fix(channels): filter malformed thinking output

src/octop/infra/gateway/gateway.py

@@ -34,6 +34,7 @@
     WebSocketChannel,
     WebSocketHub,
 )
+from octop.infra.utils.llm_text import strip_thinking
 from octop.infra.utils.locale import DEFAULT_LOCALE, Locale
 
 if TYPE_CHECKING:
@@ -452,7 +453,7 @@ async def push_text_from_session(
             async for chunk in self._agent_manager.stream(agent_id, request):
                 if chunk.get("type") in ("token", "delta"):
                     parts.append(str(chunk.get("content") or chunk.get("text") or ""))
-            outbound = "".join(parts).strip() or "(empty)"
+            outbound = strip_thinking("".join(parts)) or "(empty)"
 
         if virtual_stream:
             self._bump_dashboard_session(session, session_key, text)

src/octop/infra/gateway/process/response_mode.py

@@ -14,6 +14,8 @@
     TextContent,
 )
 
+from octop.infra.utils.llm_text import strip_thinking
+
 ChannelResponseMode = Literal["invoke", "stream"]
 
 DEFAULT_CHANNEL_RESPONSE_MODE: ChannelResponseMode = "invoke"
@@ -81,7 +83,7 @@ async def collapse_to_invoke_response(
 
         if event.type == MessageEventType.COMPLETED:
             content: list[ContentPart] = []
-            final_text = text_buffer.strip()
+            final_text = strip_thinking(text_buffer)
             if final_text:
                 content.append(TextContent(text=final_text))
             content.extend(media_buffer)

src/octop/infra/utils/llm_text.py

@@ -11,14 +11,36 @@
 from typing import Any
 
 _THINKING_RE = re.compile(
-    r"<think>[\s\S]*?</think>\s*",
+    r"<(?:think|thinking)>[\s\S]*?</(?:think|thinking)>\s*",
     re.IGNORECASE,
 )
+_THINKING_OPEN_RE = re.compile(r"<(?:think|thinking)>", re.IGNORECASE)
+_THINKING_CLOSE_RE = re.compile(r"</(?:think|thinking)>", re.IGNORECASE)
 
 
 def strip_thinking(text: str) -> str:
-    """Remove ``<think>...</think>`` blocks from model output."""
-    return _THINKING_RE.sub("", text).strip()
+    """Remove tagged or malformed thinking prefixes from model output.
+
+    Some OpenAI-compatible model routers return reasoning in ordinary
+    ``content`` and only emit the closing ``</think>`` marker.  Treat a
+    closing marker that appears before any opening marker as the boundary
+    between the hidden prefix and the user-visible answer.
+    """
+    cleaned = _THINKING_RE.sub("", text)
+
+    while True:
+        close_match = _THINKING_CLOSE_RE.search(cleaned)
+        if close_match is None:
+            break
+        open_match = _THINKING_OPEN_RE.search(cleaned)
+        if open_match is not None and open_match.start() < close_match.start():
+            break
+        cleaned = cleaned[close_match.end() :]
+
+    open_match = _THINKING_OPEN_RE.search(cleaned)
+    if open_match is not None:
+        cleaned = cleaned[: open_match.start()]
+    return cleaned.strip()
 
 
 def llm_text_content(result: Any) -> str:

tests/unit/api/test_chat_polish.py

@@ -107,6 +107,16 @@ def test_strip_thinking_removes_redacted_block() -> None:
     assert _strip_thinking(raw) == "Polished prompt"
 
 
+def test_strip_thinking_removes_orphan_closing_prefix() -> None:
+    raw = "internal reasoning without an opening tag</think>\nVisible answer"
+    assert _strip_thinking(raw) == "Visible answer"
+
+
+def test_strip_thinking_removes_unclosed_thinking_suffix() -> None:
+    raw = "Visible answer\n<thinking>truncated internal reasoning"
+    assert _strip_thinking(raw) == "Visible answer"
+
+
 def test_llm_text_content_strips_thinking_from_string_message() -> None:
     class Msg:
         content = "<think>plan</think>Final text"

tests/unit/gateway/test_gateway_push.py

@@ -229,6 +229,48 @@ async def stream_with_token(_agent_id, _request):
     assert gateway._channel_manager.push_text.await_args.args[2] == "AI reply"
 
 
+@pytest.mark.asyncio
+async def test_push_text_from_session_cron_agent_strips_orphan_thinking_prefix(
+    gateway: Gateway,
+) -> None:
+    sk = ThreadRegistry.make_key(
+        agent_id="a1",
+        channel_type="weixin",
+        channel_subject_id="wx_1",
+    )
+    gateway.thread_registry._threads.insert(
+        thread_id="thr_weixin_cron",
+        agent_id="a1",
+        user_id=1,
+        channel_type="weixin",
+        session_key=sk,
+    )
+    gateway.thread_registry._sessions.upsert(
+        session_key=sk,
+        agent_id="a1",
+        user_id=1,
+        channel_type="weixin",
+        chat_type="dm",
+        thread_id="thr_weixin_cron",
+        channel_subject_id="wx_1",
+        channel_chat_type="dm",
+        channel_metadata={"channel_type": "weixin"},
+        channel_id="ch-1",
+    )
+
+    async def stream_with_thinking_leak(_agent_id, _request):
+        yield {"type": "token", "content": "internal reasoning"}
+        yield {"type": "token", "content": "</think>"}
+        yield {"type": "token", "content": "最终学习内容"}
+
+    gateway._agent_manager.stream = stream_with_thinking_leak
+
+    await gateway.push_text_from_session("a1", sk, "run agent", task_type="agent")
+
+    gateway._channel_manager.push_text.assert_awaited_once()
+    assert gateway._channel_manager.push_text.await_args.args[2] == "最终学习内容"
+
+
 @pytest.mark.asyncio
 async def test_push_text_from_session_dashboard_agent_pushes_ws(gateway: Gateway) -> None:
     sk = ThreadRegistry.dashboard_key(agent_id="a1", user_id=1)

tests/unit/gateway/test_response_mode.py

@@ -49,6 +49,27 @@ async def test_invoke_discards_progress_before_tool_and_emits_final_once() -> No
     assert text.text == "这是最终答案。"
 
 
+@pytest.mark.asyncio
+async def test_invoke_strips_orphan_thinking_prefix_from_final_text() -> None:
+    source = _events(
+        MessageEvent.delta("Let me inspect another source. "),
+        MessageEvent.delta("This is internal reasoning."),
+        MessageEvent.delta("</think>"),
+        MessageEvent.delta("【每日指南学习】最终内容"),
+        MessageEvent.completed(),
+    )
+
+    result = [event async for event in collapse_to_invoke_response(source)]
+
+    assert [event.type for event in result] == [
+        MessageEventType.MESSAGE,
+        MessageEventType.COMPLETED,
+    ]
+    text = result[0].content[0]
+    assert isinstance(text, TextContent)
+    assert text.text == "【每日指南学习】最终内容"
+
+
 @pytest.mark.asyncio
 async def test_invoke_preserves_tool_media_with_final_text() -> None:
     attachment = FileContent(filename="report.pdf", data="cGRm")