improve

Invoke the skill on a repository and it runs recon, audits across nine categories with parallel read-only subagents, vets findings against the actual code, then writes self-contained plans to .scratch/plans/. Each plan includes code excerpts, verification commands, and done criteria so a zero-context executor can implement it.

It audits nine categories: correctness and bugs, security, performance, test coverage, tech debt and architecture, dependencies and migrations, developer experience and tooling, docs, and product direction. Every finding requires file:line evidence, impact, effort, risk, and confidence ratings.

No. It is strictly read-only on source code and only writes plan files under .scratch/plans/. For implementation, the execute variant dispatches a separate executor subagent in an isolated git worktree, then reviews the diff without ever merging or committing to your branch.

Add quick or deep anywhere in the invocation to set the effort level; standard is the default. Quick covers hotspots with top findings only, while deep audits the whole repo with up to eight subagents including low-confidence investigate items.

Yes, but only with the explicit --issues flag on a planning invocation. It verifies gh authentication, creates one issue per plan with the plan file as the body, and records each issue URL in the plan and the index.

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