What I see both in Claude Code and Codex, is that the agent sets automations to wake it, fairly often. In the example below, I don't think I prompted anything 3 hours ago. Instead, it set some schedule to look periodically for replies/reviews to the PR, and each woke the agent, and the agent reacted

Yeah, and GPT tries to do it in openhands, but it fails and the conversation just stops. We might want to open an issue for that