[Bugfix][Responses] Fix browser.find action type (#55305)

Signed-off-by: nicole-lihui <nicole.li@daocloud.io>
Co-authored-by: Chauncey <chaunceyjiang@gmail.com>

tests/entrypoints/openai/responses/test_harmony_utils.py

@@ -273,6 +273,23 @@ class TestHarmonyToResponseOutput:
         assert output_items[0].action.type == "search"
         assert output_items[0].action.query == "cursor:weather in San Francisco"
 
+    def test_browser_find_recipient_uses_responses_action_type(self):
+        """browser.find must use the Responses API find_in_page action type."""
+        message = (
+            Message.from_role_and_content(
+                Role.ASSISTANT, '{"pattern": "vLLM", "cursor": 42}'
+            )
+            .with_channel("analysis")
+            .with_recipient("browser.find")
+        )
+
+        output_items = harmony_to_response_output(message, frozenset())
+
+        assert len(output_items) == 1
+        assert isinstance(output_items[0], ResponseFunctionWebSearch)
+        assert output_items[0].action.type == "find_in_page"
+        assert output_items[0].action.pattern == "vLLM"
+
     def test_commentary_with_empty_content_and_no_recipient(self):
         """Test edge case: empty commentary with recipient=None."""
         message = Message.from_role_and_content(Role.ASSISTANT, "")

tests/entrypoints/openai/responses/test_streaming_events.py

@@ -1,6 +1,9 @@
 # SPDX-License-Identifier: Apache-2.0
 # SPDX-FileCopyrightText: Copyright contributors to the vLLM project
 
+from openai.types.responses import ResponseFunctionWebSearch
+from openai_harmony import Message, Role
+
 from vllm.entrypoints.generate.base.protocol import (
     DeltaFunctionCall,
     DeltaMessage,
@@ -8,11 +11,35 @@ from vllm.entrypoints.generate.base.protocol import (
 )
 from vllm.entrypoints.openai.responses.streaming_events import (
     SimpleStreamingEventProcessor,
+    StreamingState,
     _StateType,
+    emit_browser_tool_events,
     split_delta,
 )
 
 
+def test_browser_find_uses_responses_action_type():
+    """Both streaming output items must use the Responses find action type."""
+    message = (
+        Message.from_role_and_content(
+            Role.ASSISTANT, '{"pattern": "vLLM", "cursor": 42}'
+        )
+        .with_channel("analysis")
+        .with_recipient("browser.find")
+    )
+
+    events = emit_browser_tool_events(message, StreamingState())
+
+    added, done = events[0], events[-1]
+    assert added.type == "response.output_item.added"
+    assert done.type == "response.output_item.done"
+    for event in (added, done):
+        assert isinstance(event.item, ResponseFunctionWebSearch)
+        assert event.item.action.type == "find_in_page"
+        assert event.item.action.pattern == "vLLM"
+        assert event.item.action.url == "cursor:42"
+
+
 def _make_tool_call(
     index: int, name: str | None = None, arguments: str | None = None
 ) -> DeltaToolCall:

vllm/entrypoints/openai/responses/harmony.py

@@ -303,7 +303,7 @@ def _parse_browser_tool_call(message: Message, recipient: str) -> ResponseOutput
         action = ActionFind(
             pattern=browser_call.get("pattern", ""),
             url=f"cursor:{browser_call.get('url', '')}",
-            type="find",
+            type="find_in_page",
         )
     else:
         raise ValueError(f"Unknown browser action: {recipient}")

vllm/entrypoints/openai/responses/streaming_events.py

@@ -667,7 +667,7 @@ def emit_browser_tool_events(
         )
     elif function_name == "find":
         action = response_function_web_search.ActionFind(
-            type="find",
+            type="find_in_page",
             pattern=parsed_args["pattern"],
             # TODO: translate to url
             url=f"cursor:{parsed_args.get('cursor', '')}",