Python: Python : checkpoint restore +pending request deserialization failure when hosting workflow as agent with ResponsesHostserver in foundry #7618

Activity

1. github-project-automation

   added this to Agent Frameworkon Aug 11, 2026
2. agent-framework-automation

   added

   pythonUsage: [Issues, PRs], Target: PythonUsage: [Issues, PRs], Target: Python

   triageUsage: [Issues], Target: All issues that still need to be triagedUsage: [Issues], Target: All issues that still need to be triaged

   on Aug 11, 2026
3. github-actions

   changed the title [-]Python : checkpoint restore +pending request deserialization failure when hosting workflow as agent with ResponsesHostserver in foundry[/-] [+]Python: Python : checkpoint restore +pending request deserialization failure when hosting workflow as agent with ResponsesHostserver in foundry[/+] on Aug 11, 2026
4. Evan Mattson (moonbox3)

   assigned

   Tao Chen (TaoChenOSU)

   on Aug 11, 2026
5. Evan Mattson (moonbox3)

   added

   hostingUsage: [Issues, PRs], Target: all hosting related solutionsUsage: [Issues, PRs], Target: all hosting related solutions

   and removed

   triageUsage: [Issues], Target: All issues that still need to be triagedUsage: [Issues], Target: All issues that still need to be triaged

   on Aug 11, 2026
6. Evan Mattson (moonbox3)

   added the

   Bug

   issue type on Aug 11, 2026
7. github-actions

   added

   reproducedUsage: [Issues], Target: all issues that can be reproduced by the triage workflowUsage: [Issues], Target: all issues that can be reproduced by the triage workflow

   on Aug 12, 2026
8. github-actions commented on Aug 12, 2026

   

   github-actionsBot

   on Aug 12, 2026 – with GitHub Actions

   Contributor

   More actions

   🤖 Automated triage reproduction notes (agent-authored — trust but verify)

   Agent analysis

   Repro: _RestrictedUnpickler.find_class() in packages/core/agent_framework/_workflows/_checkpoint_encoding.py (line 200) rejects types from agent_framework_orchestrations.* because _FRAMEWORK_MODULE_PREFIX on line 72 is "agent_framework." which doesn't match "agent_framework_orchestrations." (underscore vs dot). Triggered when FileCheckpointStorage (or any storage passing allowed_types != None) deserializes a checkpoint containing HandoffAgentUserRequest or MagenticPlanReviewRequest. Minimal repro: encode a HandoffAgentUserRequest via encode_checkpoint_value then decode with allowed_types=frozenset() — the restricted unpickler raises "deserialization blocked".

   • Failing test: python/packages/core/tests/workflow/test_checkpoint_orchestrations_deserialization.py
   • Files examined: python/packages/core/agent_framework/_workflows/_checkpoint_encoding.py, python/packages/core/agent_framework/_workflows/_checkpoint.py, python/packages/foundry_hosting/agent_framework_foundry_hosting/_state_store.py, python/packages/foundry_hosting/agent_framework_foundry_hosting/_responses.py, python/packages/orchestrations/agent_framework_orchestrations/_handoff.py
   • Tests run: python/packages/core/tests/workflow/test_checkpoint_orchestrations_deserialization.py
   • Reported version: agent-framework-core==1.13.0, agent-framework-foundry-hosting==1.0.0.b260730
   • Current version: agent-framework-core==1.13.0, agent-framework-foundry-hosting==1.0.0b260730
9. uuzzrm commented on Aug 12, 2026

   

   Ruiming Zhao (uuzzrm)

   on Aug 12, 2026

   Contributor

   More actions

   I can prepare a focused fix for this. The triage notes point to the checkpoint allowlist rejecting �gent_framework_orchestrations.* because it only matches the dotted �gent_framework. prefix. I’ll verify the repro on current main, add a regression test for the affected pending request types, and keep the change limited to checkpoint deserialization.
10. Eduard van Valkenburg (eavanvalkenburg)

    closed this as completedin #8723on Sep 28, 2026
11. github-project-automation

    moved this to Done in Agent Frameworkon Sep 28, 2026
12. ktz03 commented on Oct 18, 2026

    

    LI (ktz03)

    on Oct 18, 2026

    Contributor

    More actions

    A normal HandoffBuilder user-input pause still cannot be saved with default FileCheckpointStorage on source main ebcf4db7f9462ef6c377761e0b3159c4c5a73027, although the orchestration registrations from #8258 are present. This looks like additional evidence for this existing orchestration-checkpoint issue rather than a separate issue or a competing implementation.

    The workflow itself supplies a list[Message] response type for HandoffAgentUserRequest. At the pause boundary, file checkpoint save-time validation reports types:GenericAlias; the pending-request checkpoint is not saved. The latest successfully saved checkpoint contains no pending request, so attempting the normal reply against that checkpoint then raises No pending requests found in workflow context. This is a failed pause save, not corruption of a successfully saved pause checkpoint.

    The same ordinary handoff and reply with InMemoryCheckpointStorage saves one pending request and resumes successfully. The file case saves zero pending requests in its latest available checkpoint. Both use real Agent, function/middleware layers, public HandoffBuilder, and deterministic local chat responses, with the required per-service-call history persistence enabled; there is no application state, custom codec, extra allowed type or fault injection.

    Minimal reproduction

    Run with both Core and orchestrations resolving to the source revision above:

    import asyncio
    from pathlib import Path
    from tempfile import TemporaryDirectory
    
    from agent_framework import (
        Agent, BaseChatClient, ChatMiddlewareLayer, ChatResponse, Content,
        FileCheckpointStorage, FunctionInvocationLayer, InMemoryCheckpointStorage, Message,
    )
    from agent_framework_orchestrations import HandoffAgentUserRequest, HandoffBuilder
    
    
    class LocalClient(FunctionInvocationLayer, ChatMiddlewareLayer, BaseChatClient):
        def __init__(self, name):
            super().__init__()
            self.name = name
    
        async def _inner_get_response(self, *, messages, stream, options, **kwargs):
            await self._validate_options(options)
            assert not stream
            if any(m.text == "Tuesday" for m in messages):
                content = Content.from_text("Delivery confirmed.")
            elif self.name == "dispatch":
                content = Content.from_function_call(
                    call_id="delivery-handoff", name="handoff_to_calendar", arguments={}
                )
            else:
                content = Content.from_text("Which delivery date should I record?")
            return ChatResponse(messages=[Message(role="assistant", contents=[content])])
    
    
    def build(storage):
        agents = [
            Agent(client=LocalClient(name), name=name,
                  require_per_service_call_history_persistence=True)
            for name in ("dispatch", "calendar")
        ]
        return HandoffBuilder(
            name="delivery", participants=agents, checkpoint_storage=storage,
            termination_condition=lambda conversation: any(
                m.text == "Delivery confirmed." for m in conversation
            ),
        ).with_start_agent(agents[0]).build()
    
    
    async def check(storage):
        initial = await build(storage).run("Arrange delivery of the blue desk.")
        request = initial.get_request_info_events()[0]
        checkpoint = await storage.get_latest(workflow_name="delivery")
        print(type(storage).__name__, "saved pending requests:",
              len(checkpoint.pending_request_info_events))
        result = await build(storage).run(
            checkpoint_id=checkpoint.checkpoint_id,
            responses={request.request_id: HandoffAgentUserRequest.create_response("Tuesday")},
        )
        print([value.text for value in result.get_outputs() if hasattr(value, "text")])
    
    
    async def main():
        await check(InMemoryCheckpointStorage())
        with TemporaryDirectory() as directory:
            await check(FileCheckpointStorage(Path(directory)))
    
    
    asyncio.run(main())

    Observed output:

    InMemoryCheckpointStorage saved pending requests: 1
    ['Delivery confirmed.']
    FileCheckpointStorage saved pending requests: 0
    

    The file run logs this during checkpoint creation, then the reply raises the pending-request error:

    Failed to create checkpoint at iteration 2: Failed to decode pickled checkpoint data:
    Checkpoint deserialization blocked for type 'types:GenericAlias'.
    ...
    RuntimeError: No pending requests found in workflow context.
    

    Tested on Windows / Python 3.14.3, using source Core 1.22.0 and orchestrations 1.3.2 with cached dependencies (installed Core metadata 1.19.0). A second independent local business example produced the same memory/file difference. This does not cover live models, Foundry hosting, AG-UI, streaming, other Python versions or a full suite.

    Related scope: #7618 / #7789 and the already merged #8258 cover built-in orchestration checkpoint types; #7831 owns the broader recovery/fallback behavior. #9440 needs an explicitly allowed application ZoneInfo value, whereas this example contains only framework-generated response types and ordinary text. Could maintainers confirm the direction for this remaining built-in Handoff case within the existing route? No implementation or allowlist change is included.

    AI assistance: analysis, reproduction, and writing.