[azure-ai-agentserver-responses] Failed response inputs are replayed in conversation history

[eavanvalkenburg]

• Package Name: azure-ai-agentserver-responses
• Package Version: 2.2.0b2 on current main; originally observed through agent-framework-foundry-hosting 1.0.0b260730
• Operating System: Cross-platform; reproduced with a hosted Foundry agent and locally
• Python Version: 3.13.14 in the original report

Describe the bug

AgentServer persists a stored response and its input_items when it processes the initial response.created / in_progress event. If the handler later emits response.failed, update_response() updates the response envelope but does not change the stored input references.

get_history_item_ids() subsequently includes those input IDs:

• when resolving responses in a conversation_id; and
• when a failed standalone response is later chained through previous_response_id.

As a result, invalid input that caused one request to fail is replayed into subsequent otherwise-valid requests, potentially poisoning the conversation indefinitely.

Original report and reproduction: microsoft/agent-framework#7630

Agent Framework workaround under review: microsoft/agent-framework#7637

That workaround has to buffer synchronous handler events so it knows the terminal status before AgentServer performs its initial create, and it wraps/mutates AgentServer's private providers. It cannot address streaming and background failures cleanly without delaying their streams.

To reproduce

1. Create a conversation.
2. Submit a response containing an unmatched function_call_output:

response = client.responses.create(
    conversation=conversation.id,
    input=[
        {"role": "user", "content": "Hello"},
        {
            "type": "function_call_output",
            "call_id": "call_that_does_not_exist",
            "output": "invalid output",
        },
    ],
)

3. Confirm the response fails because no matching function call exists.
4. Submit a valid request to the same conversation:

response2 = client.responses.create(
    conversation=conversation.id,
    input="Hello, how are you?",
)

5. Observe that the second request fails with the same unmatched-function-call error because the first request's input was included in resolved conversation history.

The equivalent issue occurs if step 2 creates a stored standalone failed response and step 4 uses previous_response_id=response.id.

Actual behavior

get_history_item_ids() includes the failed response's own input IDs, so subsequent handlers receive and replay the invalid input.

Expected behavior

Failed-response inputs should remain stored and retrievable through /responses/{id}/input_items for diagnostics, but should not be returned as replayable history for either conversation_id or previous_response_id. Successful-response history should remain unchanged.

Ownership rationale

This behavior is defined by azure-ai-agentserver-responses:

• its orchestrator decides when response inputs are persisted;
• ResponseProviderProtocol defines create, update, and history operations;
• its in-memory and file providers build conversation and previous-response history; and
• its Foundry provider delegates to the hosted history/item_ids storage endpoint.

Fixing this only in Agent Framework's foundry_hosting adapter leaves other AgentServer hosts and streaming/background request modes with the same behavior.

Suggested direction

Define and enforce a provider invariant that a failed response's own input_item_ids are excluded from replayable history while the stored items remain available for diagnostic retrieval.

• Apply this before history-limit truncation.
• Update the in-memory and file providers.
• Apply the same behavior in the Foundry history/item_ids backend, or expose a supported server-side filter.
• Add tests for failed conversation turns and failed standalone responses later chained with previous_response_id, across synchronous, streaming, and background modes.

Related:

• Python: [Bug]: Foundry Hosting - User messages added to conversation chat history for failed requests microsoft/agent-framework#7630
• Python: omit failed Foundry turns from conversation chat history microsoft/agent-framework#7637