From 66890bd79b09e41d71966a3740804e69ff4d37f3 Mon Sep 17 00:00:00 2001
From: pablopupo <145598901+pablopupo@users.noreply.github.com>
Date: Thu, 2 Jul 2026 08:30:16 -0400
Subject: [PATCH] [Bugfix] Clear response format constraints when tool_choice
 is auto

Fixes #39929. With tools present and tool_choice auto, unset, or null,
a response format constraint boxed decoding into plain JSON so the
model could never emit tool-call tokens. Clear response_format on
ChatCompletionRequest and the text.format analog on ResponsesRequest
in ToolParser.adjust_request for the auto case, keeping both for
tool_choice none. Adds parser-level regression tests for both request
types and an E2E test mirroring the existing required-variant test,
excluded from the Rust frontend job the same way that variant is.

Signed-off-by: pablopupo <145598901+pablopupo@users.noreply.github.com>
---
 .buildkite/test_areas/rust_frontend.yaml      |   2 +-
 tests/tool_use/test_chat_completions.py       |  40 +++++
 .../test_tool_parser_adjust_request.py        | 164 ++++++++++++++++++
 vllm/tool_parsers/abstract_tool_parser.py     |  16 ++
 4 files changed, 221 insertions(+), 1 deletion(-)
 create mode 100644 tests/tool_use/test_tool_parser_adjust_request.py

diff --git a/.buildkite/test_areas/rust_frontend.yaml b/.buildkite/test_areas/rust_frontend.yaml
index adb27c4a049b..285b02da87ca 100644
--- a/.buildkite/test_areas/rust_frontend.yaml
+++ b/.buildkite/test_areas/rust_frontend.yaml
@@ -85,7 +85,7 @@ steps:
   commands:
   - export VLLM_USE_RUST_FRONTEND=1
   - export VLLM_WORKER_MULTIPROC_METHOD=spawn
-  - pytest -v -s tool_use --ignore=tool_use/mistral --models llama3.2 -k "not test_response_format_with_tool_choice_required and not test_parallel_tool_calls_false and not test_tool_call_and_choice"
+  - pytest -v -s tool_use --ignore=tool_use/mistral --models llama3.2 -k "not test_response_format_with_tool_choice_required and not test_response_format_with_tool_choice_auto and not test_parallel_tool_calls_false and not test_tool_call_and_choice"
 
 - label: Rust Frontend Distributed
   timeout_in_minutes: 30
diff --git a/tests/tool_use/test_chat_completions.py b/tests/tool_use/test_chat_completions.py
index e5bb475875ac..c21207eb9094 100644
--- a/tests/tool_use/test_chat_completions.py
+++ b/tests/tool_use/test_chat_completions.py
@@ -5,6 +5,7 @@
 import pytest
 
 from .utils import (
+    MESSAGES_ASKING_FOR_TOOLS,
     MESSAGES_WITHOUT_TOOLS,
     SEED,
     WEATHER_TOOL,
@@ -198,3 +199,42 @@ async def test_response_format_with_tool_choice_required(
     assert choice.finish_reason == "tool_calls"
     assert choice.message.tool_calls is not None
     assert len(choice.message.tool_calls) > 0
+
+
+# Regression test for https://github.com/vllm-project/vllm/issues/39929
+# response_format: json_object suppressed tool calls when tool_choice: auto
+@pytest.mark.asyncio
+@pytest.mark.timeout(120)
+async def test_response_format_with_tool_choice_auto(
+    client: openai.AsyncOpenAI, server_config: ServerConfig
+):
+    """
+    Test that tool calls still work when response_format: json_object is
+    combined with tool_choice: auto.
+
+    Before the fix, response_format constrained decoding to plain JSON, so
+    the model could never emit tool-call tokens and answered with raw JSON
+    content instead of calling the tool.
+    """
+    models = await client.models.list()
+    model_name: str = models.data[0].id
+
+    # Same fixture and seed as the auto-choice tests in test_tool_calls.py,
+    # since calling the tool is a free model decision here, not a forced one
+    chat_completion = await client.chat.completions.create(
+        messages=ensure_system_prompt(MESSAGES_ASKING_FOR_TOOLS, server_config),
+        temperature=0,
+        max_completion_tokens=150,
+        model=model_name,
+        tools=[WEATHER_TOOL],
+        tool_choice="auto",
+        response_format={"type": "json_object"},
+        seed=SEED,
+    )
+
+    # response_format is dropped for auto, leaving the model free to call
+    # the weather tool for a weather question
+    choice = chat_completion.choices[0]
+    assert choice.finish_reason == "tool_calls"
+    assert choice.message.tool_calls is not None
+    assert len(choice.message.tool_calls) > 0
diff --git a/tests/tool_use/test_tool_parser_adjust_request.py b/tests/tool_use/test_tool_parser_adjust_request.py
new file mode 100644
index 000000000000..6cfdce3d298a
--- /dev/null
+++ b/tests/tool_use/test_tool_parser_adjust_request.py
@@ -0,0 +1,164 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+"""Regression tests for response_format handling in ToolParser.adjust_request.
+
+With tool_choice="auto" no schema is derived from the tools, so a
+user-supplied response_format used to stay on the request and constrain
+decoding to plain JSON, which prevented the model from ever emitting
+tool-call tokens (https://github.com/vllm-project/vllm/issues/39929).
+"""
+
+from __future__ import annotations
+
+from typing import Any
+
+from vllm.entrypoints.openai.chat_completion.protocol import ChatCompletionRequest
+from vllm.entrypoints.openai.responses.protocol import ResponsesRequest
+from vllm.tool_parsers.abstract_tool_parser import ToolParser
+
+WEATHER_TOOL = {
+    "type": "function",
+    "function": {
+        "name": "get_weather",
+        "description": "Get current weather for a city",
+        "parameters": {
+            "type": "object",
+            "properties": {"city": {"type": "string"}},
+            "required": ["city"],
+        },
+    },
+}
+
+WEATHER_TOOL_RESPONSES = {
+    "type": "function",
+    "name": "get_weather",
+    "description": "Get current weather for a city",
+    "parameters": {
+        "type": "object",
+        "properties": {"city": {"type": "string"}},
+        "required": ["city"],
+    },
+    "strict": True,
+}
+
+
+def _build_request(**overrides: Any) -> ChatCompletionRequest:
+    data: dict[str, Any] = {
+        "model": "test-model",
+        "messages": [{"role": "user", "content": "What is the weather in Dallas?"}],
+        "tools": [WEATHER_TOOL],
+        "response_format": {"type": "json_object"},
+        **overrides,
+    }
+    return ChatCompletionRequest.model_validate(data)
+
+
+def _build_responses_request(**overrides: Any) -> ResponsesRequest:
+    data: dict[str, Any] = {
+        "model": "test-model",
+        "input": [{"role": "user", "content": "What is the weather in Dallas?"}],
+        "tools": [WEATHER_TOOL_RESPONSES],
+        "text": {"format": {"type": "json_object"}},
+        **overrides,
+    }
+    return ResponsesRequest.model_validate(data)
+
+
+def _adjust(
+    request: ChatCompletionRequest | ResponsesRequest,
+) -> ChatCompletionRequest | ResponsesRequest:
+    parser = ToolParser.__new__(ToolParser)
+    return ToolParser.adjust_request(parser, request)
+
+
+def test_auto_clears_response_format() -> None:
+    """auto: response_format must be cleared so tool-call tokens stay
+    reachable, and no structured output constraint is added."""
+    request = _build_request(tool_choice="auto")
+
+    _adjust(request)
+
+    assert request.response_format is None
+    assert request.structured_outputs is None
+
+
+def test_unset_tool_choice_clears_response_format() -> None:
+    """Unset tool_choice with tools present defaults to auto during request
+    validation and must be treated the same."""
+    request = _build_request()
+    assert request.tool_choice == "auto", (
+        "Precondition: unset tool_choice defaults to auto when tools are set"
+    )
+
+    _adjust(request)
+
+    assert request.response_format is None
+
+
+def test_null_tool_choice_clears_response_format() -> None:
+    """An explicit null tool_choice passes validation untouched and must be
+    treated like auto."""
+    request = _build_request(tool_choice=None)
+    assert request.tool_choice is None
+
+    _adjust(request)
+
+    assert request.response_format is None
+
+
+def test_none_tool_choice_preserves_response_format() -> None:
+    """none: the caller asked for a formatted reply with tool calling
+    disabled, so response_format must survive."""
+    request = _build_request(tool_choice="none")
+
+    _adjust(request)
+
+    assert request.response_format is not None
+    assert request.response_format.type == "json_object"
+
+
+def test_required_overrides_response_format_with_tool_schema() -> None:
+    """required: the schema derived from the tools replaces response_format,
+    same as before this fix (#32006)."""
+    request = _build_request(tool_choice="required")
+
+    _adjust(request)
+
+    assert request.response_format is None
+    assert request.structured_outputs is not None
+    assert request.structured_outputs.json is not None
+
+
+def test_no_tools_preserves_response_format() -> None:
+    """Without tools adjust_request returns early and response_format is
+    untouched."""
+    request = _build_request(tools=None, tool_choice=None)
+
+    _adjust(request)
+
+    assert request.response_format is not None
+    assert request.response_format.type == "json_object"
+
+
+def test_responses_auto_clears_text_format() -> None:
+    """Responses API: text.format is the response_format analog and must be
+    dropped for auto so tool-call tokens stay reachable."""
+    request = _build_responses_request()
+    assert request.tool_choice == "auto", (
+        "Precondition: ResponsesRequest.tool_choice defaults to auto"
+    )
+    assert request.text is not None and request.text.format is not None
+
+    _adjust(request)
+
+    assert request.text is None or request.text.format is None
+
+
+def test_responses_none_tool_choice_preserves_text_format() -> None:
+    """Responses API: tool_choice none keeps the caller's text format."""
+    request = _build_responses_request(tool_choice="none")
+
+    _adjust(request)
+
+    assert request.text is not None
+    assert request.text.format is not None
diff --git a/vllm/tool_parsers/abstract_tool_parser.py b/vllm/tool_parsers/abstract_tool_parser.py
index a1c4cf1ffaeb..db7044601f4c 100644
--- a/vllm/tool_parsers/abstract_tool_parser.py
+++ b/vllm/tool_parsers/abstract_tool_parser.py
@@ -162,6 +162,22 @@ def adjust_request(
                         strict=True,
                     )
                 )
+        elif request.tool_choice in ("auto", None):
+            # tool_choice "auto" (or unset, which defaults to auto when tools
+            # are provided) must leave the model free to emit tool-call
+            # tokens, which a response format constraint would prevent.
+            # tool_choice "none" keeps the response format since tools are
+            # disabled anyway.
+            if isinstance(request, ChatCompletionRequest):
+                request.response_format = None
+            if (
+                isinstance(request, ResponsesRequest)
+                and request.text is not None
+                and request.text.format is not None
+            ):
+                # Single-shot copy for the same Pydantic v2 reason as above,
+                # keeping unrelated text settings intact
+                request.text = request.text.model_copy(update={"format": None})
 
         return request