Python: fix runtime response_format on foundry agent endpoint

python/packages/foundry/agent_framework_foundry/_agent.py

@@ -351,6 +351,20 @@ async def _prepare_options(
         # agent endpoint and tools present. FunctionTools are invoked client-side
         # by the function invocation layer, not sent to the service.
         run_options.pop("model", None)
+        # Strip text/text_format from the request body. The Foundry agent endpoint
+        # rejects per-call ``text`` configuration when an agent is bound with
+        # ``400 invalid_payload "Not allowed when agent is specified."``. The
+        # original ``response_format`` remains in ``options`` and is honored
+        # client-side via ``ChatResponse``'s lazy structured-value parsing in
+        # ``_parse_response_from_openai``. The bound agent itself must be
+        # configured to emit JSON compatible with the requested format. For a
+        # Pydantic ``response_format``, accessing ``response.value`` may raise
+        # ``pydantic.ValidationError`` if the output does not validate. For a
+        # dict / ``json_schema`` ``response_format``, the lazy path only parses
+        # JSON and may raise ``ValueError`` on invalid JSON; it does not
+        # validate against the schema.
+        run_options.pop("text", None)
+        run_options.pop("text_format", None)
         if not self.allow_preview:
             run_options.pop("tools", None)
             run_options.pop("tool_choice", None)

python/packages/foundry/tests/foundry/test_foundry_agent.py

@@ -200,7 +200,7 @@ def my_func() -> str:
 
 
 async def test_raw_foundry_agent_chat_client_prepare_options_strips_client_side_fields() -> None:
-    """Test that _prepare_options strips model and tool-loop fields from run_options."""
+    """Test that _prepare_options strips model, tool-loop, and text fields from run_options."""
 
     mock_project = MagicMock()
     mock_openai = MagicMock()
@@ -225,6 +225,8 @@ def my_func() -> str:
             "tools": [{"type": "function", "function": {"name": "my_func"}}],
             "tool_choice": "auto",
             "parallel_tool_calls": True,
+            "text": {"format": {"type": "json_schema", "name": "x", "schema": {"type": "object"}}},
+            "text_format": MagicMock(),
         },
     ):
         result = await client._prepare_options(
@@ -236,9 +238,72 @@ def my_func() -> str:
     assert "tools" not in result
     assert "tool_choice" not in result
     assert "parallel_tool_calls" not in result
+    assert "text" not in result
+    assert "text_format" not in result
     assert result == {}
 
 
+async def test_raw_foundry_agent_chat_client_prepare_options_strips_text_for_runtime_response_format() -> None:
+    """Issue #5467: per-call response_format must not be sent to the Foundry agent endpoint.
+
+    The Foundry agent endpoint rejects requests that carry per-call ``text``
+    when an agent is bound (``400 invalid_payload "Not allowed when agent is
+    specified."``). The runtime ``response_format`` must instead be honored
+    client-side via ``ChatResponse``'s lazy parsing path.
+    """
+    from pydantic import BaseModel
+
+    class OutputStruct(BaseModel):
+        location: str
+        conditions: str
+
+    mock_project = MagicMock()
+    mock_project.get_openai_client.return_value = MagicMock()
+
+    client = RawFoundryAgentChatClient(
+        project_client=mock_project,
+        agent_name="test-agent",
+    )
+
+    # Simulate what the parent OpenAI Responses client produces for a Pydantic
+    # ``response_format``: it sets ``text_format`` and would route through
+    # ``responses.parse()``.
+    with patch(
+        "agent_framework_openai._chat_client.RawOpenAIChatClient._prepare_options",
+        new_callable=AsyncMock,
+        return_value={"text_format": OutputStruct},
+    ):
+        result = await client._prepare_options(
+            messages=[Message(role="user", contents="hi")],
+            options={"response_format": OutputStruct},
+        )
+
+    assert "text" not in result
+    assert "text_format" not in result
+
+    # And for the dict / json_schema variant the parent populates ``text``.
+    with patch(
+        "agent_framework_openai._chat_client.RawOpenAIChatClient._prepare_options",
+        new_callable=AsyncMock,
+        return_value={
+            "text": {
+                "format": {
+                    "type": "json_schema",
+                    "name": "WeatherDigest",
+                    "schema": {"type": "object", "properties": {"location": {"type": "string"}}},
+                }
+            }
+        },
+    ):
+        result = await client._prepare_options(
+            messages=[Message(role="user", contents="hi")],
+            options={"response_format": {"type": "json_schema", "json_schema": {"name": "WeatherDigest"}}},
+        )
+
+    assert "text" not in result
+    assert "text_format" not in result
+
+
 async def test_raw_foundry_agent_chat_client_prepare_options_maps_agent_session_id_to_extra_body() -> None:
     """Test that service_session_id is forwarded as agent_session_id for hosted sessions."""