AI feature builder

Enterprise

The feature builder treats AI-generated code as a proposal that must pass the same checks as code written by a person. The useful unit is not a chat answer; it is a reviewable Kumiko feature that can be stored, patched, validated, and published.

The workflow

describe → generate → compile-check → review → patch → validate → publish

• ai-generate starts a new feature.ts from a natural-language description.
• ai-patch proposes focused PatternChange[] edits to existing source.
• Designer gives the author a visual editor and review surface.
• Pattern storage versions the source with tenant scope and optimistic locking.
• Publish validates and builds a stored pattern for a controlled release path.

What stays under human control

The generated source is not automatically a production deployment. The review step is where the author checks the feature boundary, access rules, data classification, dependencies, and generated UI. Compile and validation checks catch structural problems; domain review catches requirements that a compiler cannot know.

Build your own builder

An application can use the public registry and source APIs to build a narrower editor. It can use a different model, its own AST rules, or a domain-specific form. Enterprise provides the maintained generation, patch, storage, and publish composition. A custom builder owns its source isolation, validation, merge policy, review UX, and release controls.