From ca96c8a5fac49307e4688a7693a4d4e9dc639f7a Mon Sep 17 00:00:00 2001
From: thrialectics <203729272+thrialectics@users.noreply.github.com>
Date: Thu, 30 Apr 2026 12:08:56 -0400
Subject: [PATCH] =?UTF-8?q?fix(llm/openai):=20normalize=20tool=5Fchoice=20?=
 =?UTF-8?q?"any"=20=E2=86=92=20"required"?=
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

OpenAI rejects tool_choice="any" with a 400 (only accepts none/auto/required),
but Anthropic and Gemini backends both treat "any" and "required" as
equivalent (anthropic.py:321, gemini.py:568), and tool_loop.py:382 assumes
the same. Callers passing "any" — including via env/config overrides not
covered by the default-only patch in #630 — would still hit the error.

Map "any" → "required" inside the OpenAI backend so callers can use either
spelling without provider-specific awareness.
---
 src/llm/backends/openai.py             |  4 ++
 tests/llm/test_backends/test_openai.py | 53 ++++++++++++++++++++++++++
 2 files changed, 57 insertions(+)

diff --git a/src/llm/backends/openai.py b/src/llm/backends/openai.py
index 1e01e78a..c6e49f50 100644
--- a/src/llm/backends/openai.py
+++ b/src/llm/backends/openai.py
@@ -309,6 +309,10 @@ class OpenAIBackend:
         if tools:
             params["tools"] = self._convert_tools(tools)
             if tool_choice is not None:
+                # OpenAI accepts only "none" | "auto" | "required" (or a function spec).
+                # Other backends use "any" with the same semantics as "required".
+                if tool_choice == "any":
+                    tool_choice = "required"
                 params["tool_choice"] = tool_choice
         if extra_params:
             for key in (
diff --git a/tests/llm/test_backends/test_openai.py b/tests/llm/test_backends/test_openai.py
index 81838202..862fbfa9 100644
--- a/tests/llm/test_backends/test_openai.py
+++ b/tests/llm/test_backends/test_openai.py
@@ -230,6 +230,59 @@ async def test_openai_backend_converts_anthropic_style_tools() -> None:
     assert call["tool_choice"] == "required"
 
 
+@pytest.mark.asyncio
+async def test_openai_backend_normalizes_any_tool_choice_to_required() -> None:
+    """OpenAI rejects tool_choice="any" (Anthropic/Gemini's spelling for the
+    same semantics OpenAI calls "required"). The backend must translate it
+    so callers can pass either spelling without provider-specific awareness.
+    """
+    client = Mock()
+    client.chat.completions.create = AsyncMock(
+        return_value=SimpleNamespace(
+            choices=[
+                SimpleNamespace(
+                    finish_reason="stop",
+                    message=SimpleNamespace(
+                        content="ok",
+                        tool_calls=[],
+                        reasoning_details=[],
+                    ),
+                )
+            ],
+            usage=SimpleNamespace(
+                prompt_tokens=10,
+                completion_tokens=5,
+                prompt_tokens_details=None,
+            ),
+        )
+    )
+
+    backend = OpenAIBackend(client)
+    await backend.complete(
+        model="gpt-4.1",
+        messages=[{"role": "user", "content": "Hello"}],
+        max_tokens=100,
+        tools=[
+            {
+                "name": "get_weather",
+                "description": "Lookup weather",
+                "input_schema": {
+                    "type": "object",
+                    "properties": {"city": {"type": "string"}},
+                    "required": ["city"],
+                },
+            }
+        ],
+        tool_choice="any",
+    )
+
+    await_args = client.chat.completions.create.await_args
+    if await_args is None:
+        raise AssertionError("Expected OpenAI create call")
+    call = await_args.kwargs
+    assert call["tool_choice"] == "required"
+
+
 @pytest.mark.parametrize(
     "model",
     [