UI - LLM - Flag LLM_FEATURES_DISABLED to disable all LLM from the UI/system

changedetectionio/blueprint/settings/__init__.py

@@ -14,8 +14,10 @@ from changedetectionio.auth_decorator import login_optionally_required
 
 def construct_blueprint(datastore: ChangeDetectionStore):
     from changedetectionio.blueprint.settings.llm import construct_llm_blueprint
+    from changedetectionio.llm.evaluator import is_llm_features_disabled
     settings_blueprint = Blueprint('settings', __name__, template_folder="templates")
-    settings_blueprint.register_blueprint(construct_llm_blueprint(datastore), url_prefix='/llm')
+    if not is_llm_features_disabled():
+        settings_blueprint.register_blueprint(construct_llm_blueprint(datastore), url_prefix='/llm')
 
     @settings_blueprint.route("", methods=['GET', "POST"])
     @login_optionally_required

changedetectionio/blueprint/settings/templates/settings.html

@@ -34,7 +34,9 @@
             <li class="tab"><a href="#plugin-{{ tab.plugin_id }}">{{ tab.tab_label }}</a></li>
                 {% endfor %}
             {% endif %}
+            {% if not llm_features_disabled %}
             <li class="tab"><a href="#ai">{{ _('AI / LLM') }}</a></li>
+            {% endif %}
             <li class="tab"><a href="#info">{{ _('Info') }}</a></li>
         </ul>
     </div>
@@ -394,7 +396,9 @@ nav
             </div>
                 {% endfor %}
             {% endif %}
+            {% if not llm_features_disabled %}
             {% include 'settings_llm_tab.html' %}
+            {% endif %}
             <div class="tab-pane-inner" id="info">
                 <p><strong>{{ _('Uptime:') }}</strong> {{ uptime_seconds|format_duration }}</p>
                 <p><strong>{{ _('Python version:') }}</strong> {{ python_version }}</p>

changedetectionio/blueprint/ui/templates/edit.html

@@ -57,7 +57,9 @@
             {% if capabilities.supports_visual_selector %}
             <li class="tab"><a id="visualselector-tab" href="#visualselector">{{ _('Visual Filter Selector') }}</a></li>
             {% endif %}
+            {% if not llm_features_disabled %}
             <li class="tab"><a href="#ai-llm">{{ _('AI / LLM') }}</a></li>
+            {% endif %}
             {% if capabilities.supports_text_filters_and_triggers %}
             <li class="tab" id="filters-and-triggers-tab"><a href="#filters-and-triggers">{{ _('Filters & Triggers') }}</a></li>
             <li class="tab" id="conditions-tab"><a href="#conditions">{{ _('Conditions') }}</a></li>
@@ -321,9 +323,11 @@ Math: {{ 1 + 1 }}") }}
                     </div>
                 </div>
             </div>
+            {% if not llm_features_disabled %}
             <div class="tab-pane-inner" id="ai-llm">
                 {% include "edit/include_llm_intent.html" %}
             </div>
+            {% endif %}
             <div class="tab-pane-inner" id="filters-and-triggers">
 
                 <span id="activate-text-preview" class="pure-button pure-button-primary button-xsmall">{{ _('Activate preview') }}</span>
@@ -503,7 +507,7 @@ Math: {{ 1 + 1 }}") }}
                             <td>{{ _('Server type reply') }}</td>
                             <td>{{ watch.get('remote_server_reply') }}</td>
                         </tr>
-                        {% if settings_application.get('llm', {}).get('model') %}
+                        {% if not llm_features_disabled and settings_application.get('llm', {}).get('model') %}
                         <tr>
                             <td>{{ _('AI tokens (last check)') }}</td>
                             <td>{{ "{:,}".format(watch.get('llm_last_tokens_used') or 0) }}</td>

changedetectionio/flask_app.py

@@ -522,6 +522,11 @@ def changedetection_app(config=None, datastore_o=None):
         available_languages=available_languages
     )
 
+    @app.context_processor
+    def inject_llm_features_disabled():
+        from changedetectionio.llm.evaluator import is_llm_features_disabled
+        return dict(llm_features_disabled=is_llm_features_disabled())
+
     # Set up a request hook to check authentication for all routes
     @app.before_request
     def check_authentication():

changedetectionio/llm/evaluator.py

@@ -20,6 +20,8 @@ from dataclasses import dataclass
 from datetime import datetime, timezone
 from loguru import logger
 
+from changedetectionio.strtobool import strtobool
+
 from . import client as llm_client
 from .prompt_builder import (
     build_change_summary_prompt, build_change_summary_system_prompt,
@@ -31,6 +33,11 @@ from .response_parser import parse_eval_response, parse_preview_response, parse_
 
 _DEFAULT_MAX_INPUT_CHARS = 100_000
 
+
+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_max_input_chars(datastore) -> int:
     """Max input characters to send to the LLM. Resolution: env var → datastore → 100,000.
     Always returns at least 1 — unlimited is not permitted.
@@ -207,6 +214,8 @@ def get_llm_config(datastore) -> dict | None:
       1. Environment variables: LLM_MODEL, LLM_API_KEY, LLM_API_BASE
       2. Datastore settings (set via UI)
     """
+    if is_llm_features_disabled():
+        return None
     # 1. Environment variable override
     env_model = os.getenv('LLM_MODEL', '').strip()
     if env_model:
@@ -225,6 +234,8 @@ 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())
 
 

changedetectionio/processors/text_json_diff/__init__.py

@@ -35,6 +35,50 @@ def _task(watch, update_handler):
     return text_after_filter
 
 
+def _compute_ignore_line_numbers_for_preview(text_pre_extract, ignore_patterns, extract_patterns):
+    """1-indexed output line numbers in the post-extract display that correspond
+    to input lines matching ignore_text patterns.
+
+    Needed because extract_text (#4138) transforms line content — e.g. "0.54.10"
+    becomes ".54.10" — so a substring match for "0.54.10" against the post-extract
+    text fails and the preview UI can no longer mark the line as ignored. We find
+    the ignored line numbers in the pre-extract text and replay extract_by_regex
+    line-by-line to map them forward.
+    """
+    from changedetectionio import html_tools
+    from changedetectionio.processors.text_json_diff.processor import ContentTransformer
+
+    if not text_pre_extract or not ignore_patterns:
+        return []
+
+    ignored_input_lines = set(
+        html_tools.strip_ignore_text(
+            content=text_pre_extract,
+            wordlist=ignore_patterns,
+            mode='line numbers'
+        )
+    )
+    if not ignored_input_lines:
+        return []
+
+    if not extract_patterns:
+        return sorted(ignored_input_lines)
+
+    # Replay extract_by_regex per-line. Each emitted match ends with exactly one
+    # '\n', so counting newlines tells us how many output lines this input produced.
+    output_line_counter = 0
+    result = []
+    for input_idx, line in enumerate(text_pre_extract.splitlines()):
+        is_ignored = (input_idx + 1) in ignored_input_lines
+        matches_in_line = ContentTransformer.extract_by_regex(line, extract_patterns).count('\n')
+        for _ in range(matches_in_line):
+            output_line_counter += 1
+            if is_ignored:
+                result.append(output_line_counter)
+
+    return result
+
+
 def prepare_filter_prevew(datastore, watch_uuid, form_data):
     '''Used by @app.route("/edit/<uuid_str:uuid>/preview-rendered", methods=['POST'])'''
     from changedetectionio import forms, html_tools
@@ -50,6 +94,7 @@ def prepare_filter_prevew(datastore, watch_uuid, form_data):
 
     text_after_filter = ''
     text_before_filter = ''
+    text_pre_extract = ''
     trigger_line_numbers = []
     ignore_line_numbers = []
     blocked_line_numbers = []
@@ -89,15 +134,22 @@ def prepare_filter_prevew(datastore, watch_uuid, form_data):
             update_handler.fetcher.content = str(decompressed_data) # str() because playwright/puppeteer/requests return string
             update_handler.fetcher.headers['content-type'] = tmp_watch.get('content-type')
 
-            # Process our watch with filters and the HTML from disk, and also a blank watch with no filters but also with the same HTML from disk
+            # Process our watch with filters and the HTML from disk, and also a blank watch with no filters but also with the same HTML from disk.
+            # The third task runs with extract_text cleared so we can compute ignore_line_numbers
+            # against the pre-extract text (extract_text transforms lines so post-extract substring
+            # matching for ignore patterns would otherwise fail — see #4138 follow-up).
             # Do this as parallel threads (not processes) to avoid pickle issues with Lock objects
+            tmp_watch_no_extract = deepcopy(tmp_watch)
+            tmp_watch_no_extract['extract_text'] = []
             try:
-                with ThreadPoolExecutor(max_workers=2) as executor:
+                with ThreadPoolExecutor(max_workers=3) as executor:
                     future1 = executor.submit(_task, tmp_watch, update_handler)
                     future2 = executor.submit(_task, blank_watch_no_filters, update_handler)
+                    future3 = executor.submit(_task, tmp_watch_no_extract, update_handler)
 
                     text_after_filter = future1.result()
                     text_before_filter = future2.result()
+                    text_pre_extract = future3.result()
             except Exception as e:
                 x=1
 
@@ -111,10 +163,11 @@ def prepare_filter_prevew(datastore, watch_uuid, form_data):
 
     try:
         text_to_ignore = tmp_watch.get('ignore_text', []) + datastore.data['settings']['application'].get('global_ignore_text', [])
-        ignore_line_numbers = html_tools.strip_ignore_text(content=text_after_filter,
-                                                           wordlist=text_to_ignore,
-                                                           mode='line numbers'
-                                                           )
+        ignore_line_numbers = _compute_ignore_line_numbers_for_preview(
+            text_pre_extract=text_pre_extract,
+            ignore_patterns=text_to_ignore,
+            extract_patterns=tmp_watch.get('extract_text', [])
+        )
     except Exception as e:
         text_before_filter = f"Error: {str(e)}"
 

changedetectionio/static/js/watch-settings.js

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

changedetectionio/templates/_common_fields.html

@@ -112,7 +112,7 @@
                 <td><code>{{ '{{triggered_text}}' }}</code></td>
                 <td>{{ _('Text that tripped the trigger from filters') }}</td>
             </tr>
-            {% if settings_application and settings_application.get('llm', {}).get('model') %}
+            {% if not llm_features_disabled and settings_application and settings_application.get('llm', {}).get('model') %}
             <tr>
                 <td><code>{{ '{{diff}}' }}</code> <small style="opacity:0.6">{{ _('(upgraded)') }}</small></td>
                 <td>{{ _('When AI Change Summary is configured, contains the AI-generated description instead of the raw diff. Falls back to raw diff when not configured.') }}</td>

changedetectionio/templates/base.html

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

changedetectionio/templates/menu.html

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

changedetectionio/tests/llm/test_llm_features_disabled.py

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

changedetectionio/tests/test_live_preview.py

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

changedetectionio/worker.py

@@ -436,7 +436,8 @@ async def async_update_worker(worker_id, q, notification_q, app, datastore, exec
                         # 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.
-                        _llm_master_enabled = bool(datastore.data['settings']['application'].get('llm_enabled', True))
+                        from changedetectionio.llm.evaluator import is_llm_features_disabled as _is_llm_features_disabled
+                        _llm_master_enabled = bool(datastore.data['settings']['application'].get('llm_enabled', True)) and not _is_llm_features_disabled()
                         _llm_budget_action = datastore.data['settings']['application'].get('llm_budget_action', 'skip_llm')
                         if _llm_master_enabled and _llm_budget_action == 'skip_check':
                             from changedetectionio.llm.evaluator import is_global_token_budget_exceeded