Итеративный поиск и устранение пробелов

orchestration

vc-autoresearch is a Claude Code skill that runs an iterative gap-closing loop — find gaps → fix → repeat — until agents report no remaining issues or a numeric metric goal is reached. It acts as the shared loop primitive for PVL (plan-validate-fix) and EVL (execute-validate-fix) cycles, and can also run standalone to harden specs, tests, docs, and UX. The ORCHESTRATOR drives all bookkeeping: creates a task folder, maintains a `results.tsv` tracking file, spawns parallel research and fix agents, and halts on plateau (3 cycles without improvement), hard cap (10 iterations), or regression. Three subcommands extend the core loop: `probe` (8 personas interrogate the corpus until saturation), `reason` (adversarial debate with blind judges), and `evals` (TSV trend analysis). Key parameters include `domain`, `corpus`, `verify`, `guard`, `frozen_files`, and `auto_run`, with canonical defaults defined per domain in `vc-autoresearch-spec.md`.