Module optimizer 

The core optimization loop (Phase 3).

High-level optimization flow:

1. Run the agent on the dataset while collecting rich traces.
2. Score outputs (mechanical rules + optional LLM-as-Judge).
3. Diagnose failures using a strong model + policy + traces + code bundle.
4. Generate N targeted candidate fixes (different focus areas).
5. Validate candidates safely (cargo check + clippy + smoke tests in worktree).
6. Evaluate survivors on the full dataset.
7. Accept only net-positive changes with regression guards + holdout set.

This module is currently a structural skeleton. Real implementations of the individual steps will be filled in as the analysis crate and LLM client mature.

Structs

Candidate

A proposed improvement generated during an optimization iteration.

ModelProvenance

OptimizationRun

A single optimization experiment / iteration result.

OptimizeConfig

Configuration for a single optimization run.

Enums

EditStrategy

Functions

mechanical_score

Very rough mechanical scorer for the example agent. Gives higher score if the output is not the echo fallback.

run_optimization

Placeholder for the full optimization engine. In a real implementation this would orchestrate: