vc-autoresearch

Invoke vc-autoresearch with a domain (spec, tests, ux, docs, plan, or errors), corpus file paths, and optionally a verify command with a numeric target. The loop spawns research agents to find gaps and fix agents to resolve them, repeating until no gaps remain or the target is hit.

Set domain to tests, point the corpus at your source files, and provide a verify command that prints the coverage percentage plus a target value. The loop keeps adding tests each iteration until coverage reaches the target, with pnpm test as the default guard.

The loop stops on SUCCESS when agents find no gaps for the configured consecutive iterations or the metric target is reached. It also halts on plateau (no progress for 3 cycles), the max-iteration cap, or a regression where previously passing checks now fail.

Yes, use the frozen_files parameter with glob patterns to exclude files from the fix corpus. Fix agents must never modify files matching those patterns, and the harness domain uses this to protect frozen spec files.

PVL (plan-validate-fix) uses domain plan with a validate agent running gates against a plan document until it passes. EVL (execute-validate-fix) uses domain tests where a tester agent runs contract gate commands after execution, and failing gates trigger execute-fix agents.

HALT_PLATEAU means the gaps_found count stayed flat or increased for 3 consecutive iterations, indicating no progress. Review the results.tsv trend with the evals subcommand and consider changing the corpus scope, fix strategy, or severity floor.