Python: Multi‑turn with AgentFrameworkAgent + AzureAIClient to Azure AI Projects Foundry v2 agent fails on 2nd turn: agent_framework_azure_ai sends invalid input to agent/Responses

[djw-bsn]

I realise that everything here is in beta/preview but I thought it was still worth reporting the bug anyway...

When AzureAIClient is configured to use an Azure AI Projects agent (via AIProjectClient.get_openai_client() + extra_body["agent"]), the input produced by _prepare_input + _openai_content_parser is an OpenAI-style chat array, which Azure AI Projects rejects with invalid_payload (expects string / annotated string).

When using agent_framework_azure_ai.AzureAIClient with an Azure AI Projects Foundry v2 agent (via AIProjectClient.get_openai_client() ), multi‑turn chat via AgentFrameworkAgent fails on the second turn. Internally, the call to Azure AI Projects agent/Responses (/openai/responses) returns 400 invalid_payload, indicating that /input is an array but should be a string, and that /input/1/... is missing required type/annotations.

Example 2nd message:

{
  "messages": [
    {"role": "system",   "content": "System: \n\n"},
    {"role": "system",   "content": "System: \n\n"},
    {"role": "user",     "content": "test 1"},
    {"role": "assistant","content": "Hello! How can I assist you today?"},
    {"role": "user",     "content": "test 2"},
  ],
  "threadId": "XXXX",
}

Partial payload sent to agent responses api:

# ....
    "input": [
        {
            "role": "user",
            "content": [
                {
                    "type": "input_text",
                    "text": "test 1"
                }
            ]
        },
        {
            "role": "assistant",
            "content": [
                {
                    "type": "output_text",
                    "text": "Hello! How can I assist you today?"
                }
            ]
        },
        {
            "role": "user",
            "content": [
                {
                    "type": "input_text",
                    "text": "test 2"
                }
            ]
        }
    ], 
# ....

Server error:

'type: Value is "array" but should be "string"' at '/input',
'required: Required properties ["type"] are not present' at '/input/1',
'type: Value is "array" but should be "string"' at '/input/1/content',
'required: Required properties ["annotations"] are not present' at '/input/1/content/0'.

The error appears to be caused by how agent_framework_azure_ai._prepare_input and _openai_content_parser serialize messages into input.

Steps to reproduce

• Minimal FastAPI server using AzureAIClient + ChatAgent + AgentFrameworkAgent.
• Minimal httpx client posting two AG‑UI style turn bodies (messages + threadId).
• Turn 1 (system + single user) works.
• Turn 2 (system + prior user + prior assistant + new user) triggers the 400 from Azure AI Projects.

(Full server and client repro scripts can be provided if required.)

Key Code Paths

From agent_framework_azure_ai.AzureAIClient:

def _prepare_input(self, messages: MutableSequence[ChatMessage]) -> tuple[list[ChatMessage], str | None]:
    """Prepare input from messages and convert system/developer messages to instructions."""
    result: list[ChatMessage] = []
    instructions_list: list[str] = []
    instructions: str | None = None

    # System/developer messages are turned into instructions, since there is no such message roles in Azure AI.
    for message in messages:
        if message.role.value in ["system", "developer"]:
            for text_content in [content for content in message.contents if isinstance(content, TextContent)]:
                instructions_list.append(text_content.text)
        else:
            result.append(message)

    if len(instructions_list) > 0:
        instructions = "".join(instructions_list)

    return result, instructions

All system / developer messages are collapsed into an instructions string.
All other messages (user + assistant, including history) are left as a list that becomes the basis for input.

super().prepare_options (from OpenAIBaseResponsesClient) then uses _openai_content_parser to turn prepared_messages into the OpenAI Responses input payload.

def _openai_content_parser(
    self,
    role: Role,
    content: Contents,
    call_id_to_id: dict[str, str],
) -> dict[str, Any]:
    """Parse contents into the openai format."""
    match content:
        case TextContent():
            return {
                "type": "output_text" if role == Role.ASSISTANT else "input_text",
                "text": content.text,
            }
        # ...other content types omitted...

This builds OpenAI Responses-style{ "type": "input_text"/"output_text", "text": "..." } objects. Combined with OpenAIBaseResponsesClient, this yields an input that is an array of { role, content: [...] } chat messages. But the Azure AI Projects service rejects that input with the previously listed error.

So for this Foundry v2 agent, agent/Responses is validating input as a string / annotated string, not as a chat array. The OpenAI-style mapping that AzureAIClient currently uses via OpenAIBaseResponsesClient is incompatible with that.

Is it possible to adjust the mapping for this integration so that input is serialized in a way that Azure AI Projects agent/Responses will accept? Or is this a problem in the v2 Foundry agent responses API?

[jimmyshah83]

I tried the PR fix but the issue now happens in the 3rd turn