feat(thinking): toggle to enable/disable model thinking mode

Add a Thinking on/off button next to Auto-Play TTS. When OFF (default),
the client sends enable_thinking=false with each chat_message and the
server passes chat_template_kwargs.enable_thinking=false to self-hosted
OpenAI-compatible endpoints (Qwen3-style), suppressing chain-of-thought
to save tokens rather than merely hiding it. Skips the param for
api.openai.com (which would 400) and o1/o3 (always think); templates that
ignore the kwarg simply drop it. Client-side reasoning suppression remains
as a fallback. State persists to localStorage across desktop and mobile.

app.py

@@ -1899,7 +1899,14 @@ def handle_message(data):
             gevent.spawn(generate_dalle_image, room_name, message, username)
         else:
             # All other models (Groq, Together, Mistral, etc.) use OpenAI client
-            gevent.spawn(chat_gpt, username, room_name, model_name=model)
+            enable_thinking = data.get("enable_thinking", True)
+            gevent.spawn(
+                chat_gpt,
+                username,
+                room_name,
+                model_name=model,
+                enable_thinking=enable_thinking,
+            )
 
 
 @socketio.on("delete_message")
@@ -2120,7 +2127,7 @@ def chat_claude(
     socketio.emit("delete_processing_message", msg_id, room=room_name)
 
 
-def chat_gpt(username, room_name, model_name="gpt-4o-mini"):
+def chat_gpt(username, room_name, model_name="gpt-4o-mini", enable_thinking=True):
     openai_client, model_name = get_openai_client_and_model(model_name)
 
     temperature = 0
@@ -2194,23 +2201,29 @@ def chat_gpt(username, room_name, model_name="gpt-4o-mini"):
 
     first_chunk = True
 
+    create_kwargs = {
+        "model": model_name,
+        "messages": chat_history,
+        "n": 1,
+        "stream": True,
+    }
+    if "o3" not in model_name:
+        # o3 does not support temperature at all!
+        create_kwargs["temperature"] = temperature
+    if not enable_thinking:
+        # Qwen3 / vLLM-style switch to suppress chain-of-thought and save
+        # tokens. OpenAI's hosted API rejects unknown body params, so only
+        # send to self-hosted OpenAI-compatible endpoints; o1/o3 think
+        # unconditionally and would 400 on this anyway. Models whose chat
+        # template ignores the kwarg (e.g. Hermes) simply drop it.
+        base_url = str(getattr(openai_client, "base_url", "") or "")
+        if "api.openai.com" not in base_url:
+            create_kwargs["extra_body"] = {
+                "chat_template_kwargs": {"enable_thinking": False}
+            }
+
     try:
-        if "o3" in model_name:
-            # o3 does not support temperature at all!
-            chunks = openai_client.chat.completions.create(
-                model=model_name,
-                messages=chat_history,
-                n=1,
-                stream=True,
-            )
-        else:
-            chunks = openai_client.chat.completions.create(
-                model=model_name,
-                messages=chat_history,
-                n=1,
-                temperature=temperature,
-                stream=True,
-            )
+        chunks = openai_client.chat.completions.create(**create_kwargs)
     except Exception as e:
         with app.app_context():
             message_content = f"{model_name} Error: {e}"

templates/base.html

@@ -103,6 +103,11 @@
                     Auto-Play TTS: OFF
                 </button>
             </div>
+            <div id="show-thinking-mobile">
+                <button id="show-thinking-btn-mobile" onclick="toggleShowThinking()" style="width: 100%; margin-top: 10px; background-color: #f44336; color: white; border: none; padding: 8px; border-radius: 4px; cursor: pointer;">
+                    Thinking: OFF
+                </button>
+            </div>
             <div id="activity-controls-mobile">
                 <h3>Activities</h3>
                 <div id="current-activity-info-mobile" style="display: none;">

templates/chat.html

@@ -59,6 +59,11 @@
                 Auto-Play TTS: OFF
             </button>
         </div>
+        <div>
+            <button id="show-thinking-btn" onclick="toggleShowThinking()" style="width: 100%; margin-top: 10px; background-color: #f44336; color: white; border: none; padding: 8px; border-radius: 4px; cursor: pointer;">
+                Thinking: OFF
+            </button>
+        </div>
 
        <div id="activity-controls">
            <h3>Activities</h3>
@@ -148,6 +153,13 @@ let autoPlayTTS = localStorage.getItem('autoPlayTTS') === 'true' || false;
 let ttsQueue = [];
 let isPlayingTTS = false;
 
+// Thinking-mode state. Default OFF. When OFF we send enable_thinking=false
+// with each chat_message so the server asks the model (Qwen3-style) to skip
+// chain-of-thought entirely — saving tokens, not just hiding output. As a
+// fallback, any reasoning_content that still arrives is dropped client-side
+// (see message_chunk handler). Toggle persists to localStorage.
+let showThinking = localStorage.getItem('showThinking') === 'true';
+
 // Vision model state for auto alt-text
 let visionAvailable = false;
 let visionModel = null;
@@ -375,6 +387,9 @@ document.addEventListener('DOMContentLoaded', (event) => {
     // Initialize auto-play TTS button state from localStorage
     updateAutoPlayTTSDisplay();
 
+    // Initialize show-thinking button state from localStorage
+    updateShowThinkingDisplay();
+
     // Check for vision model availability (enables image hover descriptions)
     initVisionCapability();
 
@@ -660,7 +675,8 @@ function sendMessage() {
             "username": username,
             "message": messageToSend,
             "model": model,  // Pass model as a separate attribute
-            "room_name": room_name
+            "room_name": room_name,
+            "enable_thinking": showThinking  // false => ask model to skip chain-of-thought
         });
         messageTextarea.value = "";
         // Reset textarea height after sending
@@ -1112,6 +1128,29 @@ function processNextTTS() {
         });
 }
 
+// Update show-thinking button labels/colors on both desktop + mobile.
+function updateShowThinkingDisplay() {
+    const btn = document.getElementById("show-thinking-btn");
+    const btnMobile = document.getElementById("show-thinking-btn-mobile");
+    const label = showThinking ? "Thinking: ON" : "Thinking: OFF";
+    const color = showThinking ? "#4CAF50" : "#f44336";
+    if (btn) {
+        btn.textContent = label;
+        btn.style.backgroundColor = color;
+    }
+    if (btnMobile) {
+        btnMobile.textContent = label;
+        btnMobile.style.backgroundColor = color;
+    }
+}
+
+// Toggle reasoning_content display. Off by default; persisted in localStorage.
+function toggleShowThinking() {
+    showThinking = !showThinking;
+    localStorage.setItem('showThinking', showThinking.toString());
+    updateShowThinkingDisplay();
+}
+
 // Function to update auto-play TTS button display
 function updateAutoPlayTTSDisplay() {
     const autoPlayBtn = document.getElementById("auto-play-tts-btn");
@@ -1448,7 +1487,10 @@ socket.on("message_chunk", (data) => {
     // the message-content div. Only appears if the server actually streams
     // delta.reasoning_content (i.e., thinking is on AND the model is using it).
     // Auto-collapses below on the first content delta.
+    // Fallback guard: with Thinking OFF the server already suppresses reasoning
+    // at the model, but if a model ignores the switch and streams anyway, drop it.
     if (data.reasoning_content) {
+        if (!showThinking) return;
         let thinkingDetails = messageWrapper.querySelector(".message-thinking");
         if (!thinkingDetails) {
             // Ensure a message-body wrapper exists to anchor against