Repository navigation
Description
Observed Behavior
FoundryCheckpointStore.get_latest() raises TypeError when the same workflow contains both a naive timestamp
and a timezone-aware timestamp, even though both checkpoints can be saved and loaded successfully.
For example, saving 2026-01-01T00:00:00 and 2026-01-02T00:00:00+00:00 makes latest-checkpoint selection fail
instead of returning the second checkpoint.
Default checkpoint timestamps are timezone-aware. This requires supplied or existing naive timestamps; it does not
show that ordinary UTC-only workflows or loading a known checkpoint ID fail.
Expected Behavior
Return the latest checkpoint, treating naive timestamps as UTC for comparison. This follows the behavior added to
the Core memory/file stores in #9944.
Keep the saved timestamp strings unchanged and retain lineage-based tie-breaking when two representations identify
the same instant.
Steps to Reproduce
- Install the package versions below, or use unmodified source at
7a166e79e7a0025f3bc3a752e328f341cbaddf25. - Run the script below. It replaces only the external SDK storage boundary; encoding, saving, listing, decoding,
and checkpoint selection use the real implementation. - Observe the
TypeErrorfromget_latest().
No credentials, Azure resources, model, or hosted service is required.
Minimal Reproduction
# Copyright (c) Microsoft. All rights reserved.
import asyncio
from types import SimpleNamespace
from unittest.mock import AsyncMock, MagicMock, patch
from agent_framework import WorkflowCheckpoint
from agent_framework_foundry_hosting import FoundryCheckpointStore
from azure.ai.agentserver.core import FoundryAgentRequestContext
async def main() -> None:
saved = {}
backend = MagicMock()
backend.__aenter__ = AsyncMock(return_value=backend)
backend.__aexit__ = AsyncMock(return_value=None)
backend.set_item = AsyncMock(
side_effect=lambda key, value, **kwargs: saved.__setitem__(key, value)
)
backend.get_item = AsyncMock(
side_effect=lambda key, **kwargs: SimpleNamespace(value=saved[key])
)
backend.list_keys = AsyncMock(
return_value=SimpleNamespace(
keys=[SimpleNamespace(key="older"), SimpleNamespace(key="newer")],
has_more=False,
last_id=None,
)
)
context = FoundryAgentRequestContext(
call_id="repro-call", user_id="repro-user", session_id="repro-session"
)
storage = FoundryCheckpointStore("repro-context", context)
with patch(
"azure.ai.agentserver.core.storage.FoundryStateStore.get_or_create",
new=AsyncMock(return_value=backend),
):
for checkpoint_id, timestamp in [
("older", "2026-01-01T00:00:00"),
("newer", "2026-01-02T00:00:00+00:00"),
]:
await storage.save(
WorkflowCheckpoint(
workflow_name="workflow",
graph_signature_hash="graph",
checkpoint_id=checkpoint_id,
timestamp=timestamp,
)
)
latest = await storage.get_latest(workflow_name="workflow")
print("latest:", latest.checkpoint_id if latest else None)
assert latest is not None and latest.checkpoint_id == "newer"
asyncio.run(main())Error Messages and Stack Traces
TypeError: can't compare offset-naive and offset-aware datetimes
The comparison fails in _select_latest_checkpoint() at
max(datetime.fromisoformat(cp.timestamp) for cp in checkpoints).
Package Versions
Reproduced with the published wheels:
agent-framework-foundry-hosting==1.0.0b261112
agent-framework-core==1.25.0
azure-ai-agentserver-core==2.1.0
azure-ai-agentserver-responses==2.2.0b1
azure-ai-agentserver-invocations==1.1.0
Also reproduced against unmodified source at c7e50ccd90370416afc4ab466394b06395f7dccd with editable Core 1.26.0.
The Foundry selector remains unchanged at 7a166e79e7a0025f3bc3a752e328f341cbaddf25.
Python Version
Python 3.11.14 for the released-wheel check, and 3.12.12 for the source-checkout reproduction.
Operating System
macOS, Apple Silicon (arm64).
Regression
Unknown.
Additional Context
This is a Foundry-specific follow-up to #9941 /
#9944, which covered the Core memory/file implementations.
The Foundry selector
still compares the parsed datetimes without supplying a timezone for naive values.
I have prepared a local candidate and six regression cases covering naive-older, naive-newer, and same-instant
parent/child checkpoints in both save orders. I would like to contribute this Foundry-specific fix; would applying
the same comparison policy here be welcome?
AI assistance: Codex assisted with duplicate research, reproduction, diagnosis, the local candidate, regression
tests, and this report. The reproduction and tests were executed locally.
Acknowledgements
- I searched existing issues and did not find a duplicate.
- I personally verified this behavior and the reproduction details are authentic.
- I will wait for explicit maintainer agreement before starting implementation of a non-trivial change.
Metadata
Metadata
Assignees
Labels
Type
Projects
- StatusShow more project fieldsNo status
Activity
github-actions commentedon Nov 13, 2026
Reproduced on
739a387(Python): a focused test calling_select_latest_checkpointwith one naive (2026-01-01T00:00:00) and one aware (2026-01-02T00:00:00+00:00) checkpoint fails withTypeError: can't compare offset-naive and offset-aware datetimesat_state_store.py:47. I tested the selector directly, not through your full mocked-storage script.Support boundary: unclear. I found no maintained doc or decision record covering naive timestamps, so a maintainer needs to decide. #9944 is the related Core precedent.
Evidence: https://github.com/microsoft/agent-framework/actions/runs/43411797544. Routed to
foundry_hosting; maintainers will decide scope and priority.