AI Dev Pipeline
MBT is built through a structured human-AI collaboration model. The product owner defines intent and authorises releases. Claude authors functional designs and verifies each unit against them. Antigravity generates all production code from those designs. No code ships without a verified functional design; no sprint closes without E2E tests green on the OCI runner.
Sprint unit flow
Defines Scope & Vision
- ·Sprint goals & scope
- ·Architecture decisions
- ·Release authorisation
Functional Design
- ·Requirements & constraints
- ·API surface specs
- ·FD committed to aidlc-state.md
Code Generation
- ·Backend + frontend code
- ·DB migrations
- ·E2E specs
- ·Commits to master
Verification & QA
- ·FD compliance check
- ·Diff review
- ·aidlc-state.md updated
E2E Gate
- ·Playwright suite on Synology
- ·374+ live tests
- ·VM auto-stop post-run
Governing principles
Design-first, code-second
Every sprint unit has a functional design committed to aidlc-state.md before any code is written. This prevents scope drift and gives the verifier a concrete specification to check against.
E2E gate is non-negotiable
Every sprint unit must ship with Playwright E2E tests covering the golden path. The OCI VM runner validates the live Synology deployment — no in-memory mocks. Tests must pass before the sprint unit closes.
Authorization chain
Antigravity implements code; Claude reviews and verifies. Neither can unilaterally decide to implement new functionality — only the Product Owner can authorize work. Prevents uncontrolled scope expansion.
Closed architecture decisions
All architectural decisions are captured in ARCHITECTURE.md and treated as closed. Formal change requests are required to alter them — no ad-hoc redesigns during implementation.