How does human review fit into AI-assisted PR triage?

[msegec]

This morning I noticed that 36 unmerged PRs had been closed on 28 August after comments headed "GPT-5.6 Sol responding on behalf of Theo." Thirty-three were closed within 1 hour and 45 minutes~ Five were mine.

This led me to reread CONTRIBUTING.md. I understand that T3 Code is not actively accepting contributions, and that opening a PR does not oblige maintainers to review or merge it.

What I would like to understand is where human review fits into this process. Does a maintainer read each PR and approve the decision before Sol or another delegated model posts the closure, or are these models authorised to make the initial keep-or-close decision, with maintainers becoming involved only if a contributor challenges it?

I understand that maintaining the current backlog is a real workload. At the same time, when a PR represents substantial voluntary work, a bulk AI-assisted closure can feel dismissive if the contributor cannot tell whether a maintainer read it.

I maintain my own fork against nightly. Understanding the process would help me decide which changes are useful to offer upstream and which are better kept in my fork.