fix(sdk): strip leading 思考 block from LLM-generated title

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
@@ -11,6 +12,43 @@
 logger = get_logger(__name__)
 
 
+# Conservative peel of a leading 思考 reasoning block. Some providers
+# (e.g. Qwen3 behind Nebius) return reasoning inline in `content` as
+# `思考reasoning…思考` instead of the normalized `reasoning_content`
+# field, which would otherwise leak into the autogenerated title.
+# The peel only targets a leading block so mid-text literal `思考`
+# occurrences in a legitimate title are preserved.
+_THINK_BLOCK_RE = re.compile(r"^\s*思考[\s\S]*?思考\s*", re.DOTALL)
+
+
+def _strip_leading_think_block(text: str) -> str:
+    """Remove a leading ``思考`` block from ``text`` if present.
+
+    The conservative peel handles two shapes:
+
+    * A closed block ``思考…思考`` at the start of the response is
+      stripped and the remainder is returned.
+    * An unterminated block (the response begins with ``思考`` and
+      never closes) consumes the entire response, returning ``""`` so
+      the LLM title path yields no title and the existing
+      user-message truncation fallback runs.
+
+    A leading ``思考`` substring that is NOT a complete unclosed block
+    (e.g. a mid-text literal occurrence in a legitimate title) is
+    preserved unchanged.
+    """
+    if not text:
+        return ""
+    if not text.lstrip().startswith("思考"):
+        return text
+    match = _THINK_BLOCK_RE.match(text)
+    if match:
+        return text[match.end() :]
+    # Leading 思考 with no closer — the whole response is the think
+    # block, so consume it and let the caller fall back.
+    return ""
+
+
 categories = [
     {"emoji": "💄", "name": "frontend", "description": "UI and style files"},
     {"emoji": "👔", "name": "backend", "description": "Business logic"},
@@ -137,7 +175,18 @@ 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_leading_think_block(response.message.content[0].text).strip()
+
+            # A response that contained only a 思考 block yields no
+            # usable title from the LLM path; fall back to the existing
+            # user-message truncation rather than handing back the
+            # peel whitespace.
+            if not title:
+                logger.warning(
+                    "LLM title response contained only a 思考 block; "
+                    "falling back to truncation"
+                )
+                return None
 
             # Ensure the title isn't too long
             if len(title) > max_length:

tests/sdk/conversation/test_generate_title.py

@@ -287,3 +287,99 @@ def test_generate_title_disables_streaming_when_llm_streams(mock_transport):
     assert mock_transport.call_args.kwargs["enable_streaming"] is False
     assert mock_transport.call_args.kwargs["on_token"] is None
     assert streaming_llm.stream is True
+
+
+@patch("openhands.sdk.llm.llm.LLM.completion")
+def test_generate_title_strips_leading_think_block(mock_completion):
+    """A leading 思考…思考 block in the LLM title response is peeled
+    before the title is returned. Reproduces the Qwen3-behind-Nebius
+    inline-reasoning leak reported in #4530.
+    """
+    agent = create_test_agent()
+    conv = Conversation(agent=agent, visualizer=None)
+
+    user_message = create_user_message_event("Help me create a Python script")
+    conv.state.events.append(user_message)
+
+    custom_llm = LLM(model="gpt-4o-mini", api_key=SecretStr("key"), usage_id="think")
+    mock_completion.return_value = create_mock_llm_response(
+        "思考reasoning…思考🐛 Fix the login bug"
+    )
+
+    title = conv.generate_title(llm=custom_llm)
+
+    assert title == "🐛 Fix the login bug"
+
+
+@patch("openhands.sdk.llm.llm.LLM.completion")
+def test_generate_title_unclosed_think_block_falls_back(mock_completion):
+    """An unterminated 思考 block is consumed by the same peel; if
+    nothing real remains the LLM path returns ``None`` so the
+    user-message truncation fallback supplies the title.
+    """
+    agent = create_test_agent()
+    conv = Conversation(agent=agent, visualizer=None)
+
+    user_message = create_user_message_event("Help me create a Python script")
+    conv.state.events.append(user_message)
+
+    custom_llm = LLM(
+        model="gpt-4o-mini", api_key=SecretStr("key"), usage_id="think-unclosed"
+    )
+    # No closing 思考 — the peel matches nothing, the whole string is
+    # preserved, but it is the *only* content we got, so the LLM path
+    # must not hand it back as a title.
+    mock_completion.return_value = create_mock_llm_response(
+        "思考half-formed reasoning…"
+    )
+
+    title = conv.generate_title(llm=custom_llm)
+
+    # Falls back to the user message, not the raw reasoning block.
+    assert title == "Help me create a Python script"
+
+
+@patch("openhands.sdk.llm.llm.LLM.completion")
+def test_generate_title_think_only_response_falls_back(mock_completion):
+    """A response that is only a 思考 block (closed or unclosed) yields
+    no LLM title, so the truncation fallback runs.
+    """
+    agent = create_test_agent()
+    conv = Conversation(agent=agent, visualizer=None)
+
+    user_message = create_user_message_event("Help me create a Python script")
+    conv.state.events.append(user_message)
+
+    custom_llm = LLM(
+        model="gpt-4o-mini", api_key=SecretStr("key"), usage_id="think-only"
+    )
+    mock_completion.return_value = create_mock_llm_response(
+        "思考deep reasoning chain思考"
+    )
+
+    title = conv.generate_title(llm=custom_llm)
+
+    assert title == "Help me create a Python script"
+
+
+@patch("openhands.sdk.llm.llm.LLM.completion")
+def test_generate_title_preserves_mid_text_think_literal(mock_completion):
+    """A mid-text literal 思考 occurrence in a legitimate title is
+    preserved. The peel only targets a *leading* block, so a title
+    that legitimately contains the characters ``思考`` (e.g. a
+    non-English conversation title) is untouched.
+    """
+    agent = create_test_agent()
+    conv = Conversation(agent=agent, visualizer=None)
+
+    user_message = create_user_message_event("Help me create a Python script")
+    conv.state.events.append(user_message)
+
+    custom_llm = LLM(
+        model="gpt-4o-mini", api_key=SecretStr("key"), usage_id="think-mid"
+    )
+    mock_completion.return_value = create_mock_llm_response("🐛 Use 思考 in a sentence")
+
+    title = conv.generate_title(llm=custom_llm)
+
+    assert title == "🐛 Use 思考 in a sentence"