LLM - UI & Ollama tweaks (#4148)

changedetectionio/blueprint/settings/__init__.py

@@ -41,6 +41,7 @@ def construct_blueprint(datastore: ChangeDetectionStore):
             '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)),
@@ -125,6 +126,9 @@ def construct_blueprint(datastore: ChangeDetectionStore):
                 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'
                 )

changedetectionio/blueprint/settings/llm.py

@@ -113,23 +113,43 @@ def construct_llm_blueprint(datastore: ChangeDetectionStore):
     @llm_blueprint.route("/test", methods=['GET'])
     @login_optionally_required
     def llm_test():
+        from flask import request
         from changedetectionio.llm.client import completion
 
-        llm_cfg = datastore.data['settings']['application'].get('llm') or {}
-        model    = llm_cfg.get('model', '').strip()
-        api_base = llm_cfg.get('api_base', '') or ''
+        # Pull stored config as the fallback, then override with anything the
+        # form-driven JS sent as query params. Lets users test config changes
+        # without first hitting Save (matching how /settings/llm/models works).
+        stored = datastore.data['settings']['application'].get('llm') or {}
+        llm_cfg = {
+            'model':                   (request.args.get('model')                   or stored.get('model', '')).strip(),
+            'api_key':                 (request.args.get('api_key')                 or stored.get('api_key', '')).strip(),
+            'api_base':                (request.args.get('api_base')                or stored.get('api_base', '')).strip(),
+            'provider_kind':           (request.args.get('provider_kind')           or stored.get('provider_kind', '')).strip(),
+            'local_token_multiplier':   request.args.get('local_token_multiplier')  or stored.get('local_token_multiplier'),
+        }
+        model    = llm_cfg['model']
+        api_base = llm_cfg['api_base']
 
-        logger.debug(f"LLM connection test requested: model={model!r} api_base={api_base!r}")
+        logger.debug(
+            f"LLM connection test requested: model={model!r} api_base={api_base!r} "
+            f"provider_kind={llm_cfg['provider_kind']!r} "
+            f"source={'form' if request.args.get('model') else 'datastore'}"
+        )
 
         if not model:
-            logger.error("LLM connection test failed: no model configured in datastore")
+            logger.error("LLM connection test failed: no model configured")
             return jsonify({'ok': False, 'error': 'No model configured.'}), 400
 
         try:
             logger.debug(f"LLM connection test: sending test prompt to model={model!r}")
             # Reuse the same multiplier path the production calls use, so cloud providers
             # stay on a small base cap (matching upstream's pre-existing behavior) and only
-            # 'openai_compatible' endpoints opt into the reasoning-friendly headroom.
+            # reasoning-capable endpoints (Ollama, openai_compatible) opt into the extra
+            # headroom needed for chain-of-thought to complete.
+            # Timeout: omit the override so the test inherits DEFAULT_TIMEOUT (60s, tunable
+            # via LLM_TIMEOUT). A shorter test-only timeout falsely fails on cold-starting
+            # cloud reasoning models (e.g. ollama.com hosting qwen3.5:397b takes ~60s on
+            # first hit) even though the same call succeeds in production.
             from changedetectionio.llm.evaluator import apply_local_token_multiplier
             text, total_tokens, input_tokens, output_tokens = completion(
                 model=model,
@@ -137,8 +157,8 @@ def construct_llm_blueprint(datastore: ChangeDetectionStore):
                     'Respond with just the word: ready'}],
                 api_key=llm_cfg.get('api_key') or None,
                 api_base=api_base or None,
-                timeout=30,
                 max_tokens=apply_local_token_multiplier(200, llm_cfg),
+                debug=bool(datastore.data['settings']['application'].get('llm_debug', False)),
             )
             reply = text.strip()
             if not reply:

changedetectionio/blueprint/settings/templates/settings_llm_tab.html

@@ -123,14 +123,15 @@
   </div>
 
   {# Hidden field carrying the dropdown selection so the backend knows when to apply
-     reasoning-friendly token caps (only for self-hosted OpenAI-compatible endpoints). #}
+     reasoning-friendly token caps (Ollama and OpenAI-compatible endpoints, which commonly
+     serve reasoning models that need headroom for chain-of-thought to complete). #}
   {{ form.llm.form.llm_provider_kind() }}
 
   <div class="pure-control-group" id="llm-local-advanced-group" style="display:none">
     <label for="{{ form.llm.form.llm_local_token_multiplier.id }}">{{ form.llm.form.llm_local_token_multiplier.label.text }}</label>
     {{ form.llm.form.llm_local_token_multiplier() }}
     <span class="pure-form-message-inline">
-      {{ _('Local 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, lower it if you want tighter limits. Only applied to self-hosted OpenAI-compatible endpoints — cloud providers (OpenAI, Anthropic, Gemini) keep their original tight caps.', default='5x') | safe }}
+      {{ _('Reasoning models (Qwen3, DeepSeek-R1, Gemma 3, etc.) emit chain-of-thought before the final answer. This multiplier scales every <code>max_tokens</code> cap for this endpoint to leave reasoning room. Defaults to %(default)s; raise it if responses come back truncated or empty, lower it (down to 1x) if you want tighter limits on a paid endpoint. Applied to Ollama and OpenAI-compatible endpoints — other cloud providers (OpenAI, Anthropic, Gemini) keep their original tight caps.', default='5x') | safe }}
     </span>
   </div>
 
@@ -201,6 +202,17 @@
     </a>
     <span class="pure-form-message-inline">{{ _('Removes all cached AI change summaries across all watches. They will be regenerated on the next check.') }}</span>
   </div>
+
+  <div class="pure-control-group">
+    <label></label>
+    {{ form.llm.form.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.') }}
+    </span>
+  </div>
   {% endif %}{# llm_env_configured #}
 
   {% if not llm_env_configured and not (llm_config and llm_config.get('model')) %}
@@ -417,13 +429,14 @@
     }
 
     // Persist the dropdown selection so the backend can branch on provider kind
-    // (currently only 'openai_compatible' triggers the local-multiplier code path).
+    // (self-hosted endpoints — 'ollama' and 'openai_compatible' — trigger the
+    // local-multiplier code path; cloud providers do not).
     if (kindField) kindField.value = provider || '';
 
-    // Show the local-endpoint advanced settings (token multiplier) only for the
-    // OpenAI-compatible self-hosted option. Cloud providers and Ollama get the
-    // original tight caps and don't see this section at all.
-    if (localAdvGrp) localAdvGrp.style.display = (provider === 'openai_compatible') ? '' : 'none';
+    // Show the local-endpoint advanced settings (token multiplier) for self-hosted
+    // endpoints. Cloud providers get the original tight caps and don't see this
+    // section at all.
+    if (localAdvGrp) localAdvGrp.style.display = (provider === 'ollama' || provider === 'openai_compatible') ? '' : 'none';
 
     hint.textContent = KEY_HINTS[provider] || '';
     modelSelGrp.style.display = 'none';
@@ -502,8 +515,23 @@
     btn.textContent = '⏳ {{ _("Testing…") }}';
     result.style.display = 'none';
 
+    // Send the form's current values so the user doesn't have to hit Save before
+    // testing a config change. Endpoint falls back to the stored datastore values
+    // for any field we don't send.
+    const params = new URLSearchParams();
+    const model   = (document.querySelector('[name="llm-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());
+    if (kind.trim())    params.set('provider_kind', kind.trim());
+    if (mult.trim())    params.set('local_token_multiplier', mult.trim());
+
     try {
-      const resp = await fetch('{{ url_for("settings.llm.llm_test") }}');
+      const resp = await fetch('{{ url_for("settings.llm.llm_test") }}?' + params);
       const data = await resp.json();
       if (data.ok) {
         result.style.cssText = 'display:block; background:rgba(39,174,96,0.08); border:1px solid rgba(39,174,96,0.3); border-radius:5px; padding:0.6em 0.85em; font-size:0.88em; line-height:1.45;';
@@ -519,7 +547,7 @@
       result.innerHTML = '<span style="color:#c0392b; font-weight:600;">&#10007; {{ _("Request failed") }}</span>: ' + e.message.replace(/</g,'&lt;');
     } finally {
       btn.disabled = false;
-      btn.textContent = '&#9654; {{ _("Test connection") }}';
+      btn.textContent = '▶ {{ _("Test connection") }}';
     }
   };
 

changedetectionio/blueprint/ui/diff.py

@@ -270,12 +270,15 @@ def construct_blueprint(datastore: ChangeDetectionStore):
             LLMInputTooLargeError,
         )
 
-        # Diff-pref flags + system prompt are part of the cache key so prompt changes bust the cache.
+        # Diff-pref flags + system prompt + active model are part of the cache key
+        # so prompt or model changes bust the cache.
         _max_summary_tokens = datastore.data['settings']['application'].get('llm_max_summary_tokens', 3000)
+        _llm_model = (datastore.data['settings']['application'].get('llm') or {}).get('model', '')
         cache_prompt = build_summary_cache_prompt(
             effective_prompt=get_effective_summary_prompt(watch, datastore),
             max_summary_tokens=_max_summary_tokens,
             prefs=prefs,
+            model=_llm_model,
         )
 
         # Check cache — keyed by version pair + prompt hash (invalidates if prompt changes)

changedetectionio/forms.py

@@ -1125,11 +1125,14 @@ class globalSettingsLLMForm(Form):
         validators=[validators.Optional()],
         default='',
     )
-    # Multiplier applied to LLM max_tokens caps when provider_kind == 'openai_compatible'.
-    # Reasoning models (Qwen3, DeepSeek-R1, Gemma 3, etc.) emit chain-of-thought into
+    # Multiplier applied to LLM max_tokens caps when provider_kind is 'ollama' or
+    # 'openai_compatible' — endpoints that commonly serve reasoning models (Qwen3,
+    # DeepSeek-R1, Gemma 3, etc.) which emit chain-of-thought into
     # message.reasoning_content before the final answer lands in message.content.
-    # Local self-hosted models cost no per-token money, so giving them headroom is cheap;
-    # cloud providers stay on the original tight caps so existing users see no cost change.
+    # Cloud providers with non-reasoning defaults (OpenAI, Anthropic, Gemini,
+    # OpenRouter) stay on the original tight caps so existing users see no
+    # behavior or cost change. Users on paid Ollama / openai_compatible endpoints
+    # who care about cost can dial this down to 1x.
     llm_local_token_multiplier = IntegerField(
         _l('Token multiplier for local reasoning models'),
         validators=[validators.Optional(), validators.NumberRange(min=1, max=20)],
@@ -1187,6 +1190,10 @@ class globalSettingsLLMForm(Form):
         _l('Use LLM as a fallback for extracting price and restock info'),
         default=True,
     )
+    llm_debug = BooleanField(
+        _l('Enable LLM debug logging'),
+        default=False,
+    )
     llm_thinking_budget = SelectField(
         _l('AI thinking budget (tokens)'),
         choices=[

changedetectionio/llm/client.py

@@ -4,6 +4,7 @@ Keeps litellm import isolated so the rest of the codebase doesn't depend on it d
 and makes the call easy to mock in tests.
 """
 
+import logging
 import os
 from loguru import logger
 
@@ -17,9 +18,46 @@ DEFAULT_TIMEOUT = int(os.getenv('LLM_TIMEOUT', 60))
 DEFAULT_RETRIES = 3
 
 
+class _LoguruInterceptHandler(logging.Handler):
+    # Routes litellm's stdlib log records through loguru so debug output
+    # uses the same format/sink as the rest of the app.
+    def emit(self, record):
+        try:
+            level = logger.level(record.levelname).name
+        except (ValueError, AttributeError):
+            level = record.levelno
+        logger.opt(exception=record.exc_info).log(level, record.getMessage())
+
+
+_debug_installed = False
+
+
+def _install_litellm_debug():
+    # Attach our loguru intercept and clear any pre-existing handlers so litellm's
+    # own stdout StreamHandler (installed by _turn_on_debug / set_verbose) doesn't
+    # double-emit. Setting the logger level to DEBUG is enough to make litellm
+    # produce debug records — we don't call _turn_on_debug() for that reason.
+    global _debug_installed
+    if _debug_installed:
+        return
+
+    handler = _LoguruInterceptHandler()
+    handler.setLevel(logging.DEBUG)
+    for _name in ('LiteLLM', 'litellm', 'litellm.utils', 'litellm.router'):
+        _lg = logging.getLogger(_name)
+        _lg.handlers = []
+        _lg.setLevel(logging.DEBUG)
+        _lg.addHandler(handler)
+        _lg.propagate = False
+
+    _debug_installed = True
+    logger.info("LLM client: litellm debug logging routed through loguru")
+
+
 def completion(model: str, messages: list, api_key: str = None,
                api_base: str = None, timeout: int = DEFAULT_TIMEOUT,
-               max_tokens: int = None, extra_body: dict = None) -> tuple[str, int, int, int]:
+               max_tokens: int = None, extra_body: dict = None,
+               debug: bool = False) -> tuple[str, int, int, int]:
     """
     Call the LLM and return (response_text, total_tokens, input_tokens, output_tokens).
     Retries up to DEFAULT_RETRIES times on timeout or connection errors.
@@ -31,6 +69,9 @@ def completion(model: str, messages: list, api_key: str = None,
     except ImportError:
         raise RuntimeError("litellm is not installed. Add it to requirements.txt.")
 
+    if debug:
+        _install_litellm_debug()
+
     _timeout = timeout if timeout is not None else DEFAULT_TIMEOUT
 
     kwargs = {
@@ -49,7 +90,10 @@ def completion(model: str, messages: list, api_key: str = None,
 
     _retryable = (litellm.Timeout, litellm.APIConnectionError)
 
-    logger.trace("Sending payload to LLM.. ")
+    logger.debug(
+        f"LLM client: calling model={model!r} api_base={api_base!r} "
+        f"timeout={_timeout}s max_tokens={kwargs['max_tokens']}"
+    )
     logger.trace(messages)
 
     for attempt in range(1, DEFAULT_RETRIES + 1):

changedetectionio/llm/evaluator.py

@@ -120,22 +120,25 @@ def _summary_max_tokens(diff: str, max_cap: int = LLM_DEFAULT_MAX_SUMMARY_TOKENS
 
 def apply_local_token_multiplier(base_max_tokens: int, llm_cfg: dict) -> int:
     """
-    Scale max_tokens for self-hosted OpenAI-compatible endpoints (vLLM, LM Studio, llama.cpp).
+    Scale max_tokens for endpoints that commonly serve reasoning models
+    (Ollama — self-hosted or ollama.com cloud — and OpenAI-compatible servers like
+    vLLM, LM Studio, llama.cpp).
 
     Reasoning models (Qwen3, DeepSeek-R1, Gemma 3, etc.) emit chain-of-thought into
     `message.reasoning_content` BEFORE the final answer lands in `message.content`.
-    Without enough headroom the request truncates mid-thought (`finish_reason='length'`)
-    and the answer never lands — callers see an empty string and silently fall through
-    to safe defaults, hiding the problem.
+    Without enough headroom the request truncates mid-thought (`finish_reason='length'`
+    or `'stop'` with empty content) and the answer never lands — callers see an empty
+    string and silently fall through to safe defaults, hiding the problem.
 
-    Local self-hosted models cost no per-token money, so headroom is cheap; cloud
-    providers (OpenAI, Anthropic, Gemini, OpenRouter) keep their original tight caps
-    so existing users see no cost change.
+    Cloud providers with stable, non-reasoning defaults (OpenAI, Anthropic, Gemini,
+    OpenRouter) keep their original tight caps so existing users see no behavior or
+    cost change. Ollama / OpenAI-compatible users can dial the multiplier down to 1x
+    in Settings → AI → Provider if they want to keep costs tight on a paid endpoint.
 
-    Activated only when `llm_cfg['provider_kind'] == 'openai_compatible'`.
+    Activated when `llm_cfg['provider_kind']` is `'ollama'` or `'openai_compatible'`.
     Multiplier defaults to 5x and is user-configurable in Settings → AI → Provider.
     """
-    if (llm_cfg or {}).get('provider_kind') != 'openai_compatible':
+    if (llm_cfg or {}).get('provider_kind') not in ('ollama', 'openai_compatible'):
         return base_max_tokens
     try:
         multiplier = int(llm_cfg.get('local_token_multiplier') or 5)
@@ -399,6 +402,7 @@ def run_setup(watch, datastore, snapshot_text: str) -> None:
             api_base=cfg.get('api_base'),
             max_tokens=apply_local_token_multiplier(JSON_RESPONSE_MAX_TOKENS, cfg),
             extra_body=_thinking_extra_body(cfg['model'], int(datastore.data['settings']['application'].get('llm_thinking_budget', LLM_DEFAULT_THINKING_BUDGET) or 0)),
+            debug=bool(datastore.data['settings']['application'].get('llm_debug', False)),
         )
         _check_token_budget(watch, cfg, tokens)
         accumulate_global_tokens(datastore, tokens, model=cfg['model'])
@@ -472,7 +476,7 @@ class DiffPrefs:
 
 
 def build_summary_cache_prompt(effective_prompt: str, max_summary_tokens: int,
-                                prefs: DiffPrefs = None) -> str:
+                                prefs: DiffPrefs = None, model: str = '') -> str:
     """
     Compose the full cache-key string passed to save/get_llm_diff_summary.
 
@@ -480,6 +484,10 @@ def build_summary_cache_prompt(effective_prompt: str, max_summary_tokens: int,
     worker-side pre-cache is hit by an unmodified UI request. Same helper must
     be used by both the worker pre-cache write and the UI diff route read,
     otherwise the prompt hashes diverge and the cache file isn't found.
+
+    The active model name is folded into the key so switching models
+    (e.g. qwen3 → gpt-4o) invalidates stale summaries that were generated
+    by a different model with potentially different phrasing/quality.
     """
     if prefs is None:
         prefs = DiffPrefs()
@@ -488,6 +496,7 @@ def build_summary_cache_prompt(effective_prompt: str, max_summary_tokens: int,
         + prefs.cache_key_suffix()
         + f'\x00sys:{build_change_summary_system_prompt()}'
         + f'\x00max_tokens:{max_summary_tokens}'
+        + f'\x00model:{model}'
     )
 
 
@@ -551,6 +560,7 @@ def summarise_change(watch, datastore, diff: str, current_snapshot: str = '') ->
                 cfg,
             ),
             extra_body=_extra_body,
+            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
@@ -613,6 +623,7 @@ def preview_extract(watch, datastore, content: str) -> dict | None:
             api_base=cfg.get('api_base'),
             max_tokens=apply_local_token_multiplier(JSON_RESPONSE_MAX_TOKENS, cfg),
             extra_body=_thinking_extra_body(cfg['model'], int(datastore.data['settings']['application'].get('llm_thinking_budget', LLM_DEFAULT_THINKING_BUDGET) or 0)),
+            debug=bool(datastore.data['settings']['application'].get('llm_debug', False)),
         )
         accumulate_global_tokens(datastore, tokens, model=cfg['model'])
         result = parse_preview_response(raw)
@@ -697,6 +708,7 @@ def evaluate_change(watch, datastore, diff: str, current_snapshot: str = '') ->
             api_base=cfg.get('api_base'),
             max_tokens=apply_local_token_multiplier(JSON_RESPONSE_MAX_TOKENS, cfg),
             extra_body=_thinking_extra_body(cfg['model'], int(datastore.data['settings']['application'].get('llm_thinking_budget', LLM_DEFAULT_THINKING_BUDGET) or 0)),
+            debug=bool(datastore.data['settings']['application'].get('llm_debug', False)),
         )
         raw, tokens = _resp[0], _resp[1]
         input_tokens  = _resp[2] if len(_resp) > 2 else 0

changedetectionio/processors/text_json_diff/difference.py

@@ -218,9 +218,11 @@ def render(watch, datastore, request, url_for, render_template, flash, redirect,
             # Must match the cache_prompt the worker writes and the UI ajax route reads —
             # using UI default diff prefs so the initial render finds the worker's pre-cache.
             _max_summary_tokens = datastore.data['settings']['application'].get('llm_max_summary_tokens', 3000)
+            _llm_model = (datastore.data['settings']['application'].get('llm') or {}).get('model', '')
             _cache_prompt = build_summary_cache_prompt(
                 effective_prompt=_prompt,
                 max_summary_tokens=_max_summary_tokens,
+                model=_llm_model,
             )
             llm_diff_summary = watch.get_llm_diff_summary(from_version, to_version, prompt=_cache_prompt)
         except Exception as e:

changedetectionio/translations/cs/LC_MESSAGES/messages.po

@@ -906,10 +906,11 @@ msgstr ""
 #: changedetectionio/blueprint/settings/templates/settings_llm_tab.html
 #, python-format
 msgid ""
-"Local 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, lower it if you want tighter limits. Only applied to self-"
-"hosted OpenAI-compatible endpoints — cloud providers (OpenAI, Anthropic, Gemini) keep their original tight caps."
+"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."
 msgstr ""
 
 #: changedetectionio/blueprint/settings/templates/settings_llm_tab.html
@@ -983,6 +984,12 @@ msgstr ""
 msgid "Removes all cached AI change summaries across all watches. They will be regenerated on the next check."
 msgstr ""
 
+#: changedetectionio/blueprint/settings/templates/settings_llm_tab.html
+msgid ""
+"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."
+msgstr ""
+
 #: changedetectionio/blueprint/settings/templates/settings_llm_tab.html
 msgid "Default AI Change Summary"
 msgstr ""
@@ -3173,6 +3180,10 @@ msgstr ""
 msgid "Use LLM as a fallback for extracting price and restock info"
 msgstr ""
 
+#: changedetectionio/forms.py
+msgid "Enable LLM debug logging"
+msgstr ""
+
 #: changedetectionio/forms.py
 msgid "AI thinking budget (tokens)"
 msgstr "AI pracovní rozpočet (tokeny)"

changedetectionio/translations/de/LC_MESSAGES/messages.po

@@ -922,10 +922,11 @@ msgstr ""
 #: changedetectionio/blueprint/settings/templates/settings_llm_tab.html
 #, python-format
 msgid ""
-"Local 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, lower it if you want tighter limits. Only applied to self-"
-"hosted OpenAI-compatible endpoints — cloud providers (OpenAI, Anthropic, Gemini) keep their original tight caps."
+"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."
 msgstr ""
 
 #: changedetectionio/blueprint/settings/templates/settings_llm_tab.html
@@ -999,6 +1000,12 @@ msgstr ""
 msgid "Removes all cached AI change summaries across all watches. They will be regenerated on the next check."
 msgstr ""
 
+#: changedetectionio/blueprint/settings/templates/settings_llm_tab.html
+msgid ""
+"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."
+msgstr ""
+
 #: changedetectionio/blueprint/settings/templates/settings_llm_tab.html
 msgid "Default AI Change Summary"
 msgstr ""
@@ -3225,6 +3232,10 @@ msgstr ""
 msgid "Use LLM as a fallback for extracting price and restock info"
 msgstr ""
 
+#: changedetectionio/forms.py
+msgid "Enable LLM debug logging"
+msgstr ""
+
 #: changedetectionio/forms.py
 msgid "AI thinking budget (tokens)"
 msgstr ""

changedetectionio/translations/en_GB/LC_MESSAGES/messages.po

@@ -904,10 +904,11 @@ msgstr ""
 #: changedetectionio/blueprint/settings/templates/settings_llm_tab.html
 #, python-format
 msgid ""
-"Local 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, lower it if you want tighter limits. Only applied to self-"
-"hosted OpenAI-compatible endpoints — cloud providers (OpenAI, Anthropic, Gemini) keep their original tight caps."
+"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."
 msgstr ""
 
 #: changedetectionio/blueprint/settings/templates/settings_llm_tab.html
@@ -981,6 +982,12 @@ msgstr ""
 msgid "Removes all cached AI change summaries across all watches. They will be regenerated on the next check."
 msgstr ""
 
+#: changedetectionio/blueprint/settings/templates/settings_llm_tab.html
+msgid ""
+"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."
+msgstr ""
+
 #: changedetectionio/blueprint/settings/templates/settings_llm_tab.html
 msgid "Default AI Change Summary"
 msgstr ""
@@ -3167,6 +3174,10 @@ msgstr ""
 msgid "Use LLM as a fallback for extracting price and restock info"
 msgstr ""
 
+#: changedetectionio/forms.py
+msgid "Enable LLM debug logging"
+msgstr ""
+
 #: changedetectionio/forms.py
 msgid "AI thinking budget (tokens)"
 msgstr ""

changedetectionio/translations/en_US/LC_MESSAGES/messages.po

@@ -904,10 +904,11 @@ msgstr ""
 #: changedetectionio/blueprint/settings/templates/settings_llm_tab.html
 #, python-format
 msgid ""
-"Local 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, lower it if you want tighter limits. Only applied to self-"
-"hosted OpenAI-compatible endpoints — cloud providers (OpenAI, Anthropic, Gemini) keep their original tight caps."
+"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."
 msgstr ""
 
 #: changedetectionio/blueprint/settings/templates/settings_llm_tab.html
@@ -981,6 +982,12 @@ msgstr ""
 msgid "Removes all cached AI change summaries across all watches. They will be regenerated on the next check."
 msgstr ""
 
+#: changedetectionio/blueprint/settings/templates/settings_llm_tab.html
+msgid ""
+"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."
+msgstr ""
+
 #: changedetectionio/blueprint/settings/templates/settings_llm_tab.html
 msgid "Default AI Change Summary"
 msgstr ""
@@ -3167,6 +3174,10 @@ msgstr ""
 msgid "Use LLM as a fallback for extracting price and restock info"
 msgstr ""
 
+#: changedetectionio/forms.py
+msgid "Enable LLM debug logging"
+msgstr ""
+
 #: changedetectionio/forms.py
 msgid "AI thinking budget (tokens)"
 msgstr ""

changedetectionio/translations/es/LC_MESSAGES/messages.po

@@ -942,10 +942,11 @@ msgstr ""
 #: changedetectionio/blueprint/settings/templates/settings_llm_tab.html
 #, python-format
 msgid ""
-"Local 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, lower it if you want tighter limits. Only applied to self-"
-"hosted OpenAI-compatible endpoints — cloud providers (OpenAI, Anthropic, Gemini) keep their original tight caps."
+"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."
 msgstr ""
 
 #: changedetectionio/blueprint/settings/templates/settings_llm_tab.html
@@ -1019,6 +1020,12 @@ msgstr ""
 msgid "Removes all cached AI change summaries across all watches. They will be regenerated on the next check."
 msgstr ""
 
+#: changedetectionio/blueprint/settings/templates/settings_llm_tab.html
+msgid ""
+"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."
+msgstr ""
+
 #: changedetectionio/blueprint/settings/templates/settings_llm_tab.html
 msgid "Default AI Change Summary"
 msgstr ""
@@ -3240,6 +3247,10 @@ msgstr ""
 msgid "Use LLM as a fallback for extracting price and restock info"
 msgstr ""
 
+#: changedetectionio/forms.py
+msgid "Enable LLM debug logging"
+msgstr ""
+
 #: changedetectionio/forms.py
 msgid "AI thinking budget (tokens)"
 msgstr ""

changedetectionio/translations/fr/LC_MESSAGES/messages.po

@@ -910,10 +910,11 @@ msgstr ""
 #: changedetectionio/blueprint/settings/templates/settings_llm_tab.html
 #, python-format
 msgid ""
-"Local 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, lower it if you want tighter limits. Only applied to self-"
-"hosted OpenAI-compatible endpoints — cloud providers (OpenAI, Anthropic, Gemini) keep their original tight caps."
+"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."
 msgstr ""
 
 #: changedetectionio/blueprint/settings/templates/settings_llm_tab.html
@@ -987,6 +988,12 @@ msgstr ""
 msgid "Removes all cached AI change summaries across all watches. They will be regenerated on the next check."
 msgstr ""
 
+#: changedetectionio/blueprint/settings/templates/settings_llm_tab.html
+msgid ""
+"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."
+msgstr ""
+
 #: changedetectionio/blueprint/settings/templates/settings_llm_tab.html
 msgid "Default AI Change Summary"
 msgstr ""
@@ -3180,6 +3187,10 @@ msgstr ""
 msgid "Use LLM as a fallback for extracting price and restock info"
 msgstr ""
 
+#: changedetectionio/forms.py
+msgid "Enable LLM debug logging"
+msgstr ""
+
 #: changedetectionio/forms.py
 msgid "AI thinking budget (tokens)"
 msgstr ""

changedetectionio/translations/it/LC_MESSAGES/messages.po

@@ -906,10 +906,11 @@ msgstr ""
 #: changedetectionio/blueprint/settings/templates/settings_llm_tab.html
 #, python-format
 msgid ""
-"Local 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, lower it if you want tighter limits. Only applied to self-"
-"hosted OpenAI-compatible endpoints — cloud providers (OpenAI, Anthropic, Gemini) keep their original tight caps."
+"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."
 msgstr ""
 
 #: changedetectionio/blueprint/settings/templates/settings_llm_tab.html
@@ -983,6 +984,12 @@ msgstr ""
 msgid "Removes all cached AI change summaries across all watches. They will be regenerated on the next check."
 msgstr ""
 
+#: changedetectionio/blueprint/settings/templates/settings_llm_tab.html
+msgid ""
+"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."
+msgstr ""
+
 #: changedetectionio/blueprint/settings/templates/settings_llm_tab.html
 msgid "Default AI Change Summary"
 msgstr ""
@@ -3169,6 +3176,10 @@ msgstr ""
 msgid "Use LLM as a fallback for extracting price and restock info"
 msgstr ""
 
+#: changedetectionio/forms.py
+msgid "Enable LLM debug logging"
+msgstr ""
+
 #: changedetectionio/forms.py
 msgid "AI thinking budget (tokens)"
 msgstr ""

changedetectionio/translations/ja/LC_MESSAGES/messages.po

@@ -911,10 +911,11 @@ msgstr ""
 #: changedetectionio/blueprint/settings/templates/settings_llm_tab.html
 #, python-format
 msgid ""
-"Local 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, lower it if you want tighter limits. Only applied to self-"
-"hosted OpenAI-compatible endpoints — cloud providers (OpenAI, Anthropic, Gemini) keep their original tight caps."
+"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."
 msgstr ""
 
 #: changedetectionio/blueprint/settings/templates/settings_llm_tab.html
@@ -988,6 +989,12 @@ msgstr ""
 msgid "Removes all cached AI change summaries across all watches. They will be regenerated on the next check."
 msgstr ""
 
+#: changedetectionio/blueprint/settings/templates/settings_llm_tab.html
+msgid ""
+"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."
+msgstr ""
+
 #: changedetectionio/blueprint/settings/templates/settings_llm_tab.html
 msgid "Default AI Change Summary"
 msgstr ""
@@ -3186,6 +3193,10 @@ msgstr ""
 msgid "Use LLM as a fallback for extracting price and restock info"
 msgstr ""
 
+#: changedetectionio/forms.py
+msgid "Enable LLM debug logging"
+msgstr ""
+
 #: changedetectionio/forms.py
 msgid "AI thinking budget (tokens)"
 msgstr ""

changedetectionio/translations/ko/LC_MESSAGES/messages.po

@@ -912,10 +912,11 @@ msgstr "Ollama 또는 사용자 지정/자체 호스팅 엔드포인트에만
 #: changedetectionio/blueprint/settings/templates/settings_llm_tab.html
 #, python-format
 msgid ""
-"Local 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, lower it if you want tighter limits. Only applied to self-"
-"hosted OpenAI-compatible endpoints — cloud providers (OpenAI, Anthropic, Gemini) keep their original tight caps."
+"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."
 msgstr ""
 
 #: changedetectionio/blueprint/settings/templates/settings_llm_tab.html
@@ -989,6 +990,12 @@ msgstr "모든 요약 캐시 지우기"
 msgid "Removes all cached AI change summaries across all watches. They will be regenerated on the next check."
 msgstr "모든 모니터링에 저장된 AI 변경 요약 캐시를 제거합니다. 다음 확인 시 다시 생성됩니다."
 
+#: changedetectionio/blueprint/settings/templates/settings_llm_tab.html
+msgid ""
+"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."
+msgstr ""
+
 #: changedetectionio/blueprint/settings/templates/settings_llm_tab.html
 msgid "Default AI Change Summary"
 msgstr "기본 AI 변경 요약"
@@ -3177,6 +3184,10 @@ msgstr "{{diff}} 알림 토큰을 AI 요약으로 대체"
 msgid "Use LLM as a fallback for extracting price and restock info"
 msgstr "가격 및 재입고 정보 추출의 대체 수단으로 LLM 사용"
 
+#: changedetectionio/forms.py
+msgid "Enable LLM debug logging"
+msgstr ""
+
 #: changedetectionio/forms.py
 msgid "AI thinking budget (tokens)"
 msgstr "AI 추론 예산 (토큰)"

changedetectionio/translations/messages.pot

@@ -8,7 +8,7 @@ msgid ""
 msgstr ""
 "Project-Id-Version: changedetection.io 0.55.3\n"
 "Report-Msgid-Bugs-To: EMAIL@ADDRESS\n"
-"POT-Creation-Date: 2026-05-15 15:44+0200\n"
+"POT-Creation-Date: 2026-05-15 18:31+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"
@@ -903,10 +903,11 @@ msgstr ""
 #: changedetectionio/blueprint/settings/templates/settings_llm_tab.html
 #, python-format
 msgid ""
-"Local 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, lower it if you want tighter limits. Only applied to self-"
-"hosted OpenAI-compatible endpoints — cloud providers (OpenAI, Anthropic, Gemini) keep their original tight caps."
+"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."
 msgstr ""
 
 #: changedetectionio/blueprint/settings/templates/settings_llm_tab.html
@@ -980,6 +981,12 @@ msgstr ""
 msgid "Removes all cached AI change summaries across all watches. They will be regenerated on the next check."
 msgstr ""
 
+#: changedetectionio/blueprint/settings/templates/settings_llm_tab.html
+msgid ""
+"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."
+msgstr ""
+
 #: changedetectionio/blueprint/settings/templates/settings_llm_tab.html
 msgid "Default AI Change Summary"
 msgstr ""
@@ -3166,6 +3173,10 @@ msgstr ""
 msgid "Use LLM as a fallback for extracting price and restock info"
 msgstr ""
 
+#: changedetectionio/forms.py
+msgid "Enable LLM debug logging"
+msgstr ""
+
 #: changedetectionio/forms.py
 msgid "AI thinking budget (tokens)"
 msgstr ""

changedetectionio/translations/pt_BR/LC_MESSAGES/messages.po

@@ -929,10 +929,11 @@ msgstr ""
 #: changedetectionio/blueprint/settings/templates/settings_llm_tab.html
 #, python-format
 msgid ""
-"Local 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, lower it if you want tighter limits. Only applied to self-"
-"hosted OpenAI-compatible endpoints — cloud providers (OpenAI, Anthropic, Gemini) keep their original tight caps."
+"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."
 msgstr ""
 
 #: changedetectionio/blueprint/settings/templates/settings_llm_tab.html
@@ -1006,6 +1007,12 @@ msgstr ""
 msgid "Removes all cached AI change summaries across all watches. They will be regenerated on the next check."
 msgstr ""
 
+#: changedetectionio/blueprint/settings/templates/settings_llm_tab.html
+msgid ""
+"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."
+msgstr ""
+
 #: changedetectionio/blueprint/settings/templates/settings_llm_tab.html
 msgid "Default AI Change Summary"
 msgstr ""
@@ -3217,6 +3224,10 @@ msgstr ""
 msgid "Use LLM as a fallback for extracting price and restock info"
 msgstr ""
 
+#: changedetectionio/forms.py
+msgid "Enable LLM debug logging"
+msgstr ""
+
 #: changedetectionio/forms.py
 msgid "AI thinking budget (tokens)"
 msgstr ""

changedetectionio/translations/tr/LC_MESSAGES/messages.po

@@ -939,10 +939,11 @@ msgstr ""
 #: changedetectionio/blueprint/settings/templates/settings_llm_tab.html
 #, python-format
 msgid ""
-"Local 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, lower it if you want tighter limits. Only applied to self-"
-"hosted OpenAI-compatible endpoints — cloud providers (OpenAI, Anthropic, Gemini) keep their original tight caps."
+"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."
 msgstr ""
 
 #: changedetectionio/blueprint/settings/templates/settings_llm_tab.html
@@ -1016,6 +1017,12 @@ msgstr ""
 msgid "Removes all cached AI change summaries across all watches. They will be regenerated on the next check."
 msgstr ""
 
+#: changedetectionio/blueprint/settings/templates/settings_llm_tab.html
+msgid ""
+"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."
+msgstr ""
+
 #: changedetectionio/blueprint/settings/templates/settings_llm_tab.html
 msgid "Default AI Change Summary"
 msgstr ""
@@ -3220,6 +3227,10 @@ msgstr ""
 msgid "Use LLM as a fallback for extracting price and restock info"
 msgstr ""
 
+#: changedetectionio/forms.py
+msgid "Enable LLM debug logging"
+msgstr ""
+
 #: changedetectionio/forms.py
 msgid "AI thinking budget (tokens)"
 msgstr ""

changedetectionio/translations/uk/LC_MESSAGES/messages.po

@@ -919,10 +919,11 @@ msgstr ""
 #: changedetectionio/blueprint/settings/templates/settings_llm_tab.html
 #, python-format
 msgid ""
-"Local 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, lower it if you want tighter limits. Only applied to self-"
-"hosted OpenAI-compatible endpoints — cloud providers (OpenAI, Anthropic, Gemini) keep their original tight caps."
+"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."
 msgstr ""
 
 #: changedetectionio/blueprint/settings/templates/settings_llm_tab.html
@@ -996,6 +997,12 @@ msgstr ""
 msgid "Removes all cached AI change summaries across all watches. They will be regenerated on the next check."
 msgstr ""
 
+#: changedetectionio/blueprint/settings/templates/settings_llm_tab.html
+msgid ""
+"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."
+msgstr ""
+
 #: changedetectionio/blueprint/settings/templates/settings_llm_tab.html
 msgid "Default AI Change Summary"
 msgstr ""
@@ -3199,6 +3206,10 @@ msgstr ""
 msgid "Use LLM as a fallback for extracting price and restock info"
 msgstr ""
 
+#: changedetectionio/forms.py
+msgid "Enable LLM debug logging"
+msgstr ""
+
 #: changedetectionio/forms.py
 msgid "AI thinking budget (tokens)"
 msgstr ""

changedetectionio/translations/zh/LC_MESSAGES/messages.po

@@ -908,10 +908,11 @@ msgstr ""
 #: changedetectionio/blueprint/settings/templates/settings_llm_tab.html
 #, python-format
 msgid ""
-"Local 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, lower it if you want tighter limits. Only applied to self-"
-"hosted OpenAI-compatible endpoints — cloud providers (OpenAI, Anthropic, Gemini) keep their original tight caps."
+"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."
 msgstr ""
 
 #: changedetectionio/blueprint/settings/templates/settings_llm_tab.html
@@ -985,6 +986,12 @@ msgstr ""
 msgid "Removes all cached AI change summaries across all watches. They will be regenerated on the next check."
 msgstr ""
 
+#: changedetectionio/blueprint/settings/templates/settings_llm_tab.html
+msgid ""
+"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."
+msgstr ""
+
 #: changedetectionio/blueprint/settings/templates/settings_llm_tab.html
 msgid "Default AI Change Summary"
 msgstr ""
@@ -3172,6 +3179,10 @@ msgstr ""
 msgid "Use LLM as a fallback for extracting price and restock info"
 msgstr ""
 
+#: changedetectionio/forms.py
+msgid "Enable LLM debug logging"
+msgstr ""
+
 #: changedetectionio/forms.py
 msgid "AI thinking budget (tokens)"
 msgstr ""

changedetectionio/translations/zh_Hant_TW/LC_MESSAGES/messages.po

@@ -907,10 +907,11 @@ msgstr ""
 #: changedetectionio/blueprint/settings/templates/settings_llm_tab.html
 #, python-format
 msgid ""
-"Local 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, lower it if you want tighter limits. Only applied to self-"
-"hosted OpenAI-compatible endpoints — cloud providers (OpenAI, Anthropic, Gemini) keep their original tight caps."
+"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."
 msgstr ""
 
 #: changedetectionio/blueprint/settings/templates/settings_llm_tab.html
@@ -984,6 +985,12 @@ msgstr ""
 msgid "Removes all cached AI change summaries across all watches. They will be regenerated on the next check."
 msgstr ""
 
+#: changedetectionio/blueprint/settings/templates/settings_llm_tab.html
+msgid ""
+"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."
+msgstr ""
+
 #: changedetectionio/blueprint/settings/templates/settings_llm_tab.html
 msgid "Default AI Change Summary"
 msgstr ""
@@ -3171,6 +3178,10 @@ msgstr ""
 msgid "Use LLM as a fallback for extracting price and restock info"
 msgstr ""
 
+#: changedetectionio/forms.py
+msgid "Enable LLM debug logging"
+msgstr ""
+
 #: changedetectionio/forms.py
 msgid "AI thinking budget (tokens)"
 msgstr ""

changedetectionio/worker.py

@@ -539,9 +539,11 @@ async def async_update_worker(worker_id, q, notification_q, app, datastore, exec
                                     )
                                     _llm_to_version = list(watch.history.keys())[-1]
                                     _llm_max_summary_tokens = datastore.data['settings']['application'].get('llm_max_summary_tokens', 3000)
+                                    _llm_model = (datastore.data['settings']['application'].get('llm') or {}).get('model', '')
                                     _llm_cache_prompt = build_summary_cache_prompt(
                                         effective_prompt=get_effective_summary_prompt(watch, datastore),
                                         max_summary_tokens=_llm_max_summary_tokens,
+                                        model=_llm_model,
                                     )
                                     watch.save_llm_diff_summary(
                                         update_obj['_llm_change_summary'],