Assist and Generated Workspaces
Source generation and AI Assist must come from the same model descriptors and generated API inventory. Handwritten parallel API descriptions are not an independent source of truth.
Ten Assist capabilities
The current conformance model defines ten scopes for every runtime:
- Create
- Query, including field-level progressive disclosure
- Update
- Delete
- List/page
- Expression
- Debug
- Tool API
- Runtime customization
- Ensure Schema
Each scope must teach the generated language-native lifecycle and required negative cases. The publication evidence level can differ by scope: the older nine-scope matrix has retained published evidence, while Ensure Schema has its own later render, compilation, and execution gate. Consult the pinned conformance source before describing all ten as one published evidence class.
Progressive API discovery
For application work:
- Request entity/action Assist.
- Request field-level Assist only for fields being used.
- Compile the handwritten caller against generated output.
- Treat missing or contradictory guidance as an Assist defect instead of guessing a method.
Generated application AGENTS.md files define the editing boundary. Generated
domain-library source is not the normal API-discovery path.
Workspace baseline
All seven runtimes have library/core and console targets in the committed inventory. Web/server and app/mobile targets are intentionally asymmetric. The current target names are generated into the Cross-language Conformance Status page.
A target being listed proves discoverability and structure, not successful build, execution, publication, or long-term support. Console generation is the minimum common baseline; every runtime's console should consume its matching core library, install the Runtime Module, and pass a native startup gate.
Regeneration boundary
Model, generator version, and runtime selection are inputs. Generated areas must regenerate deterministically without overwriting declared application-owned code. Review dependency, public API, and schema metadata diffs before accepting a regenerated workspace.
Use the Agent Workflow for the full model → evaluate → generate → implement → verify loop.