Skip to content
loopctl

`src/reflection/llm.rs` — the model-powered reflector

Asks the model to analyze its own tool failures — the first in-tree consumer of structured output. Opt-in; one model round-trip per analyzed failure.

Key items

  • LlmReflector::new(client) + with_system_prompt(...) — the built-in analyst prompt specifies the exact FailureAnalysis JSON shape and instructs the model to prefer is_recoverable: false over inventing corrections.
  • analyze — builds one user message (Tool / Input / Schema / Error / Task / Attempt lines; schema omitted when unavailable; attempt rendered 1-indexed), calls request_structured::<FailureAnalysis>, validates.
  • validate_modified_input [feature: schema_validation] — rejects an InputFix whose modified_input fails the tool’s real schema before the retry; without the feature, corrections are trusted as-is.

Behavior notes

  • Any reflector-side failure (ReflectionError::Internal) becomes a conservative Fail in the engine.
  • The client is held as Arc<dyn ApiClient> so the boxed future can reach it without borrowing.
  • Cost model: one model call per failure analysis — the reason it isn’t the default.

Deep dive: Reflection.