens.label.text }}</label>
-    {{ form.llm.form.max_summary_tokens() }}
+    <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() }}
    <span class="pure-form-message-inline">{{ _('Upper limit on tokens the AI may use when writing a change summary. Higher values allow longer summaries but cost more.') }}</span>
  </div>

  <div class="pure-control-group">
-    <label>{{ form.llm.form.budget_action.label.text }}</label>
+    <label>{{ form.llm.form.llm_budget_action.label.text }}</label>
    <div>
-      {% for subfield in form.llm.form.budget_action %}
+      {% for subfield in form.llm.form.llm_budget_action %}
      <label class="pure-radio" style="display:block; font-weight:normal; margin-bottom:0.3em;">
        {{ subfield() }} {{ subfield.label.text }}
      </label>
@@ -348,9 +337,9 @@
        {% 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-token_budget_month" value="{{ llm_token_budget_month_env }}">
+          <input type="hidden" name="llm-llm_token_budget_month" value="{{ llm_token_budget_month_env }}">
        {% else %}
-          {{ form.llm.form.token_budget_month(placeholder=_('0 = unlimited'), value=llm_stored.get('token_budget_month', 0) or '') }}
+          {{ form.llm.form.llm_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>
@@ -365,21 +354,14 @@
      <span class="llm-usage-row-label">{{ _('Max input characters') }}</span>
      <span class="llm-usage-row-value">
        {% if llm_max_input_chars_env %}
-          {{ form.llm.form.max_input_chars(value=llm_max_input_chars_env, readonly=True, style="width:10em;opacity:0.6;cursor:not-allowed;") }}
+          {{ form.llm.form.llm_max_input_chars(value=llm_max_input_chars_env, readonly=True, style="width:10em;opacity:0.6;cursor:not-allowed;") }}
          <span class="llm-env-badge">{{ _('(set via <code>LLM_MAX_INPUT_CHARS</code>)') | safe }}</span>
        {% else %}
-          {{ form.llm.form.max_input_chars(placeholder='100000', value=llm_stored.get('max_input_chars', 100000) or '') }}
+          {{ form.llm.form.llm_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 %}
@@ -392,9 +374,9 @@
        {% 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-token_budget_month" value="{{ llm_token_budget_month_env }}">
+          <input type="hidden" name="llm-llm_token_budget_month" value="{{ llm_token_budget_month_env }}">
        {% else %}
-          {{ form.llm.form.token_budget_month(placeholder=_('0 = unlimited'), value=llm_stored.get('token_budget_month', 0) or '') }}
+          {{ form.llm.form.llm_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>
@@ -403,21 +385,14 @@
      <span class="llm-usage-row-label">{{ _('Max input characters') }}</span>
      <span class="llm-usage-row-value">
        {% if llm_max_input_chars_env %}
-          {{ form.llm.form.max_input_chars(value=llm_max_input_chars_env, readonly=True, style="width:10em;opacity:0.6;cursor:not-allowed;") }}
+          {{ form.llm.form.llm_max_input_chars(value=llm_max_input_chars_env, readonly=True, style="width:10em;opacity:0.6;cursor:not-allowed;") }}
          <span class="llm-env-badge">{{ _('(set via <code>LLM_MAX_INPUT_CHARS</code>)') | safe }}</span>
        {% else %}
-          {{ form.llm.form.max_input_chars(placeholder='100000', value=llm_stored.get('max_input_chars', 100000) or '') }}
+          {{ form.llm.form.llm_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 %}
@@ -448,8 +423,8 @@
    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-api_base"]');
-    const kindField     = document.querySelector('[name="llm-provider_kind"]');
+    const baseField     = document.querySelector('[name="llm-llm_api_base"]');
+    const kindField     = document.querySelector('[name="llm-llm_provider_kind"]');
    const hint          = document.getElementById('llm-key-hint');

    fetchGroup.style.display = LIVE_PROVIDERS.includes(provider) ? '' : 'none';
@@ -477,8 +452,8 @@

  window.llmFetchModels = async function () {
    const provider  = document.getElementById('llm-provider').value;
-    const apiKey    = document.querySelector('[name="llm-api_key"]').value.trim();
-    const apiBase   = document.querySelector('[name="llm-api_base"]').value.trim();
+    const apiKey    = document.querySelector('[name="llm-llm_api_key"]').value.trim();
+    const apiBase   = document.querySelector('[name="llm-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');
@@ -513,7 +488,7 @@
      }

      modelSel.innerHTML = '<option value="">{{ _("— choose a model —") }}</option>';
-      const currentModel = document.querySelector('[name="llm-model"]').value.trim();
+      const currentModel = document.querySelector('[name="llm-llm_model"]').value.trim();
      for (const m of data.models) {
        const opt = document.createElement('option');
        opt.value = m;
@@ -535,7 +510,7 @@
  };

  window.llmOnModelPick = function (value) {
-    if (value) document.querySelector('[name="llm-model"]').value = value;
+    if (value) document.querySelector('[name="llm-llm_model"]').value = value;
  };

  window.llmRunTest = async function () {
@@ -551,11 +526,11 @@
    // 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-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 || '';
+    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 || '';
    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());
@@ -585,7 +560,7 @@

  // On page load: detect and pre-select provider from current model
  (function detectCurrentProvider() {
-    const modelField = document.querySelector('[name="llm-model"]');
+    const modelField = document.querySelector('[name="llm-llm_model"]');
    if (!modelField) return;
    const m = modelField.value.trim();
    if (!m) return;
@@ -596,7 +571,7 @@
    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-api_base"]');
+      const baseField = document.querySelector('[name="llm-llm_api_base"]');
      guessed = (baseField && baseField.value.trim()) ? 'openai_compatible' : 'openai';
    }
    else if (m.startsWith('claude'))   guessed = 'anthropic';
@@ -272,10 +272,8 @@ 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.
-        from changedetectionio.llm.evaluator import get_llm_settings
-        _ls = get_llm_settings(datastore)
-        _max_summary_tokens = _ls.max_summary_tokens
-        _llm_model = _ls.model
+        _max_summary_tokens = datastore.data['settings']['application'].get('llm_max_summary_tokens', 3000)
+        _llm_model = (datastore.data['settings']['application'].get('llm') or {}).get('model', '')
        cache_prompt = build_summary_cache_prompt(
            effective_prompt=get_effective_summary_prompt(watch, datastore),
            max_summary_tokens=_max_summary_tokens,
@@ -57,9 +57,7 @@
            {% 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>
@@ -323,11 +321,9 @@ 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>
@@ -507,7 +503,7 @@ Math: {{ 1 + 1 }}") }}
                            <td>{{ _('Server type reply') }}</td>
                            <td>{{ watch.get('remote_server_reply') }}</td>
                        </tr>
-                        {% if not llm_features_disabled and settings_application.get('llm', {}).get('model') %}
+                        {% if settings_application.get('llm', {}).get('model') %}
                        <tr>
                            <td>{{ _('AI tokens (last check)') }}</td>
                            <td>{{ "{:,}".format(watch.get('llm_last_tokens_used') or 0) }}</td>
@@ -522,11 +522,6 @@ 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():
@@ -1108,12 +1108,12 @@ class globalSettingsLLMForm(Form):
      gemini/gemini-2.0-flash               → Google Gemini
      azure/gpt-4o                          → Azure OpenAI
    """
-    model = StringField(
+    llm_model = StringField(
        _l('Model'),
        validators=[validators.Optional()],
        render_kw={"placeholder": "gpt-4o-mini", "style": "width: 24em;"},
    )
-    api_key = PasswordField(
+    llm_api_key = PasswordField(
        _l('API Key'),
        validators=[validators.Optional()],
        render_kw={
@@ -1121,7 +1121,7 @@ class globalSettingsLLMForm(Form):
            "style": "width: 24em;",
        },
    )
-    api_base = StringField(
+    llm_api_base = StringField(
        _l('API Base URL'),
        validators=[validators.Optional(), validateLLMApiBaseSafe()],
        render_kw={
@@ -1132,7 +1132,7 @@ 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.
-    provider_kind = HiddenField(
+    llm_provider_kind = HiddenField(
        validators=[validators.Optional()],
        default='',
    )
@@ -1144,13 +1144,13 @@ 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.
-    local_token_multiplier = IntegerField(
+    llm_local_token_multiplier = IntegerField(
        _l('Token multiplier for local reasoning models'),
        validators=[validators.Optional(), validators.NumberRange(min=1, max=20)],
        default=5,
        render_kw={"placeholder": "5", "style": "width: 6em;"},
    )
-    change_summary_default = TextAreaField(
+    llm_change_summary_default = TextAreaField(
        _l('Default AI Change Summary prompt'),
        validators=[validators.Optional(), validators.Length(max=2000)],
        render_kw={
@@ -1160,8 +1160,8 @@ class globalSettingsLLMForm(Form):
        },
        default='',
    )
-    max_tokens_per_count_period = IntegerField(
-        _l('Max tokens per watch per period'),
+    llm_max_tokens_per_check = IntegerField(
+        _l('Max tokens per check'),
        validators=[validators.Optional(), validators.NumberRange(min=0)],
        default=0,
        render_kw={
@@ -1169,13 +1169,22 @@ class globalSettingsLLMForm(Form):
            "style": "width: 8em;",
        },
    )
-    token_budget_month = IntegerField(
+    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(
        _l('Monthly token budget'),
        validators=[validators.Optional(), validators.NumberRange(min=0)],
        default=0,
        render_kw={"style": "width: 10em;"},
    )
-    max_input_chars = IntegerField(
+    llm_max_input_chars = IntegerField(
        _l('Max input characters'),
        validators=[validators.Optional(), validators.NumberRange(min=1)],
        default=100000,
@@ -1184,27 +1193,19 @@ class globalSettingsLLMForm(Form):
            "style": "width: 10em;",
        },
    )
-    # Master on/off switch for ALL LLM lookups at runtime. When False, every entry point
-    # in evaluator.py (and the restock fallback) short-circuits with a logger.debug
-    # message — even if a provider+model is still configured. Saved config and the
-    # "configured" badge remain visible so the user can toggle back on without re-entering.
-    enabled = BooleanField(
-        _l('Enable AI / LLM features'),
-        default=True,
-    )
-    override_diff_with_summary = BooleanField(
+    llm_override_diff_with_summary = BooleanField(
        _l('Replace {{diff}} notification token with AI summary'),
        default=True,
    )
-    restock_use_fallback_extract = BooleanField(
+    llm_restock_use_fallback_extract = BooleanField(
        _l('Use LLM as a fallback for extracting price and restock info'),
        default=True,
    )
-    debug = BooleanField(
+    llm_debug = BooleanField(
        _l('Enable LLM debug logging'),
        default=False,
    )
-    thinking_budget = SelectField(
+    llm_thinking_budget = SelectField(
        _l('AI thinking budget (tokens)'),
        choices=[
            ('0',    _l('Off (no thinking)')),
@@ -1215,7 +1216,7 @@ class globalSettingsLLMForm(Form):
        default=str(LLM_DEFAULT_THINKING_BUDGET),
        validators=[validators.Optional()],
    )
-    max_summary_tokens = SelectField(
+    llm_max_summary_tokens = SelectField(
        _l('Max AI summary length (tokens)'),
        choices=[
            ('500',   '500'),
@@ -1228,7 +1229,7 @@ class globalSettingsLLMForm(Form):
        default=str(LLM_DEFAULT_MAX_SUMMARY_TOKENS),
        validators=[validators.Optional()],
    )
-    budget_action = RadioField(
+    llm_budget_action = RadioField(
        _l('When monthly token budget is reached'),
        choices=[
            ('skip_llm',   _l('Skip AI summarisation only (watch still checks)')),
@@ -20,8 +20,6 @@ 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,
@@ -31,29 +29,7 @@ from .prompt_builder import (
 )
 from .response_parser import parse_eval_response, parse_preview_response, parse_setup_response

-from changedetectionio.model.LLMSettings import (
-    LLMSettings,
-    LLM_DEFAULT_MAX_INPUT_CHARS as _DEFAULT_MAX_INPUT_CHARS,
-    LLM_DEFAULT_MAX_SUMMARY_TOKENS,
-    LLM_DEFAULT_THINKING_BUDGET,
-)
-
-
-def is_llm_features_disabled() -> bool:
-    """True when the LLM_FEATURES_DISABLED env var is set to a truthy value."""
-    return bool(strtobool(os.getenv('LLM_FEATURES_DISABLED', '')))
-
-
-def get_llm_settings(datastore) -> LLMSettings:
-    """Hydrate the LLM config dict at settings.application.llm into a validated model.
-
-    Returns a default-constructed LLMSettings when the dict is missing or empty —
-    callers never have to None-check the result. The storage layer remains a plain
-    dict; this is only the validation/typing layer for reads.
-    """
-    cfg = datastore.data.get('settings', {}).get('application', {}).get('llm') or {}
-    return LLMSettings.model_validate(cfg)
-
+_DEFAULT_MAX_INPUT_CHARS = 100_000

 def _get_max_input_chars(datastore) -> int:
    """Max input characters to send to the LLM. Resolution: env var → datastore → 100,000.
@@ -62,9 +38,10 @@ 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)
-    stored = get_llm_settings(datastore).max_input_chars
-    if stored and stored > 0:
-        return stored
+    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)
    return _DEFAULT_MAX_INPUT_CHARS


@@ -80,6 +57,8 @@ 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).
@@ -101,6 +80,8 @@ 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.
@@ -226,8 +207,6 @@ 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:
@@ -246,33 +225,9 @@ 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
 # ---------------------------------------------------------------------------
@@ -343,22 +298,25 @@ def accumulate_global_tokens(datastore, tokens: int,

    current_month = _get_month_key()
    cost = _estimate_cost_usd(model, input_tokens, output_tokens)
-    settings = get_llm_settings(datastore)
+
+    # 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']

    # Month rollover: reset monthly counters
-    if settings.tokens_month_key != current_month:
-        settings.tokens_this_month = 0
-        settings.cost_usd_this_month = 0.0
-        settings.tokens_month_key = current_month
+    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

-    settings.tokens_total_cumulative += tokens
-    settings.tokens_this_month       += tokens
-    settings.cost_usd_total_cumulative += cost
-    settings.cost_usd_this_month       += cost
+    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

-    # 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()
+    # Persist immediately — token accounting must survive restarts
    datastore.commit()


@@ -386,44 +344,31 @@ def is_global_token_budget_exceeded(datastore) -> bool:

 def _check_token_budget(watch, cfg, tokens_this_call: int = 0) -> bool:
    """
-    Per-watch per-period token cap.
-
-    Period is currently month (matches the global counter rollover); the field
-    name `max_tokens_per_count_period` is period-agnostic so a configurable
-    day/week/month can land later without renaming storage.
-
-    On non-zero tokens_this_call:
-      - rolls over watch['llm_tokens_this_period'] if a new period started
-      - increments the per-period counter
-      - also increments the existing lifetime counter (UI stat, unchanged)
-    Returns False once the per-period counter exceeds max_tokens_per_count_period
-    so subsequent evaluate_change calls bail out for this watch until rollover.
-
-    Note: only evaluate_change actually gates on the return value (the other
-    callers invoke this for the side-effect of accumulating tokens).
+    Check token budget limits.  Returns True if within budget, False if exceeded.
+    Also accumulates tokens_this_call into watch['llm_tokens_used_cumulative'].
    """
    if tokens_this_call > 0:
-        current_period = _get_month_key()
-        # Rollover: new period zeroes the per-period counter
-        if watch.get('llm_tokens_period_key') != current_period:
-            watch['llm_tokens_this_period'] = 0
-            watch['llm_tokens_period_key'] = current_period
-        watch['llm_tokens_this_period'] = (watch.get('llm_tokens_this_period') or 0) + tokens_this_call
-        # Informational lifetime counter (UI shows this; not used for the cap)
-        watch['llm_tokens_used_cumulative'] = (watch.get('llm_tokens_used_cumulative') or 0) + tokens_this_call
+        current = watch.get('llm_tokens_used_cumulative') or 0
+        watch['llm_tokens_used_cumulative'] = current + tokens_this_call

-    max_per_period = int(cfg.get('max_tokens_per_count_period') or 0)
-    if max_per_period:
-        # Pre-flight (tokens_this_call=0) and post-call paths both read the
-        # same counter — but a stale period key means "no usage yet this period".
-        if watch.get('llm_tokens_period_key') == _get_month_key():
-            total = watch.get('llm_tokens_this_period') or 0
-            if total > max_per_period:
-                logger.warning(
-                    f"LLM per-period token budget exceeded for {watch.get('uuid')}: "
-                    f"{total} tokens > limit {max_per_period}"
-                )
-                return False
+    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

    return True

@@ -434,7 +379,7 @@ 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 = _runtime_llm_config(datastore)
+    cfg = get_llm_config(datastore)
    if not cfg:
        return

@@ -445,7 +390,6 @@ 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(
@@ -457,8 +401,8 @@ 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'], settings.thinking_budget),
-            debug=settings.debug,
+            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)),
        )
        _check_token_budget(watch, cfg, tokens)
        accumulate_global_tokens(datastore, tokens, model=cfg['model'])
@@ -482,7 +426,11 @@ def get_effective_summary_prompt(watch, datastore) -> str:
    prompt, _ = resolve_llm_field(watch, datastore, 'llm_change_summary')
    if prompt:
        return prompt
-    global_default = get_llm_settings(datastore).change_summary_default.strip()
+    global_default = (
+        datastore.data.get('settings', {})
+        .get('application', {})
+        .get('llm_change_summary_default', '') or ''
+    ).strip()
    return global_default or DEFAULT_CHANGE_SUMMARY_PROMPT


@@ -561,7 +509,7 @@ 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 = _runtime_llm_config(datastore)
+    cfg = get_llm_config(datastore)
    if not cfg:
        return ''

@@ -592,8 +540,8 @@ def summarise_change(watch, datastore, diff: str, current_snapshot: str = '') ->
        title=title,
    )

-    settings = get_llm_settings(datastore)
-    _extra_body = _thinking_extra_body(cfg['model'], settings.thinking_budget)
+    _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)

    try:
        _resp = llm_client.completion(
@@ -605,11 +553,14 @@ 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=settings.max_summary_tokens),
+                _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),
+                ),
                cfg,
            ),
            extra_body=_extra_body,
-            debug=settings.debug,
+            debug=bool(datastore.data['settings']['application'].get('llm_debug', False)),
        )
        raw, tokens = _resp[0], _resp[1]
        input_tokens  = _resp[2] if len(_resp) > 2 else 0
@@ -646,7 +597,7 @@ def preview_extract(watch, datastore, content: str) -> dict | None:

    Returns {'found': bool, 'answer': str} or None if LLM not configured / no intent.
    """
-    cfg = _runtime_llm_config(datastore)
+    cfg = get_llm_config(datastore)
    if not cfg:
        return None

@@ -660,7 +611,6 @@ 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(
@@ -672,8 +622,8 @@ 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'], settings.thinking_budget),
-            debug=settings.debug,
+            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)),
        )
        accumulate_global_tokens(datastore, tokens, model=cfg['model'])
        result = parse_preview_response(raw)
@@ -698,7 +648,7 @@ 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 = _runtime_llm_config(datastore)
+    cfg = get_llm_config(datastore)
    if not cfg:
        return None

@@ -747,7 +697,6 @@ 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'],
@@ -758,8 +707,8 @@ 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'], settings.thinking_budget),
-            debug=settings.debug,
+            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)),
        )
        raw, tokens = _resp[0], _resp[1]
        input_tokens  = _resp[2] if len(_resp) > 2 else 0
@@ -2,6 +2,7 @@ from os import getenv
 from copy import deepcopy

 from changedetectionio.blueprint.rss import RSS_FORMAT_TYPES, RSS_CONTENT_FORMAT_DEFAULT
+from changedetectionio.llm.evaluator import LLM_DEFAULT_MAX_SUMMARY_TOKENS, LLM_DEFAULT_THINKING_BUDGET
 from changedetectionio.model.Tags import TagsDict

 from changedetectionio.notification import (
@@ -70,9 +71,8 @@ class model(dict):
                    'shared_diff_access': False,
                    'strip_ignored_lines': False,
                    'tags': None,  # Initialized in __init__ with real datastore_path
-                    # All LLM settings now live nested under application.llm.* (post-migration update_31).
-                    # Defaults come from LLMSettings.model_validate({}) at read time —
-                    # no need to pre-seed an empty {} here.
+                    'llm_thinking_budget': LLM_DEFAULT_THINKING_BUDGET,
+                    'llm_max_summary_tokens': LLM_DEFAULT_MAX_SUMMARY_TOKENS,
                    'webdriver_delay': None , # Extra delay in seconds before extracting text
                    'ui': {
                        'use_page_title_in_list': True,
@@ -1,65 +0,0 @@
-"""
-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',
-    )
@@ -1,239 +0,0 @@
-# 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
@@ -376,8 +376,7 @@ 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()
-    from changedetectionio.llm.evaluator import get_llm_settings
-    _override_diff = get_llm_settings(datastore).override_diff_with_summary
+    _override_diff = datastore.data['settings']['application'].get('llm_override_diff_with_summary', True)
    if _llm_change_summary and _override_diff:
        n_object['diff'] = _llm_change_summary

@@ -204,8 +204,6 @@ 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,
@@ -196,23 +196,22 @@ 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 _runtime_llm_config, accumulate_global_tokens, get_llm_settings
+        from changedetectionio.llm.evaluator import get_llm_config, accumulate_global_tokens
        from changedetectionio.llm import client as llm_client
    except ImportError as e:
        logger.debug(f"LLM restock fallback: LLM libraries not available ({e})")
        return None

-    # Gate on the user setting (default True — enabled out of the box)
-    if not get_llm_settings(datastore).restock_use_fallback_extract:
-        logger.debug("LLM restock fallback: disabled in settings")
-        return None
-
-    # _runtime_llm_config returns None (with a debug log) when the master 'llm_enabled'
-    # toggle is off, so this path is gated for free.
-    llm_cfg = _runtime_llm_config(datastore)
+    llm_cfg = get_llm_config(datastore)
    if not llm_cfg or not llm_cfg.get('model'):
-        logger.debug("LLM restock fallback: no LLM model configured or LLM disabled, skipping")
+        logger.debug("LLM restock fallback: no LLM model configured, skipping")
        return None

    text_content = _strip_html(content) if content else ''
@@ -35,50 +35,6 @@ 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
@@ -94,7 +50,6 @@ 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 = []
@@ -134,22 +89,15 @@ 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.
-            # 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).
+            # 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
            # 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=3) as executor:
+                with ThreadPoolExecutor(max_workers=2) 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

@@ -163,11 +111,10 @@ 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 = _compute_ignore_line_numbers_for_preview(
-            text_pre_extract=text_pre_extract,
-            ignore_patterns=text_to_ignore,
-            extract_patterns=tmp_watch.get('extract_text', [])
-        )
+        ignore_line_numbers = html_tools.strip_ignore_text(content=text_after_filter,
+                                                           wordlist=text_to_ignore,
+                                                           mode='line numbers'
+                                                           )
    except Exception as e:
        text_before_filter = f"Error: {str(e)}"

@@ -217,10 +217,8 @@ 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.
-            from changedetectionio.llm.evaluator import get_llm_settings
-            _ls = get_llm_settings(datastore)
-            _max_summary_tokens = _ls.max_summary_tokens
-            _llm_model = _ls.model
+            _max_summary_tokens = datastore.data['settings']['application'].get('llm_max_summary_tokens', 3000)
+            _llm_model = (datastore.data['settings']['application'].get('llm') or {}).get('model', '')
            _cache_prompt = build_summary_cache_prompt(
                effective_prompt=_prompt,
                max_summary_tokens=_max_summary_tokens,
@@ -9,10 +9,6 @@ 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();
    });

@@ -775,53 +775,3 @@ 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")
-
@@ -112,7 +112,7 @@
                <td><code>{{ '{{triggered_text}}' }}</code></td>
                <td>{{ _('Text that tripped the trigger from filters') }}</td>
            </tr>
-            {% if not llm_features_disabled and settings_application and settings_application.get('llm', {}).get('model') %}
+            {% if 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>
@@ -281,7 +281,6 @@
      </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">
@@ -295,7 +294,6 @@
        <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 %}
@@ -37,12 +37,10 @@
    </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">
@@ -294,82 +294,78 @@ class TestTokenBudget:

        assert _check_token_budget(watch, cfg, tokens_this_call=10_000) is True

-    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
+    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

        watch = _make_watch()
-        watch['llm_tokens_this_period'] = 900
-        watch['llm_tokens_period_key'] = _get_month_key()
-        cfg = {'max_tokens_per_count_period': 1000}
+        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}

        # This call adds 200 → total 1100 > 1000
        result = _check_token_budget(watch, cfg, tokens_this_call=200)
        assert result is False

-    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
+    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

        watch = _make_watch()
-        watch['llm_tokens_this_period'] = 500
-        watch['llm_tokens_period_key'] = _get_month_key()
-        cfg = {'max_tokens_per_count_period': 1000}
+        watch['llm_tokens_used_cumulative'] = 500
+        cfg = {'max_tokens_cumulative': 1000}

        result = _check_token_budget(watch, cfg, tokens_this_call=100)
        assert result is True

-    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
+    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

        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_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
+    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

        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_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
+    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

-        ds = _make_datastore(llm_cfg={'model': 'gpt-4o-mini', 'max_tokens_per_count_period': 100})
+        ds = _make_datastore(llm_cfg={'model': 'gpt-4o-mini', 'max_tokens_cumulative': 100})
        watch = _make_watch(llm_intent='flag price drops')
-        watch['llm_tokens_this_period'] = 500  # already far over
-        watch['llm_tokens_period_key'] = _get_month_key()
+        watch['llm_tokens_used_cumulative'] = 500  # already far over

        with patch('changedetectionio.llm.client.completion') as mock_llm:
            result = evaluate_change(watch, ds, diff='- $500\n+ $400')
@@ -378,6 +374,23 @@ 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)
@@ -1,62 +0,0 @@
-"""
-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
@@ -14,9 +14,8 @@ 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': '',
@@ -85,8 +84,8 @@ class TestLLMRestockPluginDisabled:
        ds.data = {
            'settings': {
                'application': {
-                    # No 'llm' key → get_llm_config returns None;
-                    # restock_use_fallback_extract still defaults to True via LLMSettings
+                    'llm_restock_use_fallback_extract': True,
+                    # No 'llm' key → get_llm_config returns None
                }
            }
        }
@@ -77,82 +77,3 @@ 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)
@@ -329,9 +329,9 @@ def test_settings_form_preserves_api_key_when_submitted_blank(
    res = client.post(
        url_for('settings.settings_page'),
        data={
-            'llm-model': 'gpt-4o',
-            'llm-api_key': '',           # blank — PasswordField behaviour
-            'llm-api_base': '',
+            'llm-llm_model': 'gpt-4o',
+            'llm-llm_api_key': '',           # blank — PasswordField behaviour
+            'llm-llm_api_base': '',
            'application-pager_size': '50',
            'application-notification_format': 'System default',
            'requests-time_between_check-days': '0',
@@ -452,9 +452,9 @@ def test_settings_form_rejects_private_api_base(
    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',
+            'llm-llm_model':    'gpt-4o',
+            'llm-llm_api_key':  '',
+            'llm-llm_api_base': 'http://127.0.0.1:11434',
            'application-pager_size': '50',
            'application-notification_format': 'System default',
            'requests-time_between_check-days': '0',
@@ -653,23 +653,11 @@ def test_llm_clear_summary_cache_get_does_not_wipe_cache(

 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.
-    """
+    """Confirm the legit confirm-then-POST flow still wipes LLM config."""
    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
+    assert 'llm' not in ds.data['settings']['application']
@@ -28,11 +28,7 @@ def _set_response(datastore_path, content):

 def _configure_llm(client):
    ds = client.application.config.get('DATASTORE')
-    # 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
+    ds.data['settings']['application']['llm'] = {'model': 'gpt-4o-mini', 'api_key': 'sk-test'}


 # ---------------------------------------------------------------------------
@@ -242,9 +238,7 @@ def test_llm_summary_ajax_error_displayed_not_silenced(
 # ---------------------------------------------------------------------------

 def _set_global_default(ds, prompt):
-    llm = ds.data['settings']['application'].get('llm') or {}
-    llm['change_summary_default'] = prompt
-    ds.data['settings']['application']['llm'] = llm
+    ds.data['settings']['application']['llm_change_summary_default'] = prompt


 def test_global_default_used_when_watch_and_tag_have_none(
@@ -335,7 +329,7 @@ def test_hardcoded_fallback_when_nothing_set(
    watch['llm_change_summary'] = ''

    # Ensure global default is also empty
-    _set_global_default(ds, '')
+    ds.data['settings']['application']['llm_change_summary_default'] = ''

    assert get_effective_summary_prompt(watch, ds) == DEFAULT_CHANGE_SUMMARY_PROMPT

@@ -397,8 +391,8 @@ 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 the global default prompt into
-    application.llm.change_summary_default (single nested home for all LLM settings).
+    Submitting the settings form persists llm_change_summary_default at
+    settings.application level (not inside the llm credentials dict).
    """
    from changedetectionio.tests.util import live_server_setup
    live_server_setup(live_server)
@@ -411,20 +405,21 @@ 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-change_summary_default': 'Saved global prompt.',
+            'llm-llm_change_summary_default': 'Saved global prompt.',
            # Keep existing model so llm block is retained
-            'llm-model': 'gpt-4o-mini',
+            'llm-llm_model': 'gpt-4o-mini',
        },
        follow_redirects=True,
    )
    assert b'Settings updated.' in res.data

    ds = client.application.config.get('DATASTORE')
-    llm_dict = ds.data['settings']['application'].get('llm', {})
-    assert llm_dict.get('change_summary_default') == 'Saved global prompt.', f"Got: {llm_dict!r}"
+    stored = ds.data['settings']['application'].get('llm_change_summary_default', '')
+    assert stored == 'Saved global prompt.', f"Got: {stored!r}"

-    # And the old flat key must not be re-introduced
-    assert 'llm_change_summary_default' not in ds.data['settings']['application']
+    # 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

    delete_all_watches(client)

@@ -445,11 +440,7 @@ def test_global_default_survives_llm_clear(
    res = client.post(url_for('settings.llm.llm_clear'), follow_redirects=True)
    assert res.status_code == 200

-    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
+    assert ds.data['settings']['application'].get('llm_change_summary_default') == 'Surviving prompt.'

    delete_all_watches(client)

@@ -168,9 +168,9 @@ def test_settings_form_preserves_token_counters(
        url_for('settings.settings_page'),
        data={
            # LLM sub-form fields
-            'llm-model': 'gpt-4o',
-            'llm-api_key': 'sk-different-key',
-            'llm-api_base': '',
+            'llm-llm_model': 'gpt-4o',
+            'llm-llm_api_key': 'sk-different-key',
+            'llm-llm_api_base': '',
            # Minimal required fields to pass form validation
            'application-pager_size': '50',
            'application-notification_format': 'System default',
@@ -209,9 +209,9 @@ def test_settings_form_cannot_inject_fake_token_counts(
    res = client.post(
        url_for('settings.settings_page'),
        data={
-            'llm-model': 'gpt-4o-mini',
-            'llm-api_key': 'sk-test',
-            'llm-api_base': '',
+            'llm-llm_model': 'gpt-4o-mini',
+            'llm-llm_api_key': 'sk-test',
+            'llm-llm_api_base': '',
            # Attempted injection of token counter fields
            'llm-tokens_this_month': '0',
            'llm-tokens_total_cumulative': '0',
@@ -471,9 +471,9 @@ def test_cost_fields_are_tamper_proof_via_settings_form(
    client.post(
        url_for('settings.settings_page'),
        data={
-            'llm-model': 'gpt-4o',
-            'llm-api_key': 'sk-test',
-            'llm-api_base': '',
+            'llm-llm_model': 'gpt-4o',
+            'llm-llm_api_key': 'sk-test',
+            'llm-llm_api_base': '',
            'llm-cost_usd_this_month': '0',       # injection attempt
            'llm-cost_usd_total_cumulative': '0',  # injection attempt
            'application-pager_size': '50',
@@ -1,161 +0,0 @@
-#!/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()
@@ -842,10 +842,6 @@ 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 ""
@@ -1096,12 +1092,6 @@ 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 ""
@@ -3172,8 +3162,12 @@ msgstr ""
 msgid "Default AI Change Summary prompt"
 msgstr ""

-#: changedetectionio/blueprint/settings/templates/settings_llm_tab.html changedetectionio/forms.py
-msgid "Max tokens per watch per period"
+#: changedetectionio/forms.py
+msgid "Max tokens per check"
+msgstr ""
+
+#: changedetectionio/forms.py
+msgid "Max cumulative tokens (per watch)"
 msgstr ""

 #: changedetectionio/forms.py
@@ -3184,10 +3178,6 @@ 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 ""
@@ -858,10 +858,6 @@ 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 ""
@@ -1112,12 +1108,6 @@ 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 ""
@@ -3224,8 +3214,12 @@ msgstr ""
 msgid "Default AI Change Summary prompt"
 msgstr ""

-#: changedetectionio/blueprint/settings/templates/settings_llm_tab.html changedetectionio/forms.py
-msgid "Max tokens per watch per period"
+#: changedetectionio/forms.py
+msgid "Max tokens per check"
+msgstr ""
+
+#: changedetectionio/forms.py
+msgid "Max cumulative tokens (per watch)"
 msgstr ""

 #: changedetectionio/forms.py
@@ -3236,10 +3230,6 @@ 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 ""
@@ -840,10 +840,6 @@ 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 ""
@@ -1094,12 +1090,6 @@ 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 ""
@@ -3166,8 +3156,12 @@ msgstr ""
 msgid "Default AI Change Summary prompt"
 msgstr ""

-#: changedetectionio/blueprint/settings/templates/settings_llm_tab.html changedetectionio/forms.py
-msgid "Max tokens per watch per period"
+#: changedetectionio/forms.py
+msgid "Max tokens per check"
+msgstr ""
+
+#: changedetectionio/forms.py
+msgid "Max cumulative tokens (per watch)"
 msgstr ""

 #: changedetectionio/forms.py
@@ -3178,10 +3172,6 @@ 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 ""
@@ -840,10 +840,6 @@ 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 ""
@@ -1094,12 +1090,6 @@ 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 ""
@@ -3166,8 +3156,12 @@ msgstr ""
 msgid "Default AI Change Summary prompt"
 msgstr ""

-#: changedetectionio/blueprint/settings/templates/settings_llm_tab.html changedetectionio/forms.py
-msgid "Max tokens per watch per period"
+#: changedetectionio/forms.py
+msgid "Max tokens per check"
+msgstr ""
+
+#: changedetectionio/forms.py
+msgid "Max cumulative tokens (per watch)"
 msgstr ""

 #: changedetectionio/forms.py
@@ -3178,10 +3172,6 @@ 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 ""
@@ -878,10 +878,6 @@ 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 ""
@@ -1132,12 +1128,6 @@ 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 ""
@@ -2324,11 +2314,11 @@ msgstr "Último Comprobado"

 #: changedetectionio/blueprint/watchlist/templates/watch-overview.html
 msgid "Changed"
-msgstr "Cambiado"
+msgstr "Cambiadp"

 #: changedetectionio/blueprint/watchlist/templates/watch-overview.html
 msgid "Last Changed"
-msgstr "Último Cambiado"
+msgstr "Último Cambiadp"

 #: changedetectionio/blueprint/watchlist/templates/watch-overview.html
 msgid "No web page change detection watches configured, please add a URL in the box above, or"
@@ -3239,8 +3229,12 @@ msgstr ""
 msgid "Default AI Change Summary prompt"
 msgstr ""

-#: changedetectionio/blueprint/settings/templates/settings_llm_tab.html changedetectionio/forms.py
-msgid "Max tokens per watch per period"
+#: changedetectionio/forms.py
+msgid "Max tokens per check"
+msgstr ""
+
+#: changedetectionio/forms.py
+msgid "Max cumulative tokens (per watch)"
 msgstr ""

 #: changedetectionio/forms.py
@@ -3251,10 +3245,6 @@ 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 ""
@@ -846,10 +846,6 @@ 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 ""
@@ -1100,12 +1096,6 @@ 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 ""
@@ -3179,8 +3169,12 @@ msgstr ""
 msgid "Default AI Change Summary prompt"
 msgstr ""

-#: changedetectionio/blueprint/settings/templates/settings_llm_tab.html changedetectionio/forms.py
-msgid "Max tokens per watch per period"
+#: changedetectionio/forms.py
+msgid "Max tokens per check"
+msgstr ""
+
+#: changedetectionio/forms.py
+msgid "Max cumulative tokens (per watch)"
 msgstr ""

 #: changedetectionio/forms.py
@@ -3191,10 +3185,6 @@ 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 ""
@@ -842,10 +842,6 @@ 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 ""
@@ -1096,12 +1092,6 @@ 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 ""
@@ -3168,8 +3158,12 @@ msgstr ""
 msgid "Default AI Change Summary prompt"
 msgstr ""

-#: changedetectionio/blueprint/settings/templates/settings_llm_tab.html changedetectionio/forms.py
-msgid "Max tokens per watch per period"
+#: changedetectionio/forms.py
+msgid "Max tokens per check"
+msgstr ""
+
+#: changedetectionio/forms.py
+msgid "Max cumulative tokens (per watch)"
 msgstr ""

 #: changedetectionio/forms.py
@@ -3180,10 +3174,6 @@ 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 ""
@@ -847,10 +847,6 @@ 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 ""
@@ -1101,12 +1097,6 @@ 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 ""
@@ -3185,8 +3175,12 @@ msgstr ""
 msgid "Default AI Change Summary prompt"
 msgstr ""

-#: changedetectionio/blueprint/settings/templates/settings_llm_tab.html changedetectionio/forms.py
-msgid "Max tokens per watch per period"
+#: changedetectionio/forms.py
+msgid "Max tokens per check"
+msgstr ""
+
+#: changedetectionio/forms.py
+msgid "Max cumulative tokens (per watch)"
 msgstr ""

 #: changedetectionio/forms.py
@@ -3197,10 +3191,6 @@ 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 ""
@@ -844,10 +844,6 @@ 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"
@@ -1102,12 +1098,6 @@ 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."
@@ -3176,9 +3166,13 @@ msgstr ""
 msgid "Default AI Change Summary prompt"
 msgstr "기본 AI 변경 요약 프롬프트"

-#: changedetectionio/blueprint/settings/templates/settings_llm_tab.html changedetectionio/forms.py
-msgid "Max tokens per watch per period"
-msgstr ""
+#: 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/forms.py
 msgid "Monthly token budget"
@@ -3188,10 +3182,6 @@ 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"
@@ -6,9 +6,9 @@
 #, fuzzy
 msgid ""
 msgstr ""
-"Project-Id-Version: changedetection.io 0.55.5\n"
+"Project-Id-Version: changedetection.io 0.55.4\n"
 "Report-Msgid-Bugs-To: EMAIL@ADDRESS\n"
-"POT-Creation-Date: 2026-05-25 10:09+0200\n"
+"POT-Creation-Date: 2026-05-19 11:38+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"
@@ -839,10 +839,6 @@ 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 ""
@@ -1093,12 +1089,6 @@ 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 ""
@@ -3165,8 +3155,12 @@ msgstr ""
 msgid "Default AI Change Summary prompt"
 msgstr ""

-#: changedetectionio/blueprint/settings/templates/settings_llm_tab.html changedetectionio/forms.py
-msgid "Max tokens per watch per period"
+#: changedetectionio/forms.py
+msgid "Max tokens per check"
+msgstr ""
+
+#: changedetectionio/forms.py
+msgid "Max cumulative tokens (per watch)"
 msgstr ""

 #: changedetectionio/forms.py
@@ -3177,10 +3171,6 @@ 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 ""
@@ -865,10 +865,6 @@ 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 ""
@@ -1119,12 +1115,6 @@ 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 ""
@@ -3216,8 +3206,12 @@ msgstr ""
 msgid "Default AI Change Summary prompt"
 msgstr ""

-#: changedetectionio/blueprint/settings/templates/settings_llm_tab.html changedetectionio/forms.py
-msgid "Max tokens per watch per period"
+#: changedetectionio/forms.py
+msgid "Max tokens per check"
+msgstr ""
+
+#: changedetectionio/forms.py
+msgid "Max cumulative tokens (per watch)"
 msgstr ""

 #: changedetectionio/forms.py
@@ -3228,10 +3222,6 @@ 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 ""
@@ -875,10 +875,6 @@ 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 ""
@@ -1129,12 +1125,6 @@ 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 ""
@@ -3219,8 +3209,12 @@ msgstr ""
 msgid "Default AI Change Summary prompt"
 msgstr ""

-#: changedetectionio/blueprint/settings/templates/settings_llm_tab.html changedetectionio/forms.py
-msgid "Max tokens per watch per period"
+#: changedetectionio/forms.py
+msgid "Max tokens per check"
+msgstr ""
+
+#: changedetectionio/forms.py
+msgid "Max cumulative tokens (per watch)"
 msgstr ""

 #: changedetectionio/forms.py
@@ -3231,10 +3225,6 @@ 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 ""
@@ -855,10 +855,6 @@ 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 ""
@@ -1109,12 +1105,6 @@ 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 ""
@@ -3198,8 +3188,12 @@ msgstr ""
 msgid "Default AI Change Summary prompt"
 msgstr ""

-#: changedetectionio/blueprint/settings/templates/settings_llm_tab.html changedetectionio/forms.py
-msgid "Max tokens per watch per period"
+#: changedetectionio/forms.py
+msgid "Max tokens per check"
+msgstr ""
+
+#: changedetectionio/forms.py
+msgid "Max cumulative tokens (per watch)"
 msgstr ""

 #: changedetectionio/forms.py
@@ -3210,10 +3204,6 @@ 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 ""
@@ -844,10 +844,6 @@ 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 ""
@@ -1098,12 +1094,6 @@ 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 ""
@@ -3171,8 +3161,12 @@ msgstr ""
 msgid "Default AI Change Summary prompt"
 msgstr ""

-#: changedetectionio/blueprint/settings/templates/settings_llm_tab.html changedetectionio/forms.py
-msgid "Max tokens per watch per period"
+#: changedetectionio/forms.py
+msgid "Max tokens per check"
+msgstr ""
+
+#: changedetectionio/forms.py
+msgid "Max cumulative tokens (per watch)"
 msgstr ""

 #: changedetectionio/forms.py
@@ -3183,10 +3177,6 @@ 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 ""
@@ -843,10 +843,6 @@ 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 ""
@@ -1097,12 +1093,6 @@ 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 ""
@@ -3170,8 +3160,12 @@ msgstr ""
 msgid "Default AI Change Summary prompt"
 msgstr ""

-#: changedetectionio/blueprint/settings/templates/settings_llm_tab.html changedetectionio/forms.py
-msgid "Max tokens per watch per period"
+#: changedetectionio/forms.py
+msgid "Max tokens per check"
+msgstr ""
+
+#: changedetectionio/forms.py
+msgid "Max cumulative tokens (per watch)"
 msgstr ""

 #: changedetectionio/forms.py
@@ -3182,10 +3176,6 @@ 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 ""
@@ -432,15 +432,9 @@ 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.
-                        # 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':
+                        # 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':
                            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)")
@@ -450,14 +444,9 @@ 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, _runtime_llm_config,
+                                    summarise_change, get_llm_config,
                                )
-                                # _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)
+                                _llm_cfg = get_llm_config(datastore)
                                if _llm_cfg:
                                    # Compute unified diff once — used by both intent and summary
                                    _watch_dates = list(watch.history.keys())
@@ -549,10 +538,8 @@ 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]
-                                    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_max_summary_tokens = datastore.data['settings']['application'].get('llm_max_summary_tokens', 3000)
+                                    _llm_model = (datastore.data['settings']['application'].get('llm') or {}).get('model', '')
                                    _llm_cache_prompt = build_summary_cache_prompt(
                                        effective_prompt=get_effective_summary_prompt(watch, datastore),
                                        max_summary_tokens=_llm_max_summary_tokens,
@@ -621,14 +621,6 @@ 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
@@ -148,9 +148,6 @@ 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