Openclaw allows you to effectively “shell out” to another harness for your model calls, while still using Pi as your main agentic harness. This is the claude -p workflow. Tools and skills are injected into Claude and they hack session persistence into it as well.

They also absolutely blocked OpenClaw system prompts from this path in the prior weeks, based purely on keyword detection. Seems they’ve undone that now.