[Bugfix] Fix Qwen3 prompt tool-call reasoning false positive (#45763)

Signed-off-by: Alex Bilichenko <alexbi29@users.noreply.github.com>
Co-authored-by: Alex Bilichenko <alexbi29@users.noreply.github.com>

tests/parser/engine/test_qwen3_reasoning.py

@@ -15,7 +15,12 @@ import pytest
 
 from tests.parser.engine.conftest import make_mock_tokenizer
 from tests.parser.engine.streaming_helpers import simulate_reasoning_streaming
+from vllm.parser.abstract_parser import DelegatingParser
 from vllm.parser.engine.parser_engine_config import ParserState
+from vllm.parser.engine.registered_adapters import (
+    Qwen3ParserReasoningAdapter,
+    Qwen3ParserToolAdapter,
+)
 from vllm.parser.qwen3 import Qwen3Parser, qwen3_config
 
 _THINK_START_ID = 50
@@ -32,6 +37,11 @@ _QWEN3_VOCAB = {
 }
 
 
+class _Qwen3DelegatingParser(DelegatingParser):
+    reasoning_parser_cls = Qwen3ParserReasoningAdapter
+    tool_parser_cls = Qwen3ParserToolAdapter
+
+
 @pytest.fixture
 def mock_tokenizer():
     return make_mock_tokenizer(_QWEN3_VOCAB)
@@ -162,6 +172,10 @@ class TestIsReasoningEnd:
         """Unpaired <tool_call> is implicit reasoning end."""
         assert parser.is_reasoning_end([_THINK_START_ID, 1, _TOOL_CALL_ID])
 
+    def test_prompt_tool_example_before_generation_think_not_end(self, parser):
+        """Tool examples before the generation <think> must not end reasoning."""
+        assert not parser.is_reasoning_end([_TOOL_CALL_ID, _TEXT_ID, _THINK_START_ID])
+
     def test_paired_tool_call_not_end(self, parser):
         """Paired <tool_call>...</tool_call> (from template) is NOT end."""
         assert not parser.is_reasoning_end(
@@ -178,6 +192,26 @@ class TestIsReasoningEnd:
         assert not parser.is_reasoning_end([])
 
 
+class TestDelegatingPromptDetection:
+    def test_prompt_tool_example_does_not_skip_streaming_reasoning(
+        self, mock_tokenizer, mock_request
+    ):
+        parser = _Qwen3DelegatingParser(mock_tokenizer)
+        prompt_ids = [_TOOL_CALL_ID, _TEXT_ID, _THINK_START_ID]
+
+        delta = parser.parse_delta(
+            "thinking",
+            [_TEXT_ID],
+            mock_request,
+            prompt_token_ids=prompt_ids,
+            finished=False,
+        )
+
+        assert delta is not None
+        assert delta.reasoning == "thinking"
+        assert delta.content is None
+
+
 class TestStreaming:
     def test_basic_streaming(self, parser):
         reasoning, content = simulate_reasoning_streaming(

vllm/parser/qwen3.py

@@ -206,8 +206,14 @@ class Qwen3Parser(ParserEngine):
             return True
         tool_call_id = self._tool_call_token_id
         tool_call_end_id = self._tool_call_end_token_id
+        reasoning_start_id = self._reasoning_start_token_id
         if tool_call_id is not None:
             for i in range(len(input_ids) - 1, -1, -1):
+                if (
+                    reasoning_start_id is not None
+                    and input_ids[i] == reasoning_start_id
+                ):
+                    return False
                 if input_ids[i] == tool_call_id:
                     if tool_call_end_id is not None and any(
                         input_ids[j] == tool_call_end_id