auto-iterate

Iteratively improving an AI skill requires applying targeted edits, committing changes via git branching, and evaluating outcomes to keep only higher-scoring versions. The evaluation harness scores triggers, structure, constraints, and executable content, discarding changes that do not improve the score.

Automating code quality improvements involves repeatedly applying focused edits and evaluating outcomes using a user-specified evaluation command. The system commits changes, runs the deterministic evaluation, and keeps only changes that yield higher scores, rolling back any edits that fail.

When an automated skill iteration produces a lower score, the rollback mechanism discards the git commit and reverts the skill to its previous state. The system records the discarded outcome in results.tsv and applies circuit-breaker style cooldowns after repeated failures.

Yes, running automated evaluation loops on an agent memory store requires git-based branching to isolate changes. The system uses isolated subagent execution per iteration and persistent results tracking in results.tsv to manage the keep or discard decisions.

Yes, you can use a custom evaluation command instead of the built-in scripts. The system supports arbitrary code iteration by running a user-specified eval_command to determine the score delta for each committed change.

When subagent execution crashes or times out during skill iteration, the system handles it as a discard outcome. The failed iteration is recorded in results.tsv, and circuit-breaker style cooldowns are applied after repeated failures to prevent continuous crashes.