se-optimize

Automate metric-driven optimization experiments by defining an optimization spec and running a measurement loop. The harness compares code and config variants against a fixed baseline using hard metrics or LLM-based judge scores to converge on the best solution.

Run batch optimization experiments safely using isolated worktrees and a persistent experiment log. This crash-safe setup manages phases from initial scaffolding through evaluation, enabling robust resume and recovery for long-running optimization workflows.

LLM-based judge scores are fully supported alongside hard objective metrics for measurement. You can configure the optimization harness to evaluate code variants using either quantitative data or LLM assessments based on your specific evaluation criteria.

Approval gates in automated code optimization workflows control phase transitions from setup through wrap-up by enforcing dependency approvals and stopping criteria. The orchestrator manages these gates to ensure experiments meet defined scope rules before proceeding.

Isolated worktrees are needed for optimization experiments to provide crash-safe logging and reproducible measurement environments. They prevent variant conflicts during batch execution and maintain persistent experiment logs for robust recovery of long-running optimization runs.

Limitations of automated experimentation loops include the requirement for a reproducible measurement harness and defined optimization spec. Without clear scope rules, stopping criteria, and dependency approvals, the orchestrator cannot effectively manage the iterative optimization phases.