mcore-split-pr

The mcore-split-pr skill helps split a large Megatron-LM pull request into multiple smaller PRs, each touching the fewest possible CODEOWNERS reviewer groups. It solves a concrete review-burden problem: a PR spanning core, examples, tools, and training directories pulls in many reviewer groups, slowing merges, whereas a PR scoped to a single directory needs only that directory's reviewers.

The workflow has three phases. First it analyzes the PR by fetching its details and diff stat with the gh CLI, parsing .github/CODEOWNERS to map file patterns to owner groups, and counting the distinct reviewer groups the PR currently requires. Second it proposes a split that clusters files by their CODEOWNERS groups, keeps tests with the production code they validate, flags cross-PR dependencies (for example symbols renamed in one PR that another depends on), ensures each resulting PR is independently mergeable, and presents the plan as a table for user approval. Third, only after approval, it executes by creating branches from the correct base, applying file-scoped diffs with git apply, committing, and pushing to the user's fork. It always creates PRs as drafts, never pushes directly to upstream, confirms before any force-push, and credits the original author when the current user is not the PR author.

Target users are Megatron-LM contributors and maintainers dealing with sprawling PRs who want to reduce the number of reviewer groups per PR while keeping each split independently reviewable and mergeable. It is user-invocable with a PR URL or number as its argument.