Problem

Synchronous checkpointer and store provider singletons are lazily initialized without locks. Concurrent access can race and create multiple instances or observe partially initialized/reset state.

Impact

In multithreaded gateway/runtime contexts, callers can get inconsistent singleton instances or leak resources during concurrent initialization/reset.

Suggested Fix

Protect singleton initialization and reset with a lock and double-check pattern. Ensure reset and initialization cannot interleave.

Tests

• Concurrently call the sync checkpointer provider from multiple threads and verify a single instance is created.
• Repeat for the sync store provider.
• Verify reset during initialization is serialized.

References

• backend/packages/harness/deerflow/agents/checkpointer/provider.py:103
• backend/packages/harness/deerflow/runtime/store/provider.py:103