Stop letting monster PRs die in review purgatory.
A Claude Code plugin that splits large PRs into chains of small, reviewable PRs. It analyzes your diff semantically — using tree-sitter for per-function granularity — builds a dependency DAG, verifies the split is lossless, and creates the PRs with proper base branches and merge order.
/plugin marketplace add emilhe/split-pr
/plugin install split-pr@emilhe-split-pr/split-pr # split current branch vs main
/split-pr --base develop # different base branch
/split-pr --name auth-refactor # custom prefix for PR titles
/split-pr --threshold 600 # larger PRs allowed (default: 400 lines)
/split-pr --max-files 15 # max files per PR (default: 10)
/split-pr --bulk "vendor/" # skip AST analysis for vendored paths
/split-pr --auto # skip interactive review
/split-pr --pr 42 # split an existing PR
┌─────────────────────────────────────────────────────────────┐
│ /split-pr │
│ (skill orchestrator) │
└──────────────┬──────────────────────┬───────────────────────┘
│ │
▼ ▼
┌──────────────────────┐ ┌───────────────────────┐
│ Discovery Agent │ │ Splitter Agent │
│ │ │ │
│ Classify hunks │ │ Create branches │
│ Build topic DAG │ │ Apply patches │
│ Detect vendored │ │ Run validation │
│ code, shims │ │ Create PRs via gh │
│ Recursive decomp │ │ Update descriptions │
└──────────┬───────────┘ └───────────┬────────────┘
│ │
▼ ▼
┌─────────────────────────────────────────────────────────────┐
│ split-pr-tools CLI │
│ │
│ Python library: diff_parser, analyzer, dag, state │
│ tree-sitter AST analysis, pure computation, zero git dep │
└─────────────────────────────────────────────────────────────┘
AI does the thinking (topic classification, dependency detection). Code does the math (AST analysis, DAG operations, diff parsing, lossless verification).
- Tree-sitter AST analysis — splits large files into per-function virtual hunks, so different functions can go to different PRs
- Hunk-level splitting — works with messy, entangled commits
- DAG-based dependencies — independent topics can be reviewed in parallel
- Lossless verification — every hunk accounted for before any PRs are created
- Smart grouping — vendored code stays bundled, tests travel with the code they test
- Tracking issue — auto-created checklist with clickable DAG and merge order
git clone https://github.com/emilhe/split-pr.git
cd split-pr
uv sync --all-extras # install deps
uv run pytest # run tests
uv tool install --force . # install CLI globally
# Symlink skill + agents into Claude Code
mkdir -p ~/.claude/skills/split-pr ~/.claude/agents
ln -sf "$(pwd)/commands/split-pr.md" ~/.claude/skills/split-pr/SKILL.md
ln -sf "$(pwd)/agents/pr-discovery.md" ~/.claude/agents/pr-discovery.md
ln -sf "$(pwd)/agents/pr-splitter.md" ~/.claude/agents/pr-splitter.mdAdd to ~/.claude/settings.json for permission-free CLI calls:
{
"permissions": {
"allow": [
"Bash(split-pr-tools *)",
"Write(/tmp/split-pr-*)"
]
}
}| CLI | Typer |
| Diff parsing | unidiff |
| AST analysis | tree-sitter |
| Graph algorithms | NetworkX |
| Package management | uv |
| AI orchestration | Claude Code skills + agents |
| Git operations | git + gh CLI |
MIT
