agentsclimarketplace

Split pr

Skill UMEP-dev/SUEWS/.claude/skills/split-pr

Use whenever an oversized or bundled SUEWS pull request (or a fat local branch) needs to be carved into a stacked series of smaller, single-concern PRs that each build and pass tests on their own. Trigger for requests like "split this PR", "this PR is too big, break it up", "carve this branch into stacked PRs", "separate the refactor from the fix", "split off the formatting/generated changes", or when audit-pr's Size Gate recommends a split. This skill assesses the diff, proposes a split plan, creates per-child worktrees/branches, carves by concern or by commit while preserving authorship, verifies each child independently green, checks completeness (every change lands in exactly one child, nothing dropped), opens the stacked PRs for approval, and hands off to queue-pr. It does not merge and does not rewrite the original branch without explicit approval.From its SKILL.md

Install
npx -y skills add UMEP-dev/SUEWS --skill split-pr

Assembled from the repository path, not quoted from the project. Check it against their README if it does not work.

SKILL.md

12.7 KB, as published. Nobody here has run it

Split PR

Carve one oversized or bundled SUEWS pull request (or a fat local branch) into a stacked series of smaller, single-concern PRs. Each child must build and pass tests on its own, and the union of the children must equal the original diff with nothing dropped or duplicated.

The terminal state is a verified stack of PR-ready branches, not merged. This skill does not review (audit-pr) and does not merge (queue-pr / merge queue). It preserves the original branch and PR untouched until the stack is verified.

Triggers

  • Use for: "split this PR", "this PR is too big", "carve this branch into stacked PRs", "separate the refactor from the fix", "split off the formatting/generated changes", or when audit-pr's Size Gate recommends a split.
  • Size criteria live in .claude/rules/work-sizing.md. If the user only wants a judgement on whether to split, use audit-pr (Size Gate) instead.

Safety Rules

  • Never rewrite or force-push the original branch/PR without explicit approval for that exact branch. Carve into new branches; leave the original intact until the stack is verified and the user approves retiring it.
  • Pushing the new branches and opening the child PRs requires approval. Local carving, worktree creation, and verification are allowed when clearly scoped.
  • Every change in the original must land in exactly one child (see Completeness check). Never silently drop or duplicate a hunk.
  • Each child must leave the tree green: it builds and passes make test-smoke independently (a stacked child is verified on top of its parent).
  • Do not auto-resolve semantic, scientific, or schema conflicts that arise while carving. Stop and report.
  • Preserve authorship via the by-commit/cherry-pick route, which keeps the original author. The by-path route creates fresh commits authored by you; those carry the AI-collaboration trailer for the assistant that made them (Codex or Claude), never the wrong one. If single-author attribution must be retained for a by-path child, prefer cherry-pick or pass --author=.
  • Never reuse or delete a pre-existing split/<pr>-* branch or worktree without first reporting it and getting explicit approval; never delete the original branch/PR.
  • Privacy on every outward write: never let internal tracker IDs (e.g. Linear PER-XXX), absolute local/worktree paths (/Users/..., ../split-<pr>/...), internal branch names, or private host names reach a child PR body, commit message, the stack/concern descriptions, or the retire recommendation. These surfaces are public; scrub model-authored text (operationalises the privacy invariant in .claude/rules/autonomous-workflow.md).
  • Stop if the diff cannot be cleanly separated into independent concerns, or if splitting would require a broad redesign.

Scripts

No bundled scripts. Use git worktree, gh, and the queue-pr scripts for the downstream coordination of the resulting stack.

Workflow

1. Snapshot and assess

  1. Identify the source: a PR number or a local branch. Fetch and inspect:
    git fetch origin
    gh pr view <pr> --json number,title,headRefName,headRefOid,baseRefName,author,files,url
    gh pr diff <pr> --name-only
    git diff --name-status origin/<base> <source-head>   # D = deletion, R = rename: special carve in step 3
    
  2. Group the diff into separable concerns by file/subsystem and by commit (git log --oneline origin/<base>..<head>). Watch for a mix of refactor + behaviour change + formatting, bundled unrelated fixes, or generated/mechanical changes that should be carved out first.
  3. Classify each child's base using queue-pr's shape model, reduced to the two shapes that apply when carving a fresh stack: a child is independent (off <base>, no logical dependency) or stacked (off its parent child). Shared build/schema/docs edits do not get their own shape term here; they become a dedicated high-risk child (step 6 and "SUEWS high-risk files"), which is this skill's form of queue-pr's integration PR. queue-pr's serial shape has no analogue: a carved child's dependency is fixed by how it is carved.
  4. Propose the split plan: which files/commits go in which child, the order, and independent-vs-stacked. Present it for approval before any surgery.

2. Set up isolation

  1. Keep the original branch/PR untouched. Do not check out PR branches in an active worktree.
  2. Pre-flight for leftovers from a prior interrupted run (never silent, never auto-delete):
    git worktree list
    git branch --list 'split/<pr>-*'
    
    If any split/<pr>-* branch or ../split-<pr>/* worktree already exists, STOP and report it. Then take one human-gated path: (a) resume, if the leftovers match the approved plan; or (b) clean and re-carve, only on explicit approval:
    git worktree remove --force ../split-<pr>/<x>
    git worktree prune
    git branch -D split/<pr>-<x>          # branch persists after path removal
    
  3. Create a coordinator area and per-child worktrees. Fresh carve (new branch off the agreed base):
    mkdir -p ../split-<pr>
    git worktree add -b split/<pr>-a ../split-<pr>/a origin/<base>
    
    For a stacked child, use -b split/<pr>-<x> off its parent child instead of the base. The -b form fails with fatal: a branch named '...' already exists (exit 255) if the branch is still present, so it must not be used on the resume path. Resume instead re-attaches a worktree to an EXISTING leftover branch (no -b):
    git worktree add ../split-<pr>/a split/<pr>-a
    
    After re-attaching, confirm the branch tip matches the approved plan before continuing. Tear down the per-child worktrees after the stack is pushed (git worktree remove ../split-<pr>/<x>), leaving the branches intact.

3. Carve each child

  • By path (files group cleanly): in the child worktree, bring in only that child's paths from the source head:
    git checkout <source-head> -- <path> [<path> ...]
    git commit  # focused subject. NOTE: authored by you, not the original author --
                # carving by path cannot preserve per-hunk authorship. Use the
                # by-commit/cherry-pick route below to retain attribution, or pass
                # --author="Name <email>". This fresh commit carries the
                # AI-collaboration trailer for the assistant that made it.
    
    Caveat -- by-path carve cannot capture deletions or renames. git checkout <source-head> -- <path> only copies files that exist at <source-head>, so a path DELETED between base and head fails with error: pathspec '<path>' did not match any file(s) known to git and the deletion is never staged. Carve a deletion with git rm <path> in the child; for a rename, carve both sides (git rm <old-path> and git checkout <source-head> -- <new-path>). The D/R statuses from step 1's git diff --name-status tell you which paths need this.
  • By commit (the source has clean per-concern commits): cherry-pick the commit group onto the child branch (cherry-pick preserves the original author).
  • Carve out mechanical/generated changes (formatting sweeps, regenerated artefacts, large fixture updates) into their own child PR first so the semantic children stay small and reviewable.

4. Verify each child independently

Run the green-tree check on each child on its own (a stacked child on top of its parent): build if needed, then make test-smoke, plus any targeted check the child's files warrant (pytest, ruff check, fprettify --diff). If a child is not green in isolation, the boundary is wrong; re-carve rather than papering over it.

5. Completeness check (critical gate)

Confirm the stack reproduces the original exactly, with nothing dropped or duplicated, using a single shape-general recombine that covers stacked, independent, and mixed shapes. Identify the LEAF branches that jointly cover every concern: every independent child, plus the TIP of each stacked sub-series (never a non-tip stacked child, whose commits its tip already contains). In a throwaway worktree off the base, octopus-merge exactly that leaf set, diff against the source head, then tear the worktree down:

git worktree add ../split-<pr>/_recombine origin/<base>
git -C ../split-<pr>/_recombine merge --no-edit <leaf-1> <leaf-2> ...   # each independent child + each stacked sub-series TIP
git -C ../split-<pr>/_recombine diff <source-head> HEAD                 # must be empty
git worktree remove --force ../split-<pr>/_recombine

Fast path: for a PURELY stacked stack with no independent siblings, the single tip child already contains everything, so git diff <source-head> <tip-child-branch> suffices. Do not use the fast path in a mixed shape -- it false-flags the independent child's hunks as a phantom diff.

If any recombine merge reports CONFLICT (non-zero exit, or git status shows UU), the children share an overlapping hunk -- the boundary is wrong. Run git merge --abort and re-carve; do NOT hand-resolve (that masks a bad split), and do not trust the subsequent git diff, which would run against a half-merged state. If the merge succeeds cleanly but the diff is non-empty, a change was lost or duplicated. Resolve before proceeding; never report a split complete while this diff is non-empty.

6. Open the stack

After approval, push the child branches and open the PRs:

  • Independent children: base each on <base>.
  • Stacked children: base each PR on its parent child branch; note the stacking in the PR body so a reviewer reads them in order.
  • Link every child to the original (Refs #<original>), and list the stack in each child body. Schema/sample-config/build/CHANGELOG changes should be their own child (schema bumps already warrant a dedicated change per .claude/rules/python/schema-versioning.md; see also the "one PR per logical group" rule in .claude/rules/naming-convention.md).

7. Retire the original and hand off

  • Recommend closing or converting the original PR to a tracking/umbrella issue once the stack is up, preserving its discussion. This is human-gated; do not auto-close.
  • Hand the stack to queue-pr for ordering and merge-queue entry. Handing off does not transfer merge authority: enqueue/merge stays a separate, explicitly-approved step under queue-pr's own human gate.

SUEWS high-risk files

Carve these into their own child PR rather than scattering them across children: schema/public API and sample configuration, meson.build / src/supy/meson.build, .github/workflows/*, CHANGELOG.md, dependency/lock files, generated docs indexes, and shared test fixtures/baselines. See the full list in queue-pr.

Decision Rules

  • Prefer carving by path when files group cleanly; carve by commit when the history is already concern-clean.
  • Independent children become parallel PRs; dependent children become a stack. Do not hide a dependency by basing a dependent child directly on the base.
  • A child that cannot be made green in isolation signals a wrong boundary, not a reason to bundle.
  • Stop and report on any semantic/scientific/schema conflict; do not invent a resolution to make Git clean.

Output Format

End with:

[split-pr] stack plan
Source: #<pr> <title> (or branch)
Shape: independent | stacked | mixed
Concerns: <N identified in step 1> -> Children: <M created>   # must match or explain
Children:
  1. split/<pr>-a - <concern> - checks: test-smoke pass/blocked - PR <url|not pushed>
  2. split/<pr>-b - <concern> - checks: ... - PR <url|not pushed>
Completeness: empty-diff confirmed | MISMATCH: <what>
Original: untouched | retire recommended (human) | closed (approved)
Approvals: <none | push+open approved | retire approved>
Remaining: <none | explicit blocker>
Next: hand the stack to queue-pr for ordering and merge.

If the workflow stops early, use the same format and set Remaining to the specific blocker.

Keep looking

Skills are one crate of 614,194. Ordering is by how many stacks a row turns up in, so the top of any crate is what has actually been picked rather than what has the most stars.