fix: Multi turn evaluation causes JSON serialization error

py/src/braintrust/functions/invoke.py

@@ -3,6 +3,7 @@
 from sseclient import SSEClient
 
 from .._generated_types import FunctionTypeEnum
+from ..bt_json import bt_dumps
 from ..logger import Exportable, _internal_get_global_state, get_span_parent_object, login, proxy_conn
 from ..util import response_raise_for_status
 from .constants import INVOKE_API_VERSION
@@ -201,7 +202,8 @@ def invoke(
     if org_name is not None:
         headers["x-bt-org-name"] = org_name
 
-    resp = proxy_conn().post("function/invoke", json=request, headers=headers, stream=stream)
+    request_json = bt_dumps(request)
+    resp = proxy_conn().post("function/invoke", data=request_json, headers=headers, stream=stream)
     if resp.status_code == 500:
         raise BraintrustInvokeError(resp.text)
 

py/src/braintrust/functions/test_invoke.py

@@ -1,7 +1,10 @@
 """Tests for the invoke module, particularly init_function."""
 
+import json
+from unittest.mock import MagicMock, patch
 
-from braintrust.functions.invoke import init_function
+import pytest
+from braintrust.functions.invoke import init_function, invoke
 from braintrust.logger import _internal_get_global_state, _internal_reset_global_state
 
 
@@ -59,3 +62,55 @@ def test_init_function_permanently_disables_cache(self):
         # Try to start again - should still be disabled because of explicit disable
         state.span_cache.start()
         assert state.span_cache.disabled is True
+
+
+def _invoke_with_messages(messages):
+    """Call invoke() with mocked proxy_conn; return the parsed request body."""
+    mock_resp = MagicMock()
+    mock_resp.status_code = 200
+    mock_resp.json.return_value = {}
+    mock_conn = MagicMock()
+    mock_conn.post.return_value = mock_resp
+
+    with (
+        patch("braintrust.functions.invoke.login"),
+        patch("braintrust.functions.invoke.get_span_parent_object") as mock_parent,
+        patch("braintrust.functions.invoke.proxy_conn", return_value=mock_conn),
+    ):
+        mock_parent.return_value.export.return_value = "span-export"
+        invoke(project_name="test-project", slug="test-fn", messages=messages)
+
+    kwargs = mock_conn.post.call_args.kwargs
+    assert "data" in kwargs, "invoke must use data= (bt_dumps) not json= (json.dumps) (see issue 38)"
+    assert "json" not in kwargs
+    return json.loads(kwargs["data"])
+
+
+def test_invoke_serializes_openai_messages():
+    openai_chat = pytest.importorskip("openai.types.chat")
+    msg = openai_chat.ChatCompletionMessage(role="assistant", content="The answer is X.")
+    parsed = _invoke_with_messages([msg])
+    assert isinstance(parsed, dict) and parsed
+
+
+def test_invoke_serializes_anthropic_messages():
+    anthropic_types = pytest.importorskip("anthropic.types")
+    msg = anthropic_types.Message(
+        id="msg_123",
+        type="message",
+        role="assistant",
+        content=[anthropic_types.TextBlock(type="text", text="The answer is X.")],
+        model="claude-3-5-sonnet-20241022",
+        stop_reason="end_turn",
+        stop_sequence=None,
+        usage=anthropic_types.Usage(input_tokens=10, output_tokens=20),
+    )
+    parsed = _invoke_with_messages([msg])
+    assert isinstance(parsed, dict) and parsed
+
+
+def test_invoke_serializes_google_messages():
+    google_types = pytest.importorskip("google.genai.types")
+    msg = google_types.Content(role="model", parts=[google_types.Part(text="The answer is X.")])
+    parsed = _invoke_with_messages([msg])
+    assert isinstance(parsed, dict) and parsed