Skip to content

codegen.md

Latest commit

 

History

History
85 lines (58 loc) · 3.36 KB

codegen.md

File metadata and controls

85 lines (58 loc) · 3.36 KB
Raw

Codegen

Codegen turns a validated architecture graph into a NestJS project. Two phases matter:

  1. Constructor — deterministic, no LLM. Emits modules, controllers, services, entities, DTOs, wiring from graph structure.
  2. Surgical AI fill — optional LLM pass that fills method bodies only, guided by graph metadata and file regions (via @solarch/cli subprocess).

The OSS edition has no usage caps on codegen or fill.

Opening Code mode

From the top bar, switch to Code, or use the command palette → Generate Code.

Sub-modes:

Sub-mode UI Use
Agent Fill chat Surgical AI reads/greps/edits regions; code stays out of the chat stream.
Editor Simple file editor Manual edits on generated files.

The panel opens over the canvas (same project, preserved state). First open triggers generate once; explicit regenerate runs when the diagram drifts (Update/Drift flows).

Constructor (deterministic scaffold)

The server:

  1. Loads the project graph from Neo4j.
  2. Builds an internal IR (modules, DI edges, routes, TypeORM-ish entities from Table nodes, …).
  3. Runs NestJS emitters — file tree with consistent imports and module boundaries.

Output: a downloadable zip and an in-panel file tree. Paths and class names come from node properties (ServiceName, ControllerName, TableName, …), not from LLM invention.

What Constructor does not do: invent architecture you did not draw. Missing nodes stay missing; illegal graphs should already be blocked by the Rules Engine.

Surgical AI fill

After scaffold exists, Agent fill:

  • Sends region-scoped context (method signature, related DTO/entity names from the graph).
  • Streams tool actions (read file, grep, patch region).
  • Runs @solarch/cli in an isolated subprocess (bundled in the Docker image; devs may use a sibling solarch-tools checkout locally).

Fill is rate-limited (CODEGEN_FILL_THROTTLE_LIMIT in .env) to protect the server from abuse.

Simple View and API/Docs surfaces do not trigger fill — they are read/documentation paths.

Typical workflow

flowchart LR
  draw[DrawOnCanvas] --> validate[RulesEngine]
  validate --> gen[GenerateConstructor]
  gen --> review[ReviewInEditor]
  review --> fill[OptionalSurgicalFill]
  fill --> download[DownloadZip]
Loading
  1. Complete the graph (manually or with the AI Architect).
  2. Generate Code — review module layout and empty method stubs.
  3. Optionally run Surgical fill for business logic in Services/Repositories.
  4. Download zip and drop into your monorepo, or compare with your existing codebase.

Implementation counters

The CLI can report implementation progress (implTotal, implFilled, implAi) back to the server for UI badges — structural graph mutations are separate from these counters.

Requirements

  • Valid graph (nodes connected legally).
  • For fill: working LLM provider (same as AI Architect).
  • Docker image includes CLI bits; local dev may need solarch-tools checkout for bleeding-edge CLI.

See also