improve

Invoke the skill on the repository root; it runs a recon phase to learn build and test commands, then audits across nine categories and presents a vetted findings table. After you select findings, it writes self-contained plans under plans/ that another agent can execute.

Each plan inlines all context: exact file paths, current-state code excerpts, repo conventions with exemplar files, per-step verification commands with expected output, explicit out-of-scope boundaries, and STOP conditions. The executor needs no knowledge from the original audit session.

No. The advisor is strictly read-only on source code and only creates files under plans/ or advisor-plans/. Code changes happen only through a separate executor subagent working in an isolated git worktree, whose diff the advisor reviews before any verdict.

Yes. The branch variant scopes the audit to files changed since the merge-base with the default branch plus their direct callers. Findings are tagged as introduced by the branch or pre-existing, so legacy debt is not blamed on new work.

Every plan stamps the commit SHA it was written against and includes a drift check comparing in-scope files. The reconcile variant re-verifies done plans, refreshes drifted ones, and retires findings that were fixed independently.

It is a poor fit when you want immediate code changes in the same session, since the skill refuses to implement fixes itself. It also requires a git repository for the execute variant's worktree isolation and a host agent capable of spawning subagents.