[Eduard van Valkenburg (eavanvalkenburg)]

Motivation & Context

Microsoft Foundry background responses with local tools already complete through service-side continuation when store=True. With store=False, a completed background response can instead return a provider-native reasoning item without encrypted_content. Agent Framework executed the local tool once, then rejected the follow-up before transport because stateless replay only recognized encrypted reasoning.

Foundry accepts the original completed reasoning item as the replay representation. This change preserves and replays that item so background local-tool loops work in both storage modes without duplicating tool execution.

Description & Review Guide

• What are the major changes?
  • Derive stateless replay validation from the reasoning items the provider serializer can actually reconstruct.
  • Add a private provider hook for non-encrypted reasoning replay.
  • Preserve an allowlisted completed Foundry reasoning payload for non-streaming and streaming responses.
  • Add unit coverage for successful replay, incomplete-payload rejection, and streaming .done precedence, plus a live Foundry integration test for store=True and store=False.
• What is the impact of these changes?
  • Foundry background local-tool loops with store=False now complete successfully and execute each local tool once.
  • Existing service-side continuation and generic OpenAI encrypted-reasoning replay remain unchanged.
• What do you want reviewers to focus on?
  • The replay payload validation and the requirement that streaming replay uses the completed .done item rather than the earlier .added skeleton.

Related Issue

Fixes #7538

No other open PR targets this issue.

Contribution Checklist

• The code builds clean without any errors or warnings
• All unit tests pass, and I have added new tests where possible
• The PR follows the Contribution Guidelines
• This PR is linked to an issue and there is no other open PR for this issue (see Related Issue above).
• This is not a breaking change. If it is a breaking change, add the breaking change label (or add "[BREAKING]" to the title prefix, before or after any language prefix) — a workflow keeps the label and the title prefix in sync automatically.

[Copilot]

Copilot review overview

🟡 Changes recommended

Replay validation can accept malformed payloads or associate reused reasoning IDs with the wrong tool group.

Get a fresh assessment by requesting another Copilot review.

Review effort: Balanced
Findings: 1 · 1 · 1

Open (3)

• Global reasoning IDs cause lossy cross-group replay serialization · New
• Stored reasoning payload bypasses structural validation · New
• Use provider-neutral error for missing reasoning replay data · New

What changed in this PR

Enables stateless replay of Foundry background responses containing provider-native reasoning items.

Changes:

• Adds a provider-specific reasoning replay hook.
• Preserves completed Foundry reasoning payloads in streaming and non-streaming paths.
• Adds replay validation and regression/integration tests.

File	Description
python/​packages/​openai/​agent_framework_openai/​_chat_client.py	Generalizes stateless reasoning replay.
python/​packages/​foundry/​agent_framework_foundry/​_chat_client.py	Captures and restores Foundry reasoning payloads.
python/​packages/​foundry/​tests/​foundry/​test_foundry_chat_client.py	Tests replay, streaming precedence, and background tool loops.

---

💡 Add a code-review agent skill for context-aware, tailored reviews. Learn more in the docs.

[github-actions]

MAF Automated Review — Iteration 1

Result: Findings reported
Scope: full PR (1 commit(s)): 87352e37c193
Model: gpt-5.6-sol-fast

Overview

The PR preserves an allowlisted completed Foundry reasoning payload and uses the streaming .done item, while generic OpenAI replay remains fail-closed and focused tests cover successful and rejected replay. However, deriving validation from a global set of replayable reasoning IDs weakens the documented group-atomic invariant: a payload from one logical tool group can make another group with the same ID pass validation and serialize incorrectly.

Reviewed the supplied pull-request change set across correctness, security/reliability, architecture, and failure behavior.
1 verified finding remained after source verification (1 medium) across 1 file. Details are attached to the affected lines below.

Affected areas: python/packages/openai/agent_framework_openai/_chat_client.py