[Frontend] Support CustomToolCall in the non-Harmony Responses API (#57322)

Signed-off-by: Shijin Zhang <75300765+Dovis01@users.noreply.github.com>
Co-authored-by: Chauncey <chaunceyjiang@gmail.com>
This commit is contained in:
Shijin Zhang
2026-10-12 17:22:55 +08:00
committed by GitHub
co-authored by Chauncey
parent a33d46b471
commit 83773ecdf3
22 changed files with 698 additions and 72 deletions
@@ -1,13 +1,17 @@
# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
import pytest
from openai_harmony import (
Message,
)
from pydantic import ValidationError
from vllm.entrypoints.openai.responses.protocol import (
ResponsesRequest,
serialize_message,
serialize_messages,
)
from vllm.exceptions import VLLMValidationError
def test_serialize_message() -> None:
@@ -37,3 +41,34 @@ def test_serialize_messages() -> None:
}
msg = Message.from_dict(msg_value)
assert serialize_messages([msg, dict_value]) == [msg_value, dict_value]
def test_custom_tool_grammar_format_rejected() -> None:
grammar = {"type": "grammar", "syntax": "lark", "definition": 'start: "pwd"'}
with pytest.raises(VLLMValidationError) as exc_info:
ResponsesRequest(
input="hi", tools=[{"type": "custom", "name": "emit", "format": grammar}]
)
assert exc_info.value.parameter == "tools"
request = ResponsesRequest(
input="hi",
tools=[{"type": "custom", "name": "emit", "format": {"type": "text"}}],
)
assert request.tools is not None and request.tools[0].type == "custom"
@pytest.mark.parametrize(
"tools",
[
"abc",
{"type": "custom"},
[None],
[{"type": "custom", "name": "x", "format": "text"}],
],
)
def test_malformed_tools_rejected_by_field_validation(tools) -> None:
"""Malformed tools bodies fail pydantic field validation (4xx), not an
AttributeError inside the custom-format validator (500)."""
with pytest.raises(ValidationError):
ResponsesRequest(input="hi", tools=tools)
@@ -4,6 +4,7 @@
from unittest.mock import patch
import pytest
from openai.types.responses import CustomTool, FunctionTool, ResponseCustomToolCall
from openai.types.responses.response_function_tool_call import ResponseFunctionToolCall
from openai.types.responses.response_function_tool_call_output_item import (
ResponseFunctionToolCallOutputItem,
@@ -16,10 +17,14 @@ from openai.types.responses.response_reasoning_item import (
Summary,
)
from vllm.entrypoints.generate.base.protocol import FunctionCall
from vllm.entrypoints.openai.responses.utils import (
_construct_message_from_response_item,
build_response_output_items,
construct_chat_messages_with_tool_call,
construct_input_messages,
decode_custom_tool_input,
decode_custom_tool_input_prefix,
should_continue_final_message,
)
from vllm.exceptions import VLLMValidationError
@@ -248,6 +253,89 @@ class TestResponsesUtils:
assert formatted_item["content"] == "dongyi"
class TestCustomTools:
def test_custom_tool_call_output_item(self):
tools = [
CustomTool(type="custom", name="emit_command", format={"type": "text"}),
FunctionTool(
type="function", name="get_weather", parameters={}, strict=None
),
]
custom, function = build_response_output_items(
reasoning=None,
content=None,
tool_calls=[
FunctionCall(
id="call_1", name="emit_command", arguments='{"input": "pwd"}'
),
FunctionCall(
id="call_2", name="get_weather", arguments='{"city": "Paris"}'
),
],
tools=tools,
)
assert isinstance(custom, ResponseCustomToolCall)
assert (custom.call_id, custom.input) == ("call_1", "pwd")
assert custom.id.startswith("ctc_")
assert isinstance(function, ResponseFunctionToolCall)
assert function.arguments == '{"city": "Paris"}'
def test_custom_tool_history_replays_through_function_shim(self):
messages = construct_chat_messages_with_tool_call(
[
make_reasoning_item(content_text="Run pwd."),
ResponseCustomToolCall(
type="custom_tool_call",
id="ctc_1",
call_id="call_exec",
name="emit_command",
input='echo "hi"',
),
{
"type": "custom_tool_call_output",
"call_id": "call_exec",
"output": "hi",
},
{"role": "user", "content": "Continue"},
]
)
assert messages == [
{
"role": "assistant",
"reasoning": "Run pwd.",
"tool_calls": [
{
"id": "call_exec",
"type": "function",
"function": {
"name": "emit_command",
"arguments": '{"input": "echo \\"hi\\""}',
},
}
],
},
{"role": "tool", "content": "hi", "tool_call_id": "call_exec"},
{"role": "user", "content": "Continue"},
]
def test_decode_custom_tool_input(self):
assert decode_custom_tool_input('{"input": "pwd"}') == "pwd"
assert decode_custom_tool_input('{"cmd": "pwd"}') == '{"cmd": "pwd"}'
assert decode_custom_tool_input('{"file": "a.py"}') == '{"file": "a.py"}'
assert decode_custom_tool_input('{"input": 7}') == '{"input": 7}'
assert decode_custom_tool_input("pwd") == "pwd"
def test_decode_custom_tool_input_prefix(self):
assert decode_custom_tool_input_prefix('{"inp') == ""
assert decode_custom_tool_input_prefix('{"input": "ls \\"dir') == 'ls "dir'
assert decode_custom_tool_input_prefix('{"input": "a\\') == "a"
assert decode_custom_tool_input_prefix('{"input": "\\ud83d') == ""
assert (
decode_custom_tool_input_prefix('{"input": "\\ud83d\\ude00!"}')
== "\U0001f600!"
)
class TestReasoningItemContentPriority:
"""Tests that content is prioritized over summary for reasoning items."""
@@ -2,7 +2,12 @@
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
import pytest
from openai.types.responses import ResponseFunctionWebSearch, ResponseStreamEvent
from openai.types.responses import (
CustomTool,
ResponseCustomToolCall,
ResponseFunctionWebSearch,
ResponseStreamEvent,
)
from openai_harmony import Message, Role
from pydantic import TypeAdapter
@@ -20,6 +25,7 @@ from vllm.entrypoints.openai.responses.streaming_events import (
emit_reasoning_done_events,
split_delta,
)
from vllm.entrypoints.openai.responses.utils import decode_custom_tool_input
def test_browser_find_uses_responses_action_type():
@@ -190,3 +196,111 @@ class TestProcessorCompoundDeltas:
types = [e.type for e in events]
assert "response.reasoning_text.delta" in types
assert "response.output_text.delta" in types
class TestSimpleToolCallShapes:
def test_arguments_done_emitted_without_deltas(self):
processor = SimpleStreamingEventProcessor()
events = _run_through_processor(
processor, DeltaMessage(tool_calls=[_make_tool_call(0, name="noop")])
)
events += processor.close_current()
assert [e.type for e in events] == [
"response.output_item.added",
"response.function_call_arguments.done",
"response.output_item.done",
]
class TestCustomToolStreaming:
def test_custom_tool_input_events(self):
processor = SimpleStreamingEventProcessor(
tools=[CustomTool(type="custom", name="emit_command")]
)
events = _run_through_processor(
processor,
DeltaMessage(
tool_calls=[
_make_tool_call(0, name="emit_command", arguments='{"input": "ls ')
]
),
)
events += _run_through_processor(
processor,
DeltaMessage(tool_calls=[_make_tool_call(0, arguments='-la \\"x\\""}')]),
)
events += processor.close_current()
assert [e.type for e in events] == [
"response.output_item.added",
"response.custom_tool_call_input.delta",
"response.custom_tool_call_input.delta",
"response.custom_tool_call_input.done",
"response.output_item.done",
]
done = events[-1].item
assert isinstance(done, ResponseCustomToolCall)
assert done.input == 'ls -la "x"'
assert "".join(e.delta for e in events[1:3]) == done.input
assert events[3].input == done.input
assert events[0].item.type == "custom_tool_call"
assert events[0].item.id == done.id
assert done.id.startswith("ctc_")
assert events[0].item.call_id == done.call_id
def test_unrelated_keys_are_not_streamed_as_payload(self):
"""Review repro: chunks of an object whose input key arrives last must
not leak fragments of other keys into the payload."""
processor = SimpleStreamingEventProcessor(
tools=[CustomTool(type="custom", name="emit_command")]
)
events = _run_through_processor(
processor,
DeltaMessage(
tool_calls=[
_make_tool_call(0, name="emit_command", arguments='{"a": "xy"')
]
),
)
events += _run_through_processor(
processor,
DeltaMessage(
tool_calls=[_make_tool_call(0, arguments=', "input": "hello"}')]
),
)
events += processor.close_current()
deltas = [
e for e in events if e.type == "response.custom_tool_call_input.delta"
]
done_input = next(
e for e in events if e.type == "response.custom_tool_call_input.done"
).input
item = events[-1].item
assert "".join(e.delta for e in deltas) == "hello" # "xy" never leaks
assert done_input == "hello"
assert item.input == "hello"
assert decode_custom_tool_input('{"a": "xy", "input": "hello"}') == "hello"
def test_done_reported_authoritative_payload_on_mismatch(self):
"""If the final decode is not an extension of what was streamed, the
done event and item carry the authoritative value and a correcting
delta is emitted rather than dropping the divergence."""
processor = SimpleStreamingEventProcessor(
tools=[CustomTool(type="custom", name="emit_command")]
)
processor.state.current_state = _StateType.TOOL_CALL
processor.state.tool_call_is_custom = True
processor.state.tool_call_payload = "xy" # stale fragment already streamed
processor.state.accumulated_text = '{"a": "xy", "input": "hello"}'
events = processor.close_current()
delta = next(
e for e in events if e.type == "response.custom_tool_call_input.delta"
)
done_input = next(
e for e in events if e.type == "response.custom_tool_call_input.done"
).input
assert delta.delta == "hello" # full authoritative resend, not a dropped suffix
assert done_input == "hello"
assert events[-1].item.input == "hello"
@@ -8,6 +8,7 @@ from unittest.mock import MagicMock
import pytest
from vllm.entrypoints.openai.chat_completion.protocol import ChatCompletionRequest
from vllm.entrypoints.openai.responses.protocol import ResponsesRequest
from vllm.parser.gemma4 import (
TOOL_CALL_END,
TOOL_CALL_START,
@@ -117,6 +118,21 @@ def mock_request():
return request
def test_named_custom_tool_choice_skips_structured_outputs(mock_tokenizer):
# Responses custom tools ride the function shim, so a named custom choice
# must route like a named function choice, not into guided JSON.
tools = [{"type": "custom", "name": "emit_command"}]
request = ResponsesRequest(
input="hi",
tools=tools,
tool_choice={"type": "custom", "name": "emit_command"},
)
Gemma4EngineToolParser(mock_tokenizer, tools=tools).adjust_request(request)
assert request.skip_special_tokens is False
assert request.structured_outputs is None
# ---------------------------------------------------------------------------
# Unit tests for _parse_gemma4_args (shared parser logic)
# ---------------------------------------------------------------------------
@@ -8,6 +8,7 @@ import pytest
from vllm.entrypoints.generate.base.protocol import DeltaMessage
from vllm.entrypoints.openai.chat_completion.protocol import ChatCompletionRequest
from vllm.entrypoints.openai.responses.protocol import ResponsesRequest
from vllm.parser.parser_manager import ParserManager
from vllm.tool_parsers import ToolParserManager
from vllm.tool_parsers.k2_horizon_tool_parser import K2HorizonToolParser
@@ -316,3 +317,14 @@ def test_whitespace_only_surrounding_content_is_not_preserved(tokenizer):
def test_tool_parser_registered():
assert ToolParserManager.get_tool_parser("k2_horizon") is K2HorizonToolParser
assert K2HorizonToolParser.supports_required_and_named is False
def test_named_tool_resolves_custom_choice():
# Custom tools ride the function shim; a ToolChoiceCustom must resolve to
# its name like ToolChoiceFunction does.
request = ResponsesRequest(
input="hi",
tools=[{"type": "custom", "name": "emit_command"}],
tool_choice={"type": "custom", "name": "emit_command"},
)
assert K2HorizonToolParser._named_tool(request) == "emit_command"
@@ -8,6 +8,7 @@ import pytest
from vllm.entrypoints.openai.chat_completion.protocol import (
ChatCompletionRequest,
)
from vllm.entrypoints.openai.responses.protocol import ResponsesRequest
from vllm.exceptions import VLLMValidationError
from vllm.sampling_params import StructuredOutputsParams
@@ -60,3 +61,26 @@ def test_named_choice_allowed_with_structural_tag():
def test_named_choice_rejected_without_structural_tag():
with pytest.raises(VLLMValidationError):
_parser().adjust_request(_request(with_tag=False))
def _custom_request(with_tag: bool) -> ResponsesRequest:
req = ResponsesRequest(
input="hi",
tools=[{"type": "custom", "name": "emit_command"}],
tool_choice={"type": "custom", "name": "emit_command"},
)
if with_tag:
req.structured_outputs = StructuredOutputsParams(
structural_tag='{"type": "structural_tag", "format": {}}'
)
return req
def test_named_custom_choice_allowed_with_structural_tag():
req = _parser().adjust_request(_custom_request(with_tag=True))
assert req.skip_special_tokens is False
def test_named_custom_choice_rejected_without_structural_tag():
with pytest.raises(VLLMValidationError):
_parser().adjust_request(_custom_request(with_tag=False))
@@ -78,12 +78,15 @@ def _build_chat_request(*, tool_choice: str | dict[str, Any]) -> ChatCompletionR
def _build_responses_request(
*, tool_choice: str | dict[str, Any], include: list[str] | None = None
*,
tool_choice: str | dict[str, Any],
include: list[str] | None = None,
tools: list[dict[str, Any]] | None = None,
) -> ResponsesRequest:
return ResponsesRequest(
model="poolside-test",
input=[{"role": "user", "content": "write the file"}],
tools=[_responses_write_file_tool()],
tools=tools if tools is not None else [_responses_write_file_tool()],
tool_choice=tool_choice,
stream=True,
max_output_tokens=200,
@@ -145,6 +148,19 @@ def test_named_skips_structured_outputs_responses() -> None:
assert request.skip_special_tokens is False
def test_named_custom_tool_choice_skips_structured_outputs_responses() -> None:
# A named custom choice parses to ToolChoiceCustom; custom tools ride the
# function shim, so it must route like ToolChoiceFunction.
request = _build_responses_request(
tools=[{"type": "custom", "name": "emit_command"}],
tool_choice={"type": "custom", "name": "emit_command"},
)
PoolsideV1ToolParser(_StubTokenizer()).adjust_request(request)
assert request.text is None
assert request.skip_special_tokens is False
def test_auto_still_keeps_special_tokens() -> None:
request = _build_chat_request(tool_choice="auto")
_make_parser(request).adjust_request(request)
@@ -7,7 +7,7 @@ from typing import Any
from unittest.mock import MagicMock
import pytest
from openai.types.responses import ToolChoiceFunction
from openai.types.responses import CustomTool, ToolChoiceFunction
from openai.types.responses.tool_choice_allowed import ToolChoiceAllowed
from xgrammar import Grammar, StructuralTag
from xgrammar import get_model_structural_tag as get_xgrammar_builtin_structural_tag
@@ -787,6 +787,19 @@ def test_get_model_structural_tag_supports_named_tool_choice(
assert isinstance(tag, StructuralTag)
def test_glm_4_7_forced_custom_tool_uses_function_shim():
tools = [CustomTool(type="custom", name="emit_command", description="Emit.")]
tag = get_model_structural_tag(
model="glm_4_7",
tools=tools,
tool_choice="required",
reasoning=False,
)
assert isinstance(tag, StructuralTag)
assert "emit_command" in tag.model_dump_json()
@pytest.mark.parametrize(
("parser_cls", "model"),
[
+13
View File
@@ -29,6 +29,7 @@ from typing import cast
import pytest
from vllm.entrypoints.openai.chat_completion.protocol import ChatCompletionRequest
from vllm.entrypoints.openai.responses.protocol import ResponsesRequest
from vllm.reasoning.muse_glimmer_reasoning_parser import MuseGlimmerReasoningParser
from vllm.tool_parsers.muse_glimmer_tool_parser import MuseGlimmerToolParser
@@ -400,3 +401,15 @@ def test_exact_match_kept():
T, _call("get_weather"), _req("get_weather")
)
assert out.tool_calls[0].function.name == "get_weather"
def test_named_custom_tool_choice_skips_structured_outputs():
# Custom tools ride the function shim; a named custom choice must skip
# the base adjust_request like a named function choice does.
request = ResponsesRequest(
input="hi",
tools=[{"type": "custom", "name": "emit_command"}],
tool_choice={"type": "custom", "name": "emit_command"},
)
T.adjust_request(request)
assert request.structured_outputs is None
@@ -6,7 +6,9 @@ from copy import deepcopy
import pytest
import regex as re
from openai.types.responses import (
CustomTool,
FunctionTool,
ToolChoiceCustom,
ToolChoiceFunction,
WebSearchTool,
)
@@ -21,8 +23,10 @@ from vllm.entrypoints.openai.responses.protocol import ResponsesRequest
from vllm.exceptions import VLLMValidationError
from vllm.tool_parsers.streaming import extract_required_tool_call_streaming
from vllm.tool_parsers.utils import (
custom_tool_parameters,
find_tool_properties,
get_json_schema_from_tools,
iter_response_function_tool_dicts,
)
pytestmark = pytest.mark.cpu_test
@@ -483,6 +487,37 @@ class TestNonFunctionToolsSkipped:
get_json_schema_from_tools(tools=[WEB_SEARCH_TOOL], tool_choice="required")
class TestCustomToolShim:
"""``custom`` tools are shown to the model as a single-string function."""
def test_rendered_as_single_string_function(self):
tool = CustomTool(
type="custom", name="emit_command", description="Emit a command."
)
[tool_dict] = iter_response_function_tool_dicts([tool])
assert tool_dict["type"] == "function"
assert tool_dict["name"] == "emit_command"
assert tool_dict["description"] == "Emit a command."
assert tool_dict["parameters"] == custom_tool_parameters()
assert tool_dict["parameters"]["required"] == ["input"]
def test_forced_tool_choice_sees_custom_tool(self):
tools = [CustomTool(type="custom", name="emit_command"), FUNCTION_TOOL]
schema = get_json_schema_from_tools(tools=tools, tool_choice="required")
assert isinstance(schema, dict)
names = [e["properties"]["name"]["enum"] for e in schema["items"]["anyOf"]]
assert names == [["emit_command"], ["get_weather"]]
named = get_json_schema_from_tools(
tools=tools,
tool_choice=ToolChoiceCustom(type="custom", name="emit_command"),
)
assert named == custom_tool_parameters()
assert (
find_tool_properties(tools, "emit_command")
== custom_tool_parameters()["properties"]
)
class TestMalformedToolSchemaDefs:
"""Malformed `$defs` in caller-supplied tool parameters is a 400, not a 500."""