principal-review

Point the Skill at your branch, diff, or uncommitted changes and it runs Mode A: it detects the stack, chunks the change into passes of at most 400 lines, verifies each finding against installed sources, and emits a verdict with tiered BLOCKER, MAJOR, MINOR, and NIT findings.

Every finding must include a quoted hunk, a concrete failure scenario, and evidence verified against the installed library version or official docs. BLOCKER and MAJOR findings are independently scored by a fresh subagent, and anything below 80 confidence is demoted to a question or dropped.

Yes, Mode B maps the module's architecture from first principles, researches current good practice for that stack, audits file by file, and produces a backlog of self-contained remediation tasks that subagents can execute under supervision.

Yes, all state is written to a .code-review/ directory in the repository, including an append-only findings ledger with stable IDs, an evidence log, and an INDEX.md ending with an explicit next action so a fresh session can resume after compaction.

It detects stacks from pyproject.toml, package.json, go.mod, Cargo.toml, deno.json, and similar markers, with a bundled Python standards pack. For stacks without a pack, it researches conventions and persists them as a reusable standards file.

Skip it for small emergency fixes that stop ongoing production harm, where speed outranks thoroughness and a follow-up review happens later. It is also unnecessary for changes a linter, typechecker, or compiler already fully validates.