Python: [Bug]: Declarative workflows cannot save or resume checkpoints with FileCheckpointStorage #10074

Observed Behavior

A declarative workflow that uses FileCheckpointStorage (or any storage that decodes through the restricted checkpoint unpickler, such as CosmosCheckpointStorage) saves only its first checkpoint. Every later checkpoint fails with:

Failed to create checkpoint at iteration 1: Failed to decode pickled checkpoint data: Checkpoint deserialization blocked for type 'agent_framework_declarative._workflows._declarative_base:ActionComplete'. ...
Note that this does not fail the workflow run.

The run itself continues, so nothing raises. But the checkpoint taken while the workflow waits for input is never stored, so a paused workflow can't be resumed from storage. Restoring the latest checkpoint returns the iteration-0 one, and sending the response then fails with ValueError: Response provided for unknown request ID.

The cause is the same one #8258 fixed for orchestrations. The restricted unpickler trusts the agent_framework. and agent_framework_orchestrations. prefixes, but not agent_framework_declarative. Every declarative executor sends ActionComplete / ActionTrigger envelopes, and pauses checkpoint ExternalInputRequest, ToolApprovalRequest and the other request types from this package.

The only workaround today is to pass framework-internal module paths in allowed_checkpoint_types. test_workflow_run_kwargs_survive_file_checkpoint_restore in packages/declarative/tests/test_graph_executors.py does exactly that. samples/03-workflows/declarative/function_tools/main.py uses FileCheckpointStorage(TEMP_DIR) without that workaround and hits this. The WorkflowFactory docstring's checkpointing example ("With checkpointing for pause/resume support") does too, and it also passes path=, which FileCheckpointStorage doesn't accept.

Expected Behavior

Declarative workflows checkpoint and resume with FileCheckpointStorage without a user-maintained allowlist, the same way orchestrations do after #8258.

Steps to Reproduce

1. Build a declarative workflow with a RequestExternalInput action and WorkflowFactory(checkpoint_storage=FileCheckpointStorage(dir)).
2. Run it until it pauses.
3. Count the checkpoint files and restore the latest one into a fresh workflow.
4. Send the response for the pending request.

Minimal Reproduction

import asyncio, os, tempfile
from agent_framework import FileCheckpointStorage
from agent_framework.declarative import ExternalInputResponse, WorkflowFactory

YAML = """
name: approval-flow
actions:
  - kind: SetValue
    path: Local.greeting
    value: Hello
  - kind: RequestExternalInput
    id: ask
    prompt: Approve?
    variable: Local.answer
  - kind: SendActivity
    activity:
      text: '=Concatenate(Local.greeting, " - answer ", Local.answer)'
"""

async def main():
    with tempfile.TemporaryDirectory() as d:
        storage = FileCheckpointStorage(d)
        wf = WorkflowFactory(checkpoint_storage=storage).create_workflow_from_yaml(YAML)
        request = (await wf.run({})).get_request_info_events()[0]
        print("checkpoint files:", len([f for f in os.listdir(d) if f.endswith(".json")]))
        latest = await storage.get_latest(workflow_name=wf.name)
        restored = WorkflowFactory(checkpoint_storage=storage).create_workflow_from_yaml(YAML)
        await restored.run(checkpoint_id=latest.checkpoint_id)
        result = await restored.run(responses={request.request_id: ExternalInputResponse(user_input="yes")})
        print(result.get_outputs())

asyncio.run(main())

Error Messages and Stack Traces

WARNING Failed to create checkpoint at iteration 1: Failed to decode pickled checkpoint data: Checkpoint deserialization blocked for type 'agent_framework_declarative._workflows._declarative_base:ActionComplete'. ...
WARNING Failed to create checkpoint at iteration 2: ...
WARNING Failed to create checkpoint at iteration 3: ...
checkpoint files: 1
ValueError: Response provided for unknown request ID: cc60e75d-...

With the declarative types allowed, the same script writes 4 checkpoints and prints ['Hello - answer yes'].

Package Versions

agent-framework-declarative: 1.4.1; main at dc6617a

Python Version

Python 3.13

Operating System

macOS

Regression

Unknown

Additional Context

Proposed fix, mirroring #8258: register the framework-owned declarative types that cross a checkpoint boundary through the public register_checkpoint_type when agent_framework_declarative is imported. These are the executor message envelopes (ActionTrigger, ActionComplete, ConditionResult, LoopIterationResult, LoopControl); the request_info payloads and responses (ExternalInputRequest/Response, AgentExternalInputRequest/Response, ToolApprovalRequest/Response, MCPToolApprovalRequest); and ExternalLoopState, which is kept in workflow state while an agent's external input loop is paused. This keeps the core prefix allowlist unchanged. I have a PR ready and will link it here.