fix: strip <think> tags from reporter output to prevent thinking text leakage (#781) (#862)

* fix: strip <think> tags from LLM output to prevent thinking text leakage (#781)

Some models (e.g. DeepSeek-R1, QwQ via ollama) embed reasoning in
content using <think>...</think> tags instead of the separate
reasoning_content field. This causes thinking text to leak into
both streamed messages and the final report.

Fix at two layers:
- server/app.py: strip <think> tags in _create_event_stream_message
  so ALL streamed content is filtered (coordinator, planner, etc.)
- graph/nodes.py: strip <think> tags in reporter_node before storing
  final_report (which is not streamed through the event layer)

The regex uses a fast-path check ("<think>" in content) to avoid
unnecessary regex calls on normal content.

* refactor: add defensive check for think tag stripping and add reporter_node tests (#781)

- Add isinstance and fast-path check in reporter_node before regex, consistent with app.py
- Add TestReporterNodeThinkTagStripping with 5 test cases covering various scenarios

* chore: re-trigger review

src/graph/nodes.py

@@ -4,6 +4,7 @@
 import json
 import logging
 import os
+import re
 from functools import partial
 from typing import Annotated, Any, Literal
 
@@ -900,6 +901,12 @@ def reporter_node(state: State, config: RunnableConfig):
     logger.debug(f"Current invoke messages: {invoke_messages}")
     response = get_llm_by_type(AGENT_LLM_MAP["reporter"]).invoke(invoke_messages)
     response_content = response.content
+    # Strip <think>...</think> tags that some models (e.g. QwQ, DeepSeek) embed
+    # directly in content instead of using the reasoning_content field (#781)
+    if isinstance(response_content, str) and "<think>" in response_content:
+        response_content = re.sub(
+            r"<think>[\s\S]*?</think>", "", response_content
+        ).strip()
     logger.info(f"reporter response: {response_content}")
 
     return {

src/server/app.py

@@ -6,6 +6,7 @@ import base64
 import json
 import logging
 import os
+import re
 from typing import Annotated, Any, List, Optional, cast
 from uuid import uuid4
 
@@ -423,6 +424,11 @@ def _create_event_stream_message(
     if not isinstance(content, str):
         content = json.dumps(content, ensure_ascii=False)
 
+    # Strip <think>...</think> tags that some models (e.g. DeepSeek-R1, QwQ via ollama)
+    # embed directly in content instead of using the reasoning_content field (#781)
+    if isinstance(content, str) and "<think>" in content:
+        content = re.sub(r"<think>[\s\S]*?</think>", "", content).strip()
+
     event_stream_message = {
         "thread_id": thread_id,
         "agent": agent_name,