fix(sdk): strip inline reasoning from LLM-generated titles

openhands-sdk/openhands/sdk/conversation/title_utils.py

@@ -1,5 +1,6 @@
 """Utility functions for generating conversation titles."""
 
+import re
 from collections.abc import Callable, Sequence
 
 from openhands.sdk.event import MessageEvent
@@ -59,6 +60,21 @@ def extract_first_user_message(events: Sequence[Event]) -> str | None:
     return None
 
 
+_REASONING_BLOCK = re.compile(r"<think>.*?</think>", re.DOTALL | re.IGNORECASE)
+_UNCLOSED_REASONING = re.compile(r"<think>.*", re.DOTALL | re.IGNORECASE)
+
+
+def strip_reasoning_blocks(text: str) -> str:
+    """Remove inline ``<think>`` reasoning from a model's text.
+
+    Providers that do not split chain-of-thought into ``reasoning_content`` return it
+    inline in ``content``. An unterminated block means the response was cut mid-thought,
+    so everything from the opening tag on is reasoning too.
+    """
+    text = _REASONING_BLOCK.sub("", text)
+    return _UNCLOSED_REASONING.sub("", text)
+
+
 def generate_title_with_llm(
     message: str,
     llm: LLM,
@@ -132,7 +148,13 @@ def generate_title_with_llm(
         if response.message.content and isinstance(
             response.message.content[0], TextContent
         ):
-            title = response.message.content[0].text.strip()
+            title = strip_reasoning_blocks(response.message.content[0].text).strip()
+
+            if not title:
+                logger.warning(
+                    "LLM returned only reasoning content for title generation"
+                )
+                return None
 
             # Ensure the title isn't too long
             if len(title) > max_length:

tests/sdk/conversation/test_generate_title.py

@@ -428,3 +428,45 @@ def test_title_uses_real_http_transport(title_http_server, tmp_path, mode):
     assert bool(body.get("stream")) == (mode == "subscription")
     if mode != "chat":
         assert body["store"] is False
+
+
+@patch("openhands.sdk.llm.llm.LLM.completion")
+def test_generate_title_strips_inline_reasoning(mock_completion):
+    """Guards #4530.
+
+    Providers that do not split chain-of-thought into `reasoning_content` return it
+    inline as `<think>...</think>`. The title is consumed verbatim, so without
+    stripping, truncation to `max_length` keeps the reasoning and discards the title.
+    """
+    llm = LLM(model="qwen3-32b", api_key=SecretStr("test-key"), usage_id="t")
+    mock_completion.return_value = create_mock_llm_response(
+        "<think>The user wants a CSV summary script. I will pick the features "
+        "emoji and keep it short.</think>✨ Summarise a CSV in Python"
+    )
+
+    title = generate_title_with_llm("Help me summarise a CSV", llm)
+
+    assert title == "✨ Summarise a CSV in Python"
+
+
+@patch("openhands.sdk.llm.llm.LLM.completion")
+def test_generate_title_strips_unterminated_reasoning(mock_completion):
+    """An unterminated block means the response was cut mid-thought, so there is no
+    title to salvage and the caller falls back to a truncated message title."""
+    llm = LLM(model="qwen3-32b", api_key=SecretStr("test-key"), usage_id="t")
+    mock_completion.return_value = create_mock_llm_response(
+        "<think>Let me consider what this conversation is really about"
+    )
+
+    assert generate_title_with_llm("Help me summarise a CSV", llm) is None
+
+
+@patch("openhands.sdk.llm.llm.LLM.completion")
+def test_generate_title_keeps_text_without_reasoning(mock_completion):
+    """A normal response is unaffected."""
+    llm = LLM(model="gpt-4o-mini", api_key=SecretStr("test-key"), usage_id="t")
+    mock_completion.return_value = create_mock_llm_response("✨ Create Python Script")
+
+    assert generate_title_with_llm("Help me write a script", llm) == (
+        "✨ Create Python Script"
+    )