Scope
What we're building. What we're not. What's deferred.
The Goal
A toolchain that makes "forge new → forge deploy" a complete workflow for agent-built applications.
Success = You can describe a feature, AI generates correct code against known patterns, tests verify it works, deploy ships it. The loop is tight and repeatable.
What We're Building
Phase 1: Foundation (Current)
Build rally-hq by hand, following Forge conventions. No automation yet.
Deliverables:
- Working rally-hq app with tournament management
- Component registry (documented, finite set)
- Primitives library (typed functions for all operations)
- Workflow definitions (tournament lifecycle as state machine)
- Property-based tests derived from primitives
- Deploy pipeline that "just works"
Exit criteria: rally-hq is deployed and functional. Conventions are proven, not theoretical.
Phase 2: Extraction
Extract patterns from rally-hq into reusable tooling.
Deliverables:
-
forge new- scaffolds project structure -
forge test- runs property-based tests -
forge deploy- deploys to chosen target - Project template with conventions baked in
Exit criteria: Can scaffold a new app and deploy it with three commands.
Phase 3: Generation
Add AI-assisted feature generation.
Deliverables:
- Feature spec format (structured, not freeform prompts)
-
forge feature- AI generates against known patterns - Feedback loop (test failures inform fixes)
Exit criteria: Can describe a feature and get working, tested code.
Phase 4: Polish
Make it usable by others.
Deliverables:
- Documentation
-
npx create-forge-app - Example apps beyond rally-hq
- Blog post / launch
Exit criteria: Someone else can use Forge without your help.
What We're NOT Building
| Not This | Why |
|---|---|
| A new framework | Using existing tools (Next.js/SvelteKit/etc) with opinions |
| A component library | Using shadcn/ui, adding registry metadata |
| A database ORM | Using Drizzle, adding primitive wrappers |
| A hosting platform | Using Vercel/Fly/Railway, adding deploy config |
| An AI model | Using Claude/GPT, adding structured prompts |
Forge is glue + opinions, not novel technology.
What's Deferred
| Deferred | Revisit When |
|---|---|
| Multiple runtime support (SvelteKit, Phoenix) | After Next.js version works |
| Visual editor / GUI | After CLI works |
| Team collaboration features | After single-developer flow works |
| Enterprise features (SSO, audit logs) | After open source adoption |
| Paid/hosted version | After proving value |
Non-Goals
- Competing with Vercel's v0 (different problem)
- Replacing human judgment (AI assists, doesn't replace)
- Supporting every framework (opinionated means choices)
- Backward compatibility with existing codebases (fresh starts only)