Crucible · improvements
Start an improvement run
We'll try edits against your agent's known failures and known-good calls, check every edit for safety, and score what survives — so you only ever see the improvements worth reviewing.
1. Load a workflow's regression checks
2. Pick regression checks
Test cases already classified pool=regression_seed for workflow #9001 — regressions a mutated improvement draft must not break. Optional; a run can proceed with none selected.
3. Run scale
All three scale numbers are configurable per run.
How many improvement drafts each generation proposes.
Upper bound on generations. The run still stops early on budget exhaustion or a stop request. Reaching the combined-edit merge step after a plateau needs at least 4.
Total rollouts this run may spend across every generation.