plan-pr-stack

Return dependency waves with evidence for every edge and visible gaps. Use this for a set of PRs; a single-PR readiness check belongs to the repository's delivery workflow.

1. Bind the target repository, selected PR numbers, observation time and requested finish. Read the repository's stacking and delivery rules. Planning authorizes reads of the identified targets, without granting publication or merge authority.
2. Read the snapshot contract before collecting input. Capture full head and base repository, branch and SHA identities. Resolve declared dependencies from source evidence, including PRs outside the selected set. Preserve missing values as null.
3. Run python3 <skill-dir>/scripts/plan_stack.py snapshot.json. Inspect unresolved external dependencies, unknown identity or file coverage, and cycle witnesses before using waves. Exit 1 means incomplete evidence or ordering; exit 2 means invalid input or a read failure.
4. Inspect shared contracts, migrations, generated artifacts and workflows beyond identical file paths. Add only evidenced orderings for required sequencing. File overlap alone has no direction. Explain the chosen order; never derive it from PR number. Re-run after changing the snapshot. Keep independent work independent.
5. Deliver the input snapshot and generated JSON, edge reasons, candidate waves, unresolved decisions and unknowns. A wave describes known dependency order, not safe parallel merges. Refresh identity and invalidated evidence after any head/base or dependency change. Hand off actual landing to the repository's live review, checks, protection and queue gates.

The bundled tool reads only its input file and writes JSON to stdout. It performs no Git or network operations. It does not infer semantic coupling, parse PR prose, judge reviews, validate test execution, restack branches, operate a merge queue or estimate savings.

Source context and intentional limits: sources.