Fix agent-openai-agents-sdk multi-turn failure on 2nd+ prompt (id-less assistant history item)

agent-openai-agents-sdk/agent_server/agent.py

@@ -16,6 +16,7 @@
     ResponsesAgentStreamEvent,
 )
 
+from agent_server.history import normalize_history_items
 from agent_server.utils import (
     build_mcp_url,
     get_session_id,
@@ -99,7 +100,7 @@ async def invoke_handler(request: ResponsesAgentRequest) -> ResponsesAgentRespon
     # for on-behalf-of user authentication.
     async with AsyncExitStack() as stack:
         agent = create_agent()
-        messages = [i.model_dump() for i in request.input]
+        messages = normalize_history_items([i.model_dump() for i in request.input])
         result = await Runner.run(agent, messages)
         return ResponsesAgentResponse(output=[item.to_input_item() for item in result.new_items])
 
@@ -121,7 +122,7 @@ async def stream_handler(
     # for on-behalf-of user authentication.
     async with AsyncExitStack() as stack:
         agent = create_agent()
-        messages = [i.model_dump() for i in request.input]
+        messages = normalize_history_items([i.model_dump() for i in request.input])
         result = Runner.run_streamed(agent, input=messages)
 
         async for event in process_agent_stream_events(result.stream_events()):

agent-openai-agents-sdk/agent_server/history.py

@@ -0,0 +1,40 @@
+"""Conversation-history normalization helpers.
+
+Kept dependency-free (no Databricks SDK / MLflow imports) so it can be unit
+tested in isolation.
+"""
+
+
+def normalize_history_items(messages: list[dict]) -> list[dict]:
+    """Normalize replayed assistant history items for Chat Completions.
+
+    openai-agents >= 0.19 tightened ``Converter.maybe_response_output_message``
+    to require ``{"id", "content"}`` on assistant history items. The built-in
+    chat UI replays the prior assistant turn as an id-less
+    ``{"type": "message", "role": "assistant", "content": [output_text]}`` item,
+    and MLflow's ``ResponsesAgentRequest`` strips any client-supplied ``id``, so
+    on the second and later prompts the item fails recognition and hits the
+    catch-all ``UserError: Unhandled item type or structure``.
+
+    Collapse those replayed assistant items to the easy-input
+    ``{"role", "content"}`` form, which is recognized by
+    ``maybe_easy_input_message`` (no id required), sidestepping the tightened
+    guard regardless of SDK version. Multiple ``output_text`` segments are joined
+    with ``"\\n"`` to match the SDK converter's own behavior byte-for-byte.
+    """
+    normalized: list[dict] = []
+    for m in messages:
+        if (
+            m.get("type") == "message"
+            and m.get("role") == "assistant"
+            and isinstance(m.get("content"), list)
+        ):
+            text = "\n".join(
+                part.get("text", "")
+                for part in m["content"]
+                if isinstance(part, dict) and part.get("type") == "output_text"
+            )
+            normalized.append({"role": "assistant", "content": text})
+        else:
+            normalized.append(m)
+    return normalized

agent-openai-agents-sdk/tests/test_history_normalization.py

@@ -0,0 +1,85 @@
+"""Regression tests for multi-turn assistant-history normalization.
+
+openai-agents >= 0.19 tightened ``Converter.maybe_response_output_message`` to
+require ``{"id", "content"}`` on assistant history items. The built-in chat UI
+replays the prior assistant turn as an id-less ``{"type": "message",
+"role": "assistant", "content": [output_text]}`` item, and MLflow's
+``ResponsesAgentRequest`` strips any client-supplied ``id``, so on the second
+and later prompts the item failed recognition and hit the catch-all
+``UserError: Unhandled item type or structure``.
+
+``normalize_history_items`` collapses those replayed assistant items to the
+easy-input ``{"role", "content"}`` form, which is recognized without an id.
+These tests lock in that transformation so the regression cannot return silently.
+"""
+
+import unittest
+
+from agent_server.history import normalize_history_items
+
+
+class NormalizeHistoryItemsTest(unittest.TestCase):
+    def test_idless_assistant_output_text_collapsed_to_easy_input(self):
+        # Exactly what the chat UI replays on the 2nd prompt (no "id").
+        messages = [
+            {"role": "user", "content": "hello"},
+            {
+                "type": "message",
+                "role": "assistant",
+                "content": [{"type": "output_text", "text": "Hi! How can I help?"}],
+            },
+            {"role": "user", "content": "test"},
+        ]
+        self.assertEqual(
+            normalize_history_items(messages),
+            [
+                {"role": "user", "content": "hello"},
+                {"role": "assistant", "content": "Hi! How can I help?"},
+                {"role": "user", "content": "test"},
+            ],
+        )
+
+    def test_multiple_output_text_segments_joined_with_newline(self):
+        # Matches the SDK converter, which joins segments with "\n".
+        messages = [
+            {
+                "type": "message",
+                "role": "assistant",
+                "content": [
+                    {"type": "output_text", "text": "part A"},
+                    {"type": "output_text", "text": "part B"},
+                ],
+            }
+        ]
+        self.assertEqual(
+            normalize_history_items(messages),
+            [{"role": "assistant", "content": "part A\npart B"}],
+        )
+
+    def test_user_and_plain_assistant_items_are_untouched(self):
+        messages = [
+            {"role": "user", "content": "hi"},
+            {"role": "assistant", "content": "already easy-input"},
+        ]
+        self.assertEqual(normalize_history_items(messages), messages)
+
+    def test_non_output_text_content_parts_are_dropped_from_text(self):
+        # Only output_text parts contribute to the collapsed string.
+        messages = [
+            {
+                "type": "message",
+                "role": "assistant",
+                "content": [
+                    {"type": "output_text", "text": "keep"},
+                    {"type": "something_else", "foo": "bar"},
+                ],
+            }
+        ]
+        self.assertEqual(
+            normalize_history_items(messages),
+            [{"role": "assistant", "content": "keep"}],
+        )
+
+
+if __name__ == "__main__":
+    unittest.main()