4r-review

Explicitly request a 4r-review and optionally name a target such as commits, a range, a branch, a PR, or paths. Without a named target, the Skill reviews uncommitted changes or the current branch against its base, freezing an immutable reference first.

The four lenses are R1 Risk (security and stability), R2 Readability (maintainability), R3 Reliability (correctness and testing), and R4 Resilience (recovery and observability). Standard reviews use exactly one lens chosen by the dominant risk signal; full-4r reviews use all four.

Full-4r applies when the diff touches auth, security, payments, sensitive data, or migrations, or exceeds 400 changed lines. It runs all four lenses plus exactly one refuter pass over severe inferential findings.

No, the Skill is strictly read-only and never edits code, tests, or configs. It describes suggested fixes in the report and can show minimal before/after snippets on request, but only the user applies changes.

Only BLOCKER and CRITICAL findings with causal_disposition of introduced, behavior-activated, or worsened can block. WARNING and SUGGESTION rows become info entries, and pre-existing or unknown findings are reported but never block.

Yes, after the report you can opt in to saving a review ledger at docs/reviews/{target-slug}.md in English. The ledger records findings, corroboration outcomes, and up to two scoped fix rounds against the frozen reference.