pr-shepherd

Use merge-shepherd.sh with the PR number and repo; it gates on green checks, MERGEABLE and CLEAN state, then enqueues via the GraphQL enqueuePullRequest mutation or squash-merges directly with --direct. It is stateless and idempotent, so re-running it after an interruption safely resumes the step.

The reliable path is the GraphQL enqueuePullRequest mutation with the PR node id, because some queue configurations reject every gh pr merge flag including --auto. If using the CLI fallback, pass --auto with no method flag and no --delete-branch, then confirm isInMergeQueue via GraphQL.

A CONFLICTING mergeable state silently blocks CI from firing, and an empty statusCheckRollup also occurs when the base branch has no required checks, such as an intermediate stacked PR. Check mergeable and mergeStateStatus first, and verify the rollup is non-empty before trusting a CLEAN state.

It detects stack membership with stack-probe.sh using both the REST stacks endpoint and GraphQL, and blocks merging any layer with open lower layers (exit 23). Stacks merge bottom-up one layer per run; a stack under a merge queue is refused and surfaced for a human decision.

Run probe-platform-health.sh, which compares workflow runs across prior heads of the branch and checks the status page scoped to check-relevant components. It returns healthy, degraded (attributed), degraded (unattributed), or unknown, and every verdict exits 0 so it never gates a merge decision.

It never auto-rebases CONFLICTING PRs, never decides what to merge, and the stateless writer cannot distinguish a queue eject from never-enqueued, so repeated ejects need manual rebase and regeneration. The platform probe also cannot emit healthy on branches with 100 or more head-triggered runs due to pagination.