[Slai.life (huong8373tt-beep)]

Summary

Fixes #49475.

Checkpoint writes from a ResponseEventStream(response_id=...) now use the same response-snapshot normalization as terminal persistence. This stamps the resolved response identity before update_response, including:

• response_id
• agent reference
• model
• agent session
• conversation
• persisted background mode

The change is limited to checkpoint persistence and does not expose checkpoint events on the wire. Snapshot preparation remains inside the existing failure guard, so an extension value that cannot be deep-copied (for example, a generator) safely skips the checkpoint without breaking the live response.

Validation

• tests/unit/test_checkpoint.py — 9 passed.
• tests/unit/test_event_copy_reuse.py — 27 passed.
• Combined checkpoint/event-copy regression — 36 passed.
• MyPy passed for the 70 affected source files after the explicit provider-boundary cast.
• python -m compileall -q sdk/agentserver/azure-ai-agentserver-responses/azure tests — passed.
• python -m build --wheel — passed.
• git diff --check — passed.
• Pytest used an E-drive --basetemp because the Windows host had a stale C-drive temporary link with insufficient access; no production dependency was changed.

The targeted regression covers an unseeded checkpoint and asserts that the independently persisted payload contains the resolved identity and background mode. It also covers generator-valued extensions and verifies that checkpoint preparation failures are swallowed without affecting the live response.

Disclosure and authorization

Slai Agent Runtime performed the investigation, implementation, and test execution. The account owner authorized this contribution and is responsible for its content.

[github-actions]

Thank you for your contribution Slai.life (@huong8373tt-beep)! We will review the pull request and get back to you soon.

[azure-pipelines]

Azure Pipelines:
Successfully started running 1 pipeline(s).
10 pipeline(s) were filtered out due to trigger conditions.
There may be pipelines that require an authorized user to comment /azp run to run.

[Copilot]

🟡 Changes recommended

The new deepcopy can raise an uncaught TypeError for generator-valued extensions, causing live responses to fail instead of safely skipping the checkpoint.

1 open issue

• Handle generator-valued extensions during response checkpointing · New

What changed in this PR

This PR addresses #49475 by applying terminal-response identity normalization to checkpoint persistence, making checkpoints independently storable without exposing checkpoint events on the wire.

Changes:

• Forward resolved agent, model, session, and conversation identity to checkpoint persistence.
• Normalize checkpoint snapshots and persist the background mode flag.
• Add a regression test asserting the persisted checkpoint payload includes identity fields.

File	Description
sdk/​agentserver/​azure-ai-agentserver-responses/​tests/​unit/​test_checkpoint.py	Adds checkpoint identity and background-mode assertions.
sdk/​agentserver/​azure-ai-agentserver-responses/​azure/​ai/​agentserver/​responses/​hosting/​_orchestrator.py	Normalizes checkpoint payloads and forwards resolved identity from both checkpoint call sites.

_{Review effort: Balanced}

[Slai.life (huong8373tt-beep)]

Addressed the Copilot review finding in e7809a0974.

Checkpoint snapshot preparation is now inside the existing serialization guard. If a response extension contains a generator or another value that cannot be deep-copied, checkpointing returns the prior snapshot without failing the live response. Added test_checkpoint_preparation_failure_is_swallowed to cover this path.

Validation:

• tests/unit/test_checkpoint.py: 9 passed
• tests/unit/test_event_copy_reuse.py: 27 passed
• combined: 36 passed
• python -m compileall -q azure tests: passed
• wheel build: passed
• git diff --check: passed

The test environment required the repository's current httpx2 dependency for Starlette's test client; no production dependency was changed.

[Slai.life (huong8373tt-beep)]

Addressed the Azure Pipelines MyPy failure reported by the analyzer: checkpoint_response is normalized as a dictionary at runtime, but the response provider protocol expects the generated ResponseObject TypedDict. The persistence call now uses an explicit cast at the provider boundary. Local validation: MyPy passed for 70 source files; checkpoint and event-copy tests passed (36 tests). Commit: 660a5759d1.

[Slai.life (huong8373tt-beep)]

Revalidated the checkpoint change at the current head: test_checkpoint.py plus test_event_copy_reuse.py pass (36 passed), and python -m compileall -q sdk/agentserver/azure-ai-agentserver-responses/azure passes. The checkpoint preparation path retains the generator-safe failure guard; the current type-boundary update is included in the refreshed head. Azure Pipelines has been restarted.

[Copilot]

🟢 Approval recommended

No active review issues were identified for the checkpoint persistence normalization change.

No open issues 🎉

1 resolved since last review

• Handle generator-valued extensions during response checkpointing

_{Review effort: Balanced}

[Slai.life (huong8373tt-beep)]

The current Azure pipeline run completed all build, documentation, MyPy, lint, and test jobs successfully. The only failure is the repository dependency analyzer, which reports that azure-ai-agentserver-core >=2.3.0, <3.0.0 has no resolvable combination. This PR does not change dependency declarations. Could a maintainer confirm whether this analyzer failure is a repository/package-index baseline issue or whether an additional metadata update is expected for this change?