Split vibe-coded diffs into clean, ordered PR stacks.
When you vibe-code with AI, it dumps everything into one branch — types, services, routes, middleware, tests — all tangled together. Nobody wants to review a 400-line PR with 14 files.
Stacksmith takes that messy diff, uses AI to analyze the dependency graph, and splits it into a clean stack of small, focused PRs. Each one passes lint and tests independently. The final result is byte-identical to your original code.
Before: 1 PR · 14 files · 422 lines · mixed concerns
After: 5 PRs · each focused · lint passing · tests passing
npm install -g stacksmith
# On your feature branch with a large diff:
stacksmith split # AI analyzes diff → generates stack plan
stacksmith apply # Creates ordered branch stack
stacksmith push # Pushes branches → opens GitHub PRsStacksmith uses AI to analyze your diff. Configure one of:
- Claude CLI (recommended): Install Claude Code — Stacksmith detects it automatically
- API Key: Set
ANTHROPIC_API_KEY,OPENAI_API_KEY, orGOOGLE_GENERATIVE_AI_KEY
| Command | What it does |
|---|---|
stacksmith split |
Analyze diff, build dependency graph, generate a stack plan with AI. Verifies each slice passes lint + tests. |
stacksmith apply |
Create one branch per slice. Each builds on the previous, forming an ordered stack. Verifies final state matches original. |
stacksmith push |
Push branches and open GitHub PRs. PR #1 is Ready; the rest are Drafts with dependency links. |
Feature branch (messy) → stacksmith split → stacksmith apply → stacksmith push
┌──────────────────────┐ ┌──────────────┐ ┌──────────────┐ ┌──────────────┐
│ 14 files, 422 lines │ → │ AI planner │ → │ 5 branches │ → │ 5 GitHub PRs │
│ types + services + │ │ dep graph │ │ stack/01-... │ │ #1 Ready │
│ routes + middleware + │ │ lint + test │ │ stack/02-... │ │ #2 Draft │
│ tests — all mixed │ │ per slice │ │ ... │ │ ... │
└──────────────────────┘ └──────────────┘ └──────────────┘ └──────────────┘
Key guarantees:
- Byte-identical: the final stacked branch matches your original code exactly
- Each slice compiles: lint and typecheck pass at every boundary
- Tests colocated: test files ship in the same PR as the code they test
- Tests run per slice: if configured, tests pass at each boundary with progressive counts
git clone https://github.com/jonzarecki/stacksmith.git
cd stacksmith
pnpm install
pnpm build
# Run
pnpm test # 162 tests across 18 files
pnpm lint # biome check
pnpm typecheck # tsc --noEmit
# Demo
pnpm demo # full demo (real AI call, ~90s)
pnpm demo --replay # replay with cached plan (no AI, ~10s)
pnpm demo:record # record demo video (requires VHS)See ARCH.md for the full architecture, file tree, and data model.
src/
├── cli/ Command handlers (split, apply, push)
├── ai/ LLM adapters (Claude CLI, Vercel AI SDK), prompts, planner
├── core/ Diff parser, dependency graph, heuristic slicer
├── git/ Git operations, branch manager, plan applier
├── github/ PR manager (Octokit)
├── ci/ Boundary checker, test runner, auto-collapse
└── utils/ Display formatting, spinners, logger