* Behavior Change:

- Only CoT can be edited now - to reduce confusion
- Cognitive parameters only applies to CoT

This is meant to be an experimental interface focusing on co-thinking with reasoning model, not a full-spec interface. We recommend other open source LLM frontend (e.g. open-webui) for complex cases for now.

app.py

@@ -20,9 +20,7 @@ class AppConfig:
     DEFAULT_THROUGHPUT = 10
     SYNC_THRESHOLD_DEFAULT = 0
     API_TIMEOUT = 20
-    LOADING_DEFAULT = (
-        "✅ Ready! <br> Think together with AI. Use Shift+Enter to toggle generation"
-    )
+
 
 
 class DynamicState:
@@ -185,7 +183,7 @@ class ConvoState:
             )
             for chunk in response_stream:
                 chunk_content = chunk.choices[0].delta.content
-                if coordinator.should_pause_for_human(full_response):
+                if coordinator.should_pause_for_human(full_response) and dynamic_state.in_cot:
                     dynamic_state.should_stream = False
                 if not dynamic_state.should_stream:
                     break
@@ -211,7 +209,7 @@ class ConvoState:
                         else lang_data["loading_output"]
                     )
                     editor_label = f"{lang_data['editor_label']} - {status}"
-                    yield full_response, gr.update(
+                    yield self.current["cot"] + ("</think>" if think_complete else ""), gr.update(
                         label=editor_label
                     ), self.flatten_output()
 
@@ -219,7 +217,7 @@ class ConvoState:
                     start_time = time.time()
                     while (
                         time.time() - start_time
-                    ) < interval and dynamic_state.should_stream:
+                    ) < interval and dynamic_state.should_stream and dynamic_state.in_cot:
                         time.sleep(0.005)
 
         except Exception as e:
@@ -247,9 +245,9 @@ class ConvoState:
             )
             editor_label = f"{lang_data['editor_label']} - {final_status}"
             if not self.is_error:
-                yield full_response, gr.update(label=editor_label), self.flatten_output()
+                yield self.current["cot"] + ("</think>" if not dynamic_state.in_cot else ""), gr.update(label=editor_label), self.flatten_output()
             else:
-                yield full_response, gr.update(label=editor_label_error), self.flatten_output() + [
+                yield self.current["cot"] + ("</think>" if not dynamic_state.in_cot else ""), gr.update(label=editor_label_error), self.flatten_output() + [
                 {
                     "role": "assistant",
                     "content": error_msg,

lang.py

@@ -14,7 +14,7 @@ LANGUAGE_CONFIG = {
         "throughput_info": "Tokens/s - 5:Learn, 10:Follow, 50:Skim",
         "language_label": "Language",
         "loading_thinking": "🤖 AI Thinking ↓ Shift+Enter to Pause",
-        "loading_output": "🖨️ Result Writing ↓ Shift+Enter to Pause",
+        "loading_output": "🖨️ Result Writing → (Pause & submit to reroll the result)",
         "interrupted": "🤔 Pause, Human thinking time - **EDIT THOUGHTS BELOW**",
         "completed": "✅ Completed → Check overview",
         "error": "Error",
@@ -51,8 +51,8 @@ LANGUAGE_CONFIG = {
         "throughput_label": "⏱ 同步思考速度",
         "throughput_info": "词元/秒 - 5:学习, 10:跟读, 50:跳读",
         "language_label": "界面语言",
-        "loading_thinking": "🤖 AI思考中... **Shift+Enter**可暂停",
-        "loading_output": "🖨️ 结果输出中... **Shift+Enter**可暂停",
+        "loading_thinking": "🤖 AI思考中 ↓ Shift+Enter可暂停",
+        "loading_output": "🖨️ 结果输出中 → (暂停再提交会重新生成结果部分)",
         "interrupted": "🤔 暂停，人类思考回合 **下面的思考过程可以编辑**",
         "completed": "✅ 已完成 → 查看完整对话",
         "error": "错误",