I do RE mostly, and they do lock up for me. But what works for me is: warm up context with non-RE things with American models > switch to GLM 5.3 with actual request, let it fill context with some RE work > switch to American > switch back to GLM 5.3 and keep switching back and forth if anything gets flagged.

Solves the guardrail problem - check Feeds China data to US - check Feeds US data to China - check

But I think OpenAI has caught up with this, and it's why they're pushing so many things server-side and encrypt it (subagent messages, compaction, new context history thing).