Python: fix: clear service_session_id in _agent_wrapper when propagate_session=True

[Benke Qu (benke520)]

Summary

When propagate_session=True, the child agent inherits the parent's session.service_session_id. After the parent's first LLM call, MAF auto-populates this from the Responses API's conversation_id. The child sends it as previous_response_id, which the server rejects because the parent's ool_call is still pending — resulting in a 400 error.

Changes

• _agents.py: In _agent_wrapper (inside Agent.as_tool()), save and clear service_session_id before calling the child agent, restore it in a inally block. This preserves session.state sharing while isolating the server-side conversation pointer.

Tests

Added 3 new tests:

• est_chat_agent_as_tool_propagate_session_clears_service_session_id — verifies child sees service_session_id=None and parent gets it restored
• est_chat_agent_as_tool_propagate_session_restores_service_session_id_on_error — verifies restore happens even if child raises
• est_chat_agent_as_tool_propagate_session_no_service_session_id — verifies no-op when service_session_id is already None

All 106 tests in est_agents.py pass.

Fixes #5874

[Evan Mattson (moonbox3)]

Python Test Coverage Report •

File	Stmts	Miss	Cover	Missing
packages/core/agent_framework
_agents.py	447	49	89%	540, 552, 607, 1124, 1173, 1244–1248, 1347, 1377, 1414, 1447–1448, 1453, 1460, 1512, 1540, 1553, 1605, 1607, 1616–1621, 1626, 1628, 1634–1635, 1642, 1644–1645, 1653–1654, 1657–1659, 1669–1674, 1678, 1683, 1685
TOTAL	44124	5270	88%

Python Unit Test Overview

Tests	Skipped	Failures	Errors	Time
8842	33 💤	0 ❌	0 🔥	2m 16s ⏱️

[Copilot]

Pull request overview

Fixes a Python agent-as-tool session propagation issue where a child agent can inherit the parent’s session.service_session_id, causing invalid previous_response_id usage and 400 errors with Responses-style APIs.

Changes:

• Update Agent.as_tool(... propagate_session=True) wrapper to temporarily clear session.service_session_id during the child agent run and restore it afterward.
• Add tests covering save/clear/restore behavior (including restore on child error).

Reviewed changes

Copilot reviewed 2 out of 2 changed files in this pull request and generated 2 comments.

File	Description
python/packages/core/agent_framework/_agents.py	Clears and restores service_session_id around delegated run() when propagate_session=True.
python/packages/core/tests/core/test_agents.py	Adds tests validating service_session_id isolation/restoration behavior for agent-as-tool.

Comments suppressed due to low confidence (1)

python/packages/core/agent_framework/_agents.py:577

• The finally block only restores service_session_id when the original value was non-None. If the parent session starts with service_session_id=None, the child run can still set session.service_session_id (e.g., from streaming conversation_id propagation) and it will leak back to the parent. Consider always restoring the original value (including None), using a sentinel to distinguish “no session” vs “saved None”.

            finally:
                if session is not None and saved_service_session_id is not None:
                    session.service_session_id = saved_service_session_id

[Benke Qu (benke520)]

Hi Evan Mattson (@moonbox3) Eduard van Valkenburg (@eavanvalkenburg) — I've addressed both Copilot review comments in eb7bc4e (replaced in-place mutation with an isolated child session copy to avoid races under concurrent asyncio.gather). Could you take a look when you get a chance?

[Evan Mattson (moonbox3)]

Benke Qu (@benke520) sorry for the delay. Can you please address the failing CI/CD checks? Thanks.

[Evan Mattson (moonbox3)]

Helping get this past the finish line: I updated the branch with the latest main and fixed the test-typing errors. The refreshed CI is now running.

[Benke Qu (benke520)]

Eduard van Valkenburg (@eavanvalkenburg) Sorry for the accidental re-request — no changes since your approval. The PR keeps getting kicked from the merge queue due to a flaky durabletask integration test (test_multiple_calls_to_same_agent timing out after 90 attempts). This test is unrelated to my changes. Could you advise on next steps — re-run, bypass, or something else? Thanks!