Fix multi-turn chat: lock composer during generation, always stream into a fresh end-of-history bubble

app.py

@@ -724,6 +724,11 @@ def _parse_think(raw: str) -> tuple[str, str, bool]:
     close_idx = rest.find(THINK_CLOSE)
     before = raw[:open_idx]
     if close_idx == -1:
+        # Hide a partially streamed closing tag at the end of the think block.
+        for i in range(len(THINK_CLOSE) - 1, 0, -1):
+            if rest.endswith(THINK_CLOSE[:i]):
+                rest = rest[:-i]
+                break
         return rest.strip(), before.strip(), True
     think = rest[:close_idx].strip()
     answer = (before + rest[close_idx + len(THINK_CLOSE) :]).strip()
@@ -853,20 +858,27 @@ def generate_reply(history, max_new_tokens, temperature, top_p, enable_thinking)
         return
     conversation = _history_for_model(history)
 
+    # Stream only into a bubble at the END of the history. Reaching for any
+    # older assistant message would overwrite a previous turn in place.
     think_msg = None
-    for msg in reversed(history):
-        if msg.get("role") == "assistant" and isinstance(msg.get("metadata"), dict):
-            think_msg = msg
-            break
-    if think_msg is None:
-        think_msg = {
-            "role": "assistant",
-            "content": "",
-            "metadata": {"title": "Logicizing…", "status": "pending"},
-        }
+    last = history[-1] if history else None
+    if (
+        last is not None
+        and last.get("role") == "assistant"
+        and isinstance(last.get("metadata"), dict)
+        and last["metadata"].get("status") == "pending"
+    ):
+        think_msg = last
+        think_msg["content"] = ""
+    elif last is not None and last.get("role") == "user":
+        # Gradio strips the empty pending bubble in preprocessing; recreate it.
+        think_msg = {"role": "assistant", "content": ""}
         history.append(think_msg)
     else:
-        think_msg["metadata"] = {"title": "Logicizing…", "status": "pending"}
+        # No new user turn (e.g. an empty submit); don't regenerate old answers.
+        yield history
+        return
+    think_msg["metadata"] = {"title": "Logicizing…", "status": "pending"}
     yield history
 
     answer_msg = {"role": "assistant", "content": ""}
@@ -958,39 +970,21 @@ with gr.Blocks(fill_height=True, fill_width=True, elem_id="app-root") as demo:
                                     gr.Button("Analyze it", size="sm", elem_classes=["twil-card-btn"])
                                 )
 
-    send.click(
-        queue_message,
-        [prompt, chatbot],
-        [prompt, chatbot, chips_row],
-        show_progress="hidden",
-    ).then(
-        generate_reply,
-        [chatbot, max_new_tokens, temperature, top_p, enable_thinking],
-        chatbot,
-        concurrency_limit=1,
-        show_progress="hidden",
-    )
-    prompt.submit(
-        queue_message,
-        [prompt, chatbot],
-        [prompt, chatbot, chips_row],
-        show_progress="hidden",
-    ).then(
-        generate_reply,
-        [chatbot, max_new_tokens, temperature, top_p, enable_thinking],
-        chatbot,
-        concurrency_limit=1,
-        show_progress="hidden",
-    )
-    for btn, (_, template) in zip(chip_btns, CHIPS):
-        btn.click(lambda t=template: t, outputs=prompt)
+    # Lock the composer while a turn is running: overlapping events make the
+    # finishing stream overwrite the chat with its own stale history.
+    lockable = [prompt, send, *card_btns]
 
-    for btn, case in zip(card_btns, SHOWCASE):
-        case_prompt = SHOWCASE_TEMPLATE.format(argument=case["quote"])
-        btn.click(
-            lambda p=case_prompt: (p, gr.Tabs(selected="chat")),
-            outputs=[prompt, tabs],
-        ).then(
+    def _lock():
+        return [gr.update(interactive=False)] * len(lockable)
+
+    def _unlock():
+        return [gr.update(interactive=True)] * len(lockable)
+
+    def _wire(event, prep_fn=None, prep_outputs=None):
+        chain = event(_lock, None, lockable, show_progress="hidden")
+        if prep_fn is not None:
+            chain = chain.then(prep_fn, outputs=prep_outputs, show_progress="hidden")
+        chain = chain.then(
             queue_message,
             [prompt, chatbot],
             [prompt, chatbot, chips_row],
@@ -1002,6 +996,20 @@ with gr.Blocks(fill_height=True, fill_width=True, elem_id="app-root") as demo:
             concurrency_limit=1,
             show_progress="hidden",
         )
+        chain.then(_unlock, None, lockable, show_progress="hidden")
+
+    _wire(send.click)
+    _wire(prompt.submit)
+    for btn, (_, template) in zip(chip_btns, CHIPS):
+        btn.click(lambda t=template: t, outputs=prompt)
+
+    for btn, case in zip(card_btns, SHOWCASE):
+        case_prompt = SHOWCASE_TEMPLATE.format(argument=case["quote"])
+        _wire(
+            btn.click,
+            prep_fn=lambda p=case_prompt: (p, gr.Tabs(selected="chat")),
+            prep_outputs=[prompt, tabs],
+        )
 
 if __name__ == "__main__":
     demo.launch(