From 28ec3af6974c0f04ed9eb3cb4667a307a4493645 Mon Sep 17 00:00:00 2001
From: Frank Song <franksong2702@gmail.com>
Date: Thu, 14 May 2026 08:42:57 +0800
Subject: [PATCH] fix: strip only leading user-asking wrapper line

Refs #2215 Fix B: remove the mid-response stripping hazard without losing leading multi-line wrapper cleanup.

The pattern now strips only a leading 'the user is asking' wrapper line and preserves the visible answer that follows. Add regression coverage for both the leading-wrapper and mid-response prose cases.
---
 api/streaming.py       |  2 +-
 tests/test_issue607.py | 14 ++++++++++++++
 2 files changed, 15 insertions(+), 1 deletion(-)

diff --git a/api/streaming.py b/api/streaming.py
index bf473d90..0d1334e0 100644
--- a/api/streaming.py
+++ b/api/streaming.py
@@ -695,7 +695,7 @@ def _strip_thinking_markup(text: str) -> str:
     s = re.sub(r'^\s*<think>.*?</think>\s*', ' ', s, flags=re.IGNORECASE | re.DOTALL)
     s = re.sub(r'^\s*<\|channel\|?>thought\n?.*?<channel\|>\s*', ' ', s, flags=re.IGNORECASE | re.DOTALL)
     s = re.sub(r'^\s*<\|turn\|>thinking\n.*?<turn\|>\s*', ' ', s, flags=re.IGNORECASE | re.DOTALL)  # Gemma 4
-    s = re.sub(r'^\s*(the|ther)\s+user\s+is\s+asking.*$', ' ', s, flags=re.IGNORECASE | re.MULTILINE)
+    s = re.sub(r'^\s*(the|ther)\s+user\s+is\s+asking[^\n]*(?:\n|$)', ' ', s, flags=re.IGNORECASE)
     # Strip plain-text thinking preambles from models that don't use <think> tags (e.g. Qwen3).
     # These appear as the very first sentence of the assistant response and are not useful as titles.
     s = re.sub(
diff --git a/tests/test_issue607.py b/tests/test_issue607.py
index 0e950c3d..112fe916 100644
--- a/tests/test_issue607.py
+++ b/tests/test_issue607.py
@@ -88,6 +88,20 @@ class TestGemma4ThinkingTokenStrip:
         result = _strip_thinking_markup(raw)
         assert result == raw
 
+    def test_leading_the_user_is_asking_wrapper_line_is_stripped(self):
+        """A leading title-wrapper line is metadata; the visible answer should remain."""
+        raw = "The user is asking about Python imports.\nUse importlib for dynamic imports."
+        result = _strip_thinking_markup(raw)
+        assert result == "Use importlib for dynamic imports."
+
+    def test_mid_response_the_user_is_asking_line_is_preserved(self):
+        """Only the leading wrapper line is stripped; normal mid-response prose survives."""
+        raw = "First, define the context.\nThe user is asking us to be patient.\nThen answer."
+        result = _strip_thinking_markup(raw)
+        assert "First, define the context." in result
+        assert "The user is asking us to be patient." in result
+        assert "Then answer." in result
+
     def test_minimax_channel_without_second_pipe_still_stripped_at_start(self):
         """The client-facing MiniMax <|channel>thought shape is stripped when leading."""
         raw = "<|channel>thought\nSome reasoning<channel|>Answer"