Python: Fix reasoning replay when store=False

[eavanvalkenburg]

Motivation and Context

Reasoning models can fail when OpenAIChatClient-based clients run with store=False and replay local tool-loop history. In that mode the Responses API cannot resolve response-scoped reasoning items from local history, which breaks scenarios like the Foundry suspend/resume sample.

Description

This updates the shared Responses serialization path to distinguish service-side storage from local storage when replaying prior messages. Reasoning items are now only sent when the request is continuing service-side storage, while local-storage replay keeps the existing function-call replay behavior without reusing service-scoped reasoning items. The change also adds regression coverage for the stateless reasoning/tool-loop path and a sample that reproduces the original Foundry scenario.

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.

[Copilot]

Pull request overview

Note

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

Fixes failures in stateless (store=False) reasoning/tool-loop replay for OpenAIChatClient by preventing replay of service-scoped reasoning items, and adds regression coverage plus a sample reproducer for the Foundry suspend/resume scenario.

Changes:

• Distinguish “service-side storage continuation” vs “local history replay” when serializing Responses input, omitting reasoning items for stateless replay.
• Add regression tests covering tool-loop replay and _prepare_options behavior for store=False with/without a conversation/previous-response identifier.
• Add a Python sample demonstrating suspend/resume using local session/history.

Reviewed changes

Copilot reviewed 3 out of 3 changed files in this pull request and generated 6 comments.

File	Description
python/samples/02-agents/conversations/suspend_resume_local_session.py	Adds a suspend/resume sample intended to reproduce the Foundry local-session scenario.
python/packages/openai/tests/openai/test_openai_chat_client.py	Adds regression tests ensuring reasoning items are omitted for stateless replay, with coverage for conversation-id continuation.
python/packages/openai/agent_framework_openai/_chat_client.py	Updates request serialization to omit service-scoped reasoning during local replay while preserving function-call replay.

[moonbox3]

Python Test Coverage Report •

File	Stmts	Miss	Cover	Missing
packages/openai/agent_framework_openai
_chat_client.py	870	123	85%	522–525, 529–530, 536–537, 547–548, 555, 570–576, 597, 605, 628, 746, 845, 904, 906, 908, 910, 976, 990, 1070, 1080, 1085, 1128, 1244, 1425, 1430, 1434–1436, 1440–1441, 1507, 1536, 1542, 1552, 1558, 1563, 1569, 1574–1575, 1636, 1658–1659, 1674–1675, 1693–1694, 1737, 1900, 1938–1939, 1955, 1957, 2036–2044, 2074, 2181, 2216, 2231, 2251–2261, 2274, 2285–2289, 2303, 2317–2328, 2337, 2369–2372, 2380–2381, 2383–2385, 2399–2401, 2411–2412, 2418, 2433
TOTAL	27547	3193	88%

Python Unit Test Overview

Tests	Skipped	Failures	Errors	Time
5574	20 💤	0 ❌	0 🔥	1m 29s ⏱️

[chetantoshniwal]

Automated Code Review

Reviewers: 4 | Confidence: 83%

✗ Correctness

The PR refactors reasoning-item handling to correctly omit response-scoped reasoning items (rs_*) when replaying tool loops without service-side storage. The core logic is sound: it computes request_uses_service_side_storage by checking for conversation_id, previous_response_id, or conversation keys, then threads that flag through message preparation to skip reasoning items in stateless replay. However, the PR accidentally includes three .worktrees/ submodule entries that should not be committed, and the conversation key check only matches strings, which could miss non-string conversation objects from the OpenAI SDK.

✗ Security Reliability

The core logic change—conditionally omitting response-scoped reasoning items when the request does not use service-side storage—is sound and well-tested. The Message.additional_properties attribute is always initialized to a dict, so the removal of the defensive getattr guard is safe. However, the PR accidentally includes three .worktrees/ gitlink entries that are local development artifacts and should not be committed. Additionally, the isinstance(value, str) type guard for the conversation options key may silently miss non-string conversation identifiers (e.g., objects), causing reasoning items to be incorrectly omitted in that path.

✓ Test Coverage

The PR adds three new tests covering the main behavioral change (reasoning items omitted for stateless replay, kept when conversation_id is present, and an integration test for the full tool loop). These tests adequately cover the core happy paths. However, there are gaps: (1) no test for the _attribution override path where replays_local_storage=True causes reasoning items to be omitted even when request_uses_service_side_storage=True, (2) no tests for previous_response_id or conversation as the trigger key (only conversation_id is tested), and (3) the integration test doesn't assert that the fc_id from additional_properties is properly used (vs call_id) in the stateless path, which is the other half of the replays_local_storage logic.

✓ Design Approach

The PR correctly identifies that response-scoped reasoning item IDs (rs_*) are only valid within a live service-managed context and must be omitted when replaying conversation history stateless (store=False without a prior conversation reference). The request-level flag propagation through _prepare_options → _prepare_messages_for_openai → _prepare_message_for_openai → _prepare_content_for_openai is consistent with how the tool loop works: with store=False, no conversation_id is ever injected into mutable_options, so prepped_messages always contains the full history and the flag correctly gates reasoning item inclusion. One minor issue: the isinstance(value, str) check for the 'conversation' key is overly narrow — while ChatOptions only defines conversation_id (a string), raw caller dicts could in theory include a non-string conversation value (e.g. an OpenAI Conversation object), and the check would silently miss it. Additionally, committing .worktrees/ git-submodule entries is an accidental inclusion that should not be in this PR.

---

Automated review by chetantoshniwal's agents