[Bugfix][Frontend] Fix Inkling tool name leaking into content after reasoning (#58792)

Signed-off-by: Baljinder Hothi <baljinder.hothi@cohere.com>
Co-authored-by: Flora Feng <4florafeng@gmail.com>

tests/parser/engine/test_inkling.py

@@ -23,8 +23,13 @@ from vllm.entrypoints.openai.chat_completion.protocol import (
     ChatCompletionToolsParam,
     FunctionDefinition,
 )
+from vllm.parser.abstract_parser import DelegatingParser
 from vllm.parser.engine.events import EventType
 from vllm.parser.engine.parser_engine_config import ParserState
+from vllm.parser.engine.registered_adapters import (
+    InklingParserReasoningAdapter,
+    InklingParserToolAdapter,
+)
 from vllm.parser.engine.streaming_parser_engine import StreamingParserEngine
 from vllm.parser.inkling import InklingParser, _inkling_arg_converter, inkling_config
 from vllm.parser.parser_manager import ParserManager
@@ -166,6 +171,11 @@ def _delegating(mock_tokenizer, tools=None):
     return parser_cls(mock_tokenizer, tools or [])
 
 
+class _TwoPassInklingParser(DelegatingParser):
+    reasoning_parser_cls = InklingParserReasoningAdapter
+    tool_parser_cls = InklingParserToolAdapter
+
+
 def _stream_delegating(parser, request, text, chunk_size, prompt_token_ids):
     """Stream ``text`` through ``DelegatingParser.parse_delta``, ``chunk_size``
     tokens per delta; return ``(content, reasoning, ordered tool names,
@@ -759,6 +769,30 @@ class TestDelegatingTwoPass:
         assert TOOL_JSON not in content
         assert END_MESSAGE not in content
 
+    def test_function_name_not_in_content_after_reasoning(
+        self, mock_tokenizer, mock_request
+    ):
+        name = "search_documents_by_keyword"
+        tools = [_function_tool(name)]
+        mock_request.tools = tools
+        text = (
+            f"{THINK_START} context.{END_MESSAGE}{MSG_MODEL}"
+            f"{name}{_tool_block(name, '{}')}"
+        )
+
+        content, reasoning, names, args = _stream_delegating(
+            _TwoPassInklingParser(mock_tokenizer, tools),
+            mock_request,
+            text,
+            chunk_size=12,
+            prompt_token_ids=self.GEN_PROMPT,
+        )
+
+        assert reasoning == " context."
+        assert content == ""
+        assert names == [name]
+        assert json.loads(args[0]) == {}
+
     @pytest.mark.parametrize("opener", [TEXT_START, TOOL_TEXT, TOOL_ERROR])
     @pytest.mark.parametrize("chunk_size", [1, 3, 64])
     def test_visible_text_then_tool_streaming(

vllm/parser/inkling.py

@@ -167,11 +167,12 @@ def inkling_config() -> ParserEngineConfig:
             (EventType.REASONING_END, EventType.TOOL_CALL_START),
         ),
     }
-    # Block-end terminals behave identically regardless of label. A
-    # closed tool block returns to CONTENT (Inkling has no section wrapper;
+    # A closed tool block returns to CONTENT (Inkling has no section wrapper;
     # blocks of any kind may follow), which also keeps the block-kind
     # and role tokens out of the engine's tool-terminal set so the
     # skip_tool_parsing reasoning pass still classifies reasoning.
+    # In a reasoning block, THINK_END defers REASONING_END to the next block
+    # opener so the function-name header is not emitted as content.
     for end in ("THINK_END", "END_SAMPLING"):
         transitions[(ParserState.CONTENT, end)] = Transition(
             ParserState.CONTENT,
@@ -179,7 +180,7 @@ def inkling_config() -> ParserEngineConfig:
         )
         transitions[(ParserState.REASONING, end)] = Transition(
             ParserState.CONTENT,
-            (EventType.REASONING_END,),
+            () if end == "THINK_END" else (EventType.REASONING_END,),
         )
         transitions[(ParserState.TOOL_ARGS, end)] = Transition(
             ParserState.CONTENT,