Python: fix reasoning model workflow handoff and history serialization

[eavanvalkenburg]

Summary

Fixes multiple related failures when using reasoning models (gpt-5-mini, gpt-5.2) in multi-agent workflows. The root issues are all about how reasoning items from the Responses API are emitted, serialized, and carried into subsequent agent runs.

Closes #4047

---

Problems Fixed

1. "reasoning was provided without its required following item"

The Responses API only accepts a reasoning item in input when it directly precedes a function_call. Sending a reasoning item that preceded a text response (no tool call) causes an API error.

Fix: _prepare_message_for_openai now checks whether the message contains a function_call. text_reasoning content is only serialized as a reasoning input item when a function_call is also present in the same message.

2. Reasoning items never emitted for encrypted/hidden reasoning

When a reasoning model produces encrypted or hidden reasoning, the output_item.added event fires with an empty content list and no reasoning_text.delta events follow. Previously, no text_reasoning Content was emitted — making it invisible to downstream serialization logic.

Fix: Both _parse_response_from_openai (non-streaming) and the output_item.added handler (streaming) now always emit at least one text_reasoning Content, even when the text is empty. The reasoning_id and encrypted_content (if present) are stored in additional_properties.

3. summary field must be an array, not an object

The summary field on a reasoning input item must be an array of objects ([{"type": "summary_text", "text": ...}]), not a single object. This caused a 400 invalid_type error.

Fix: _prepare_content_for_openai now wraps summary in a list. summary is omitted entirely when there is no visible text (e.g. encrypted reasoning, where only encrypted_content is sent).

---

Files Changed

File	Change
packages/core/agent_framework/openai/_responses_client.py	Always emit text_reasoning on reasoning output items; fix summary to be an array; skip reasoning serialization when no function_call in same message
packages/core/agent_framework/_workflows/_agent_executor.py	Clear service_session_id in run and from_response handlers; remove no-op _prepare_handoff_messages
packages/core/tests/workflow/test_full_conversation.py	Add test_run_request_with_full_history_clears_service_session_id and test_from_response_clears_service_session_id (TDD: fail without fix, pass with fix)

[Copilot]

Pull request overview

This PR fixes workflow + function-calling failures when using reasoning-capable models with the OpenAI/Azure Responses API by tightening how reasoning items are emitted/serialized and by preventing duplicate history replay across agent handoffs.

Changes:

• Adjusts Responses API parsing/serialization to (a) only include reasoning input items when paired with a function_call, (b) always emit a text_reasoning marker (even empty) for hidden/encrypted reasoning, and (c) serialize summary as an array.
• Updates workflow execution to clear service_session_id when explicitly replaying full history to avoid “Duplicate item found” errors.
• Improves function-invocation behavior across multi-message responses and adds/expands tests (unit + integration) covering these scenarios.

Reviewed changes

Copilot reviewed 8 out of 8 changed files in this pull request and generated 6 comments.

Show a summary per file

File	Description
python/packages/core/agent_framework/openai/_responses_client.py	Updates reasoning item parsing and input serialization rules for Responses API.
python/packages/core/agent_framework/_workflows/_agent_executor.py	Clears service_session_id when replaying explicit history into an executor.
python/packages/core/agent_framework/_tools.py	Improves function-call extraction across multiple messages and adjusts stop-path handling.
python/packages/core/tests/workflow/test_full_conversation.py	Adds workflow tests for handoff history and service_session_id clearing.
python/packages/core/tests/core/test_function_invocation_logic.py	Adds tests for multi-message function calls and stop-path conversation_id behavior.
python/packages/core/tests/azure/test_azure_responses_client.py	Adds an integration test that validates minimal workflow handoff across reasoning vs non-reasoning deployments.
python/samples/05-end-to-end/workflow_evaluation/run_evaluation.py	Updates the default workflow deployment name to a reasoning model for the evaluation sample.
python/samples/02-agents/conversations/redis_chat_message_store_session.py	Makes Redis URL configurable via REDIS_URL env var and updates sample messaging.

[markwallace-microsoft]

Python Test Coverage Report •

File	Stmts	Miss	Cover	Missing
packages/core/agent_framework
_tools.py	846	91	89%	165–166, 303, 305, 323–325, 332, 350, 364, 371, 378, 394, 396, 403, 440, 465, 469, 486–488, 535–537, 600, 622, 685–691, 727, 738–749, 771–773, 778, 782, 796–798, 837, 906, 916, 926, 982, 1013, 1032, 1294, 1351, 1371, 1442–1446, 1568, 1572, 1596, 1622, 1624, 1640, 1642, 1727, 1757, 1777, 1779, 1830, 1893, 2077–2078, 2115, 2128, 2138–2139, 2174–2175, 2235
_types.py	998	87	91%	49, 58–59, 113, 118, 137, 139, 143, 147, 149, 151, 153, 171, 175, 201, 223, 228, 233, 237, 263, 267, 615–616, 987, 1049, 1066, 1084, 1089, 1107, 1117, 1134–1135, 1137, 1155–1156, 1158, 1165–1166, 1168, 1203, 1214–1215, 1217, 1255, 1482, 1534, 1625–1630, 1652, 1657, 1823, 1835, 2078, 2087, 2108, 2203, 2428, 2635, 2705, 2717, 2735, 2933–2935, 2938–2940, 2944, 2949, 2953, 3037–3039, 3068, 3122, 3141–3142, 3145–3149, 3155
packages/core/agent_framework/_workflows
_agent_executor.py	200	26	87%	97, 113, 168–169, 221–222, 224–225, 255–257, 265–267, 275–277, 279, 283, 287, 291–292, 391–392, 438, 456
packages/core/agent_framework/openai
_responses_client.py	639	87	86%	290–293, 297–298, 301–302, 308–309, 314, 327–333, 354, 362, 385, 548, 551, 606, 610, 612, 614, 616, 692, 702, 707, 750, 829, 846, 859, 920, 1011, 1016, 1020–1022, 1026–1027, 1050, 1119, 1141–1142, 1157–1158, 1176–1177, 1218–1221, 1330–1331, 1347, 1349, 1428–1436, 1555, 1610, 1625, 1668–1671, 1679–1680, 1682–1684, 1698–1700, 1710–1711, 1717, 1732
TOTAL	21261	3314	84%

Python Unit Test Overview

Tests	Skipped	Failures	Errors	Time
4189	240 💤	0 ❌	0 🔥	1m 14s ⏱️