feat(llm): user-controlled think mode (default off); remove qwen3 hardcode

The chat generation pipeline previously forced think=True unconditionally
to match qwen3's combined think+tools template, locking the system into
that model family. Bench data (2026-05-21, qwen3:30b-a3b/qwen3:32b on
CPU) showed thinking adds 1-2 minutes per turn for unclear quality
benefit — qwen3:30b-a3b even produced more rambling with think on.

This decouples think from the model family by reading a per-user
`think_enabled` setting (default `false`). Non-qwen3 models can now run
through the same pipeline without the silent-generation failure mode
that content-gated thinking would have caused — they just don't think.
qwen3 users who still want thinking can opt in via the Settings UI.

Settings UI:
- New "Enable model thinking" checkbox in General → Assistant section.
- Help text explains the default-off rationale and when to opt in.
- Persists via the existing settings API; no schema migration needed
  (Setting is key/value text).

Telemetry to confirm whether this regresses tool-call reliability on
qwen3 (the current model) is in a follow-up commit (generation_tool_log).

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

frontend/src/views/SettingsView.vue

@@ -21,6 +21,8 @@ const savingTimezone = ref(false);
 const timezoneSaved = ref(false);
 const autoConsolidateTasks = ref(true);
 const savingAutoConsolidate = ref(false);
+const thinkEnabled = ref(false);
+const savingThinkEnabled = ref(false);
 
 async function saveAutoConsolidate() {
   savingAutoConsolidate.value = true;
@@ -35,6 +37,19 @@ async function saveAutoConsolidate() {
   }
 }
 
+async function saveThinkEnabled() {
+  savingThinkEnabled.value = true;
+  try {
+    await apiPut('/api/settings', {
+      think_enabled: thinkEnabled.value ? "true" : "false",
+    });
+  } catch {
+    toastStore.show('Failed to save setting', 'error');
+  } finally {
+    savingThinkEnabled.value = false;
+  }
+}
+
 function detectTimezone() {
   userTimezone.value = Intl.DateTimeFormat().resolvedOptions().timeZone;
 }
@@ -771,6 +786,9 @@ onMounted(async () => {
   userTimezone.value = allSettings.user_timezone ?? "";
   // Default true if unset; explicit "false" disables auto-consolidation.
   autoConsolidateTasks.value = (allSettings.auto_consolidate_tasks ?? "true") !== "false";
+  // Default false: think mode is opt-in. Bench showed it costs 1-2 min/turn
+  // for unclear benefit; qwen3 users can re-enable if tool calls regress.
+  thinkEnabled.value = (allSettings.think_enabled ?? "false") === "true";
   chatRetentionDays.value = allSettings.chat_retention_days !== undefined
     ? Number(allSettings.chat_retention_days)
     : 90;
@@ -1518,6 +1536,22 @@ function formatUserDate(iso: string): string {
               </span>
             </p>
           </div>
+          <div class="field">
+            <label class="checkbox-label">
+              <input
+                type="checkbox"
+                v-model="thinkEnabled"
+                :disabled="savingThinkEnabled"
+                @change="saveThinkEnabled"
+              />
+              Enable model thinking
+            </label>
+            <p class="field-hint">
+              When on, the chat model runs in reasoning mode before responding.
+              Default off — bench data showed thinking adds 1–2 minutes per turn for unclear quality benefit.
+              Turn on for qwen3-family models if tool-call reliability drops without it.
+            </p>
+          </div>
         </div>
         <div class="actions">
           <button class="btn-save" @click="saveAssistant" :disabled="saving">

src/fabledassistant/services/generation_task.py

@@ -231,12 +231,15 @@ async def run_generation(
     # Emit context event
     buf.append_event("context", {"context": context_meta})
 
-    # Always think on qwen3-class models: reasoning mode is the only reliable
+    # Think mode is per-user-setting (default off). Historically forced on for
-    # path for the tool-call template. Content-based gating was tried in 87fcaa6
+    # qwen3 because content-gated thinking (87fcaa6) exposed silent-generation
-    # but exposed silent-generation failures on short tool-intent prompts, since
+    # failures on short tool-intent prompts. After moving to a model-family
-    # the classifier had no way to tell that "create a task" needs a tool call.
+    # decoupled architecture, the bench data (May 2026) showed think costs
+    # 1-2 min/turn for unclear quality benefit; default off, opt in via
+    # the Settings UI. The generation_tool_log captures per-turn outcomes
+    # so reliability regressions surface empirically.
     think_requested = think
-    think = True
+    think = (await get_setting(user_id, "think_enabled", "false")).lower() == "true"
 
     t_start = time.monotonic()
     timing: dict = {