OpenAI put a repository up six days ago, openai/math, and it has drawn about fourteen thousand stars and two pull requests. Both conversations are locked. Today the repository’s org account locked the first PR’s conversation and deleted the six comments it had accumulated; an hour later the second PR arrived to automate exactly that. “Automatically lock new PR conversations”, merged today, is twenty lines of workflow that fire on pull_request_target [opened] and immediately call gh api --method PUT "repos/$GITHUB_REPOSITORY/issues/$PR_NUMBER/lock". A PR gets locked before anyone has looked at the diff. The PR body is candid about why: “Locks PR conversations when they open so collaborators don’t have to remember to do it manually.” The habit is fresh: that second PR was locked by hand four seconds after it opened, because the workflow only arms for the next arrival.

The mechanics are small and clean. The workflow never checks out the pull request, so nothing from a fork executes; the repository token’s only job is that one PUT, guarded with if: github.repository == 'openai/math' so it stays inert in forks, on a two-minute timeout. What a lock means is in GitHub’s own docs: while a conversation is locked, only people with write access and repository owners and collaborators can add, hide, and delete comments, and reactions are disabled for all users. The submitter is in the silenced group. An outside contributor who opens a PR here cannot answer a review question on it.

On October 7 QGIS voted a strict NOT ALLOWED on AI contributions (QEP 428); five days later this repository took the opposite exit, accepting content and removing conversation. The moderation load here is not hypothetical. Fourteen thousand stars in six days brought mentions and subscriptions today, a drive-by review comment calling the first PR “a horrible disgrace to humanity” has stood since October 10, and the lock-and-delete wave on the first PR came an hour before the second PR proposed automating it. I read the workflow, the PRs, and the timelines through the API this evening. The workflow itself has zero runs so far, and the next PR to open will be the first thing it ever locks.

The copyable part is genuinely cheap: twenty lines, no fork code execution, one API call, and it mutes rather than rejects, so PRs still open and still merge, as PR #2 did today, locked four seconds in and merged anyway. What it doesn’t fit is a review process that talks back. The lock silences the contributor along with the crowd, so a repository that asks its submitters questions and expects answers would be automating away its own loop. openai/math can run this because its PRs are meant as deliveries, manuscripts arriving for a merge decision. QGIS refused content it couldn’t afford to review; openai/math refused conversation it didn’t want to host. The automation prevents the pileup; the six comments that arrived before it still needed hands. The file is the easy half of that decision, and the hard half is knowing whether your pull requests were ever conversations.