[h05632hs-create]

Summary

Fix event-loop lifetime management in OpenEvolve's process-parallel workers.

Repeated asyncio.run() calls can close event loops while cached asynchronous clients retain resources bound to those loops. Subsequent worker operations may then fail with RuntimeError: Event loop is closed.

Changes

• Introduce a persistent, worker-owned event loop for asynchronous generation and evaluation.
• Ensure orderly cleanup, including pending-task cancellation, asynchronous generator shutdown, and executor shutdown.
• Add five regression tests covering repeated calls, loop sharing, exception propagation, cleanup, and task cancellation.

Validation

Tested against upstream commit bcf63da435ed6e89b751bddaf01dbfc0744a5320.

• Original upstream: 912 tests passed.
• Patched upstream: 917 tests passed.
• New lifecycle tests: 5/5 passed.
• Both full suites used the same isolated Docker dependency environment and were executed without external network access.

Earlier isolated investigations reproduced the event-loop lifecycle failure and supported the repair.

Scope and limitations

The change is limited to openevolve/process_parallel.py and tests/test_worker_loop_lifecycle.py.

The full local unit suite passed, but compatibility with every provider and deployment configuration remains unverified. No performance improvement is claimed.

Feedback on the worker lifecycle and cleanup approach is welcome.

[CLAassistant]

All committers have signed the CLA.