fix(graph): prevent reset_executor_state from corrupting MultiAgentBase state

[giulio-leone]

Summary

Fixes #1775 — GraphNode.reset_executor_state() corrupts MultiAgentBase executor state by overwriting GraphState with AgentState.

Root Cause

reset_executor_state() checks hasattr(self.executor, 'state') without verifying the state type:

# Before (broken)
if hasattr(self.executor, 'state'):
    self.executor.state = AgentState(self._initial_state.get())  # Overwrites GraphState!

__post_init__ already had the correct guard (hasattr(self.executor.state, 'get')), but reset_executor_state() was missing it.

Impact

Two call sites affected:

Call site	Trigger
_execute_node	reset_on_revisit=True with nested graph in a cycle
deserialize_state	Always — iterates all nodes when restoring a completed run

Fix

Mirror the __post_init__ guard:

# After (fixed)
if hasattr(self.executor, 'state') and hasattr(self.executor.state, 'get'):
    self.executor.state = AgentState(self._initial_state.get())

Test

Added test_reset_executor_state_preserves_graph_state_for_nested_graph — creates a GraphNode with a nested Graph executor and verifies reset_executor_state() preserves the GraphState type.

All 49 graph tests pass.

⚠️ This reopens #1896 which was accidentally closed due to fork deletion.

[mkmeral]

/strands review

[github-actions]

Assessment: Approve

Clean, well-scoped bugfix that correctly mirrors the existing __post_init__ guard into reset_executor_state(). The root cause analysis in the PR description is excellent, and the regression test covers the exact scenario that was broken.

Review Details

• Correctness: The fix is sound — GraphState (a dataclass) lacks a get() method while AgentState (JSONSerializableDict) has one, so the hasattr(self.executor.state, "get") guard correctly prevents the type corruption.
• Robustness: The duck-typing guard (hasattr(..., "get")) is fragile by nature — if GraphState ever gains a get() method, this breaks silently. An isinstance check or executor-type check would be more future-proof. That said, this is a pre-existing pattern, and this PR correctly keeps the two call sites in sync. See inline comment for a suggestion to improve both sites together in a follow-up.
• Testing: The new test is focused and verifies the exact regression scenario.

Nice fix — minimal and well-targeted. 👍

[github-actions]

Suggestion (non-blocking): The hasattr(self.executor.state, "get") guard relies on duck typing to distinguish AgentState from GraphState. This works today because GraphState is a dataclass without a get() method, but it would silently break if GraphState ever gains one.

A more explicit guard — such as isinstance(self.executor.state, AgentState) or not isinstance(self.executor, MultiAgentBase) — would make the intent clearer and be more resilient to future changes.

Since __post_init__ uses the same pattern, this could be addressed as a follow-up to improve both sites together.

[mkmeral]

+1 on this actually, can we update the condition? This opens the door for a possible bug later on