The Compiler Loop

The "magic" of Intend is not just prompting an LLM. It is the loops and validation that occur around it.

The Generation Pipeline

When a cache miss occurs, the AICodeGenerator takes over.

Phase 1: Context & Prompting

The PromptBuilder accepts the AST. It constructs a specialized system prompt that includes:

• Role Definition: "You are an expert TypeScript engineer..."
• Invariants: Constraints are injected as high-priority rules.
• Steps: The logical flow is presented clearly.
• Imports: Type definitions of imported items are resolved and included in the context window.

Phase 2: Generation

We call the provider (Gemini or Ollama). We request JSON or Markdown output depending on the model's capabilities, favoring structured output where possible.

Phase 3: Validation (Self-Healing)

This is the critical step. We assume the AI makes mistakes.

1. Extraction: Code is extracted from markdown code blocks.
2. Syntax Validation: We use ts.createSourceFile from the TypeScript Compiler API to check for syntax errors.
3. Compilation Check: We may run a partial type-check (if enabled).

Phase 4: Retry Loop

If Phase 3 finds errors:

1. We do not fail the build yet.
2. We construct a Diagnostic Prompt.
   • "You generated this code: [Code]"
   • "It produced these errors: [Error List]"
   • "Please fix it."
3. We recursively call the generator with this new context.

This repeats up to maxAttempts (default: 3).

Phase 5: Final Output

Only if the code passes all checks is it written to disk and committed to the CAS.