Expand description
A3 — Dynamic Reflection & Self-Healing Loop.
A 3-stage reflection pipeline for LLM-generated VibeScript:
- Stage 1 — Search Match: Parse the generated source and verify it contains expected constructs (e.g., required capabilities, function declarations) using the A4 AST query engine.
- Stage 2 — Semantic Shape Linting: Enforce structural policies on
the parsed AST (e.g., mandatory
take:budget, forbidden calls, effect-class constraints) using the A4 policy engine. - Stage 3 — Dry-Run State Injection: Attempt a sandboxed evaluation with injected state, catching runtime errors before committing.
Each stage produces diagnostics. If any stage fails, the loop retries (up to a configurable budget), feeding the accumulated diagnostics back as context for the next generation attempt.
Structs§
- Reflection
Config - Configuration for the reflection loop.
- Reflection
Loop - The reflection engine. Runs the 3-stage pipeline on generated VibeScript.
- Reflection
Result - The overall result of the reflection loop.
- Stage
Result - The outcome of a single reflection stage.
Type Aliases§
- Reflection
Engine - Alias for
ReflectionLoopmatching multi-agent orchestration terminology.