Python: Persist hosted MCP call/results as canonical mcp_call output

[Hameed Kunkanoor (Hameedkunkanoor)]

Motivation and Context

This PR introduces a standalone fix for hosted MCP relay persistence and replay reliability in Agent Framework foundry hosting.
Fix for #5546

Problem addressed:

• Follow-up turns could fail with a 400 due to replayed hosted-MCP history containing an unbalanced tool-output shape.
• Historically, hosted MCP call and result could be persisted in split form, which could replay as orphaned tool output.

Scenario supported:

• Cross-turn hosted MCP conversations where tool call identity and tool result must remain paired in persisted/replayed history.

Dependency and rollout note:

• This change depends on SDK APIs added in azure-ai-agentserver-responses 1.0.0b6.
• Merge and publish SDK changes first Add MCP builder support for output/error and caller-supplied item IDs Azure/azure-sdk-for-python#46985 , then merge this PR.

Description

Overall approach:

• Move foundry_hosting write-side persistence to canonical single-item mcp_call representation and keep replay reconstruction aligned with that shape.

Changes made:

1. Preserve original MCP call id when opening MCP builders.
2. Streaming conversion path: mcp_server_tool_result now completes the active mcp_call builder instead of emitting custom_tool_call_output fallback.
3. Non-streaming conversion path: adjacent mcp_server_tool_call + mcp_server_tool_result are coalesced into one completed mcp_call output item.
4. Replay reconstruction:
   • persisted mcp_call items with output are reconstructed back into MCP call/result content.
5. Dependency safety:
   • bump foundry_hosting dependency floor from azure-ai-agentserver-responses>=1.0.0b5,<2 to >=1.0.0b6,<2.
6. Test coverage:
   • Added focused and scenario-level regressions for streaming/non-streaming canonical persistence, replay reconstruction for both item shapes, and multi-turn round-trip behavior.

Testing performed:

• Working directory:

   Code/agent-framework/python/packages/foundry_hosting
  

• Commands:

   python -m pip install -e ../../
   python -m pytest tests/test_responses.py -q -k "hosted_mcp_call_and_result_persist_as_single_mcp_call or mcp_tool_call_and_result_streaming_emit_single_completed_mcp_call or mcp_call_with_output_reconstructs_mcp_result_content or hosted_mcp_call_round_trip_does_not_orphan_function_call_output"
   python -m pytest tests/test_responses.py -v -k "mcp_call_with_output_reconstructs_mcp_result_content or test_mcp_call"
   python -m pytest tests/test_responses.py tests/test_responses_int.py -q
  

• Results:
  • Focused regression selection: 4 passed
  • Additional replay-side check: 4 passed, 134 deselected
  • foundry_hosting suite: 135 passed, 13 skipped

Compatibility verification:

• Working directory:

   Code/agent-framework/python/packages/openai
  

• Commands:

   python -m pytest tests/openai/test_openai_chat_client.py -q -k "hosted_MCP_round_trip or coalesces_mcp_call_and_result_into_single_item or drops_orphan_mcp_server_tool_result or serializes_mcp_server_tool_call_as_mcp_call_input_item"
   python -m pytest tests/openai/test_openai_chat_client.py -q -k "coalesces_mcp_call_and_result_into_single_item or drops_orphan_mcp_server_tool_result or serializes_mcp_server_tool_call_as_mcp_call_input_item"
  

• Results:
  • 3 selected tests passed
  • 3 passed on rerun

Files changed:

• python/packages/foundry_hosting/agent_framework_foundry_hosting/_responses.py
• python/packages/foundry_hosting/tests/test_responses.py
• python/packages/foundry_hosting/pyproject.toml

Contribution Checklist

• The code builds clean without any errors or warnings
• The PR follows the Contribution Guidelines
• All unit tests pass, and I have added new tests where possible
• Is this a breaking change? If yes, add "[BREAKING]" prefix to the title of the PR.

[Hameed Kunkanoor (Hameedkunkanoor)]

The tests shall fail because we are using a unreleased version in our code

[Copilot]

Pull request overview

Note

Copilot was unable to run its full agentic suite in this review.

Persists hosted MCP call/result pairs as a single canonical mcp_call output item (with output populated) on the foundry hosting write path, and reconstructs them back into paired MCP call/result content on replay. Fixes follow-up turn 400 errors caused by orphaned tool-output items in replayed history (#5546).

Changes:

• Streaming and non-streaming conversion paths now coalesce mcp_server_tool_call + mcp_server_tool_result into a single completed mcp_call item keyed by the original call_id.
• Replay (_item_to_message / _output_item_to_message) expands mcp_call items carrying output back into both an MCP call and MCP result content.
• Bumps azure-ai-agentserver-responses floor to >=1.0.0b6 to pick up the required SDK APIs.

Reviewed changes

Copilot reviewed 3 out of 3 changed files in this pull request and generated 4 comments.

File	Description
python/packages/foundry_hosting/agent_framework_foundry_hosting/_responses.py	Implements coalescing, MCP builder id propagation, output stringification, and reconstruction logic.
python/packages/foundry_hosting/tests/test_responses.py	Adds regression tests for single-item persistence, streaming coalescing, replay reconstruction, and multi-turn round-trip.
python/packages/foundry_hosting/pyproject.toml	Raises azure-ai-agentserver-responses minimum to 1.0.0b6.

[Hameed Kunkanoor (Hameedkunkanoor)]

Copilot apply changes based on this feedback

[Evan Mattson (moonbox3)]

Per our offline discussion, please re-open the PR when the underlying Python SDK changes are in and the new package is released. Thanks.