Python: preserve Responses replay metadata across AG-UI continuations

[eavanvalkenburg]

Motivation & Context

AG-UI supports both stateless snapshot replay and provider-managed conversation continuation. These modes must not leak into each other: stateless runs should replay complete message history without restoring a provider response ID, while service-session runs should send only the incremental turn and restore private provider continuation state.

A cross-provider multi-turn matrix also found that real Responses output_text.logprobs were parsed into response-level metadata but were not retained on assistant content for direct replay. This loses provider-supplied data and prevents conformant replay when the receiving endpoint requires that field.

Description & Review Guide

• What are the major changes?

  • Persist and restore provider service-session IDs only when AG-UI service-session mode is active.
  • Reject use_service_session=True with an explicit store=False using a clear configuration error.
  • Preserve real assistant output_text.logprobs on parsed content and mirror them during direct Responses replay without fabricating empty values.
  • Add live two-turn AG-UI integration matrices for OpenAI Responses, OpenAI Chat Completions, FoundryChatClient, Foundry Prompt Agents, and Foundry Hosted Agents across stateless, conversation, and previous-response modes where applicable.
  • Keep the known Hosted Agent stateless replay incompatibility as a strict expected failure with a TODO to remove it when the upstream service accepts standard output-text replay without logprobs.

• What is the impact of these changes?

  • Stateless AG-UI history no longer accidentally combines full replay with a stored provider response ID.
  • Invalid disabled-storage service-session configurations fail before a provider request.
  • Provider-supplied logprobs survive direct assistant-message replay; messages without logprobs remain unchanged.
  • Multi-turn continuation behavior is continuously exercised across the supported OpenAI and Foundry client surfaces.

• What do you want reviewers to focus on?

  • The boundary between stateless snapshot state and provider-owned continuation state.
  • Selective logprobs preservation without synthetic defaults.
  • Matrix coverage and cleanup of temporary provider conversations and Prompt Agents.
  • The strict xfail scope for the known Hosted Agent stateless replay issue.

Related Issue

Fixes microsoft#7905

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 title prefix in sync automatically.

[eavanvalkenburg]

Closed because this draft was intended for the upstream microsoft/agent-framework repository.

[github-actions]

Python Test Coverage Report •

File	Stmts	Miss	Cover	Missing
packages/ag-ui/agent_framework_ag_ui
_agent.py	48	2	95%	82, 145
_agent_run.py	1477	147	90%	161, 221, 240, 287, 305, 321–322, 444, 497–498, 505, 614, 618, 620, 637, 664–665, 719, 748, 762, 766, 770, 773, 778, 783, 795, 802–808, 819, 833, 840, 850, 865, 868, 900, 935, 959, 972, 1021, 1039, 1041, 1067, 1069, 1071, 1074–1078, 1102, 1133, 1169, 1224–1225, 1267, 1281, 1326–1332, 1341, 1349, 1377, 1399, 1403, 1406, 1408, 1468–1469, 1475–1481, 1584, 1588–1589, 1592, 1597–1598, 1601, 1607–1608, 1612–1614, 1632–1634, 1659–1660, 1683–1684, 1697, 1739–1741, 1752–1754, 1756, 1770–1772, 1900, 1908, 1938, 1942, 1987, 2152–2153, 2239, 2348, 2353, 2425, 2427, 2434, 2450, 2455, 2512, 2517, 2521, 2570, 2620, 2695, 2841, 2879, 2887, 2889, 2892, 2898, 2953, 2956, 2966–2967, 2974, 3021
_run_common.py	662	30	95%	100, 175–176, 178, 180, 183, 189, 191, 205, 212, 236–237, 250, 269, 292–294, 305, 349, 361, 369, 372, 560, 581, 875–876, 1200–1201, 1267, 1270
_snapshots.py	83	4	95%	139, 222, 233–234
packages/anthropic/agent_framework_anthropic
_chat_client.py	542	34	93%	483, 486, 572, 709, 809, 843, 933, 972–973, 1051, 1053, 1083–1084, 1166–1168, 1172–1174, 1178–1181, 1216, 1332, 1342, 1394, 1542–1543, 1560, 1573, 1586, 1611–1612
packages/core/agent_framework
_agents.py	462	44	90%	600, 655, 1225, 1270, 1341–1345, 1444, 1474, 1511, 1606, 1634, 1647, 1696, 1698, 1707–1712, 1717, 1719, 1725–1726, 1733, 1735–1736, 1744–1745, 1748–1750, 1760–1765, 1769, 1774, 1776
packages/mistral/agent_framework_mistral
_chat_client.py	462	10	97%	285, 287–289, 296, 300, 304, 307–308, 310
packages/openai/agent_framework_openai
_chat_client.py	1409	114	91%	361, 374, 728–734, 743–746, 752–756, 764, 808–813, 817–820, 822–824, 831–833, 836, 893, 901, 924, 1141, 1200, 1202, 1204, 1206, 1272, 1286, 1366, 1376, 1381, 1424, 1536–1537, 1552, 1831, 1943, 1948–1949, 2032, 2042, 2069, 2075, 2085, 2091, 2096, 2102, 2107–2108, 2188, 2232, 2235–2238, 2252, 2262–2263, 2275, 2317, 2382, 2399, 2402, 2429–2431, 2470, 2487, 2490, 2552, 2559, 2596–2597, 2638, 2676–2677, 2695–2696, 2868–2869, 2887, 2978–2981, 3159, 3174, 3263–3265, 3275–3276, 3282, 3297, 3430–3431
TOTAL	48234	4481	90%

Python Unit Test Overview

Tests	Skipped	Failures	Errors	Time
9764	36 💤	0 ❌	0 🔥	2m 40s ⏱️