Review and simplify code changes with parallel analysis

Reviews changed code for reuse, quality, efficiency, and clarity issues by launching four parallel review sub-agents (code reuse, code quality, efficiency, clarity and standards), each reporting file, line or symbol, the problem, a recommended fix, and a confidence level; findings are then aggregated, normalized, and stripped of weak, duplicative, or instruction-conflicting items before any fix is applied. It runs in one of three modes: review-only (report findings and stop), safe-fixes (apply only high-confidence, behavior-preserving fixes and skip subjective refactors requiring product or architectural judgment), or fix-and-validate (same as safe-fixes, plus running the smallest relevant validation - targeted tests, a typecheck or compile, or a lint/format check - afterward). Mode defaults from the user's own wording: "review", "audit", or "check" default to review-only; "simplify", "clean up", or "refactor" default to safe-fixes.

Use it when the user asks to simplify, clean up, refactor, or review changed code, or wants high-confidence, behavior-preserving improvements on a scoped diff. Scope is determined in a fixed priority order - files explicitly named by the user, current git changes, files edited earlier in the same turn, or (only if the user asked for a review with no diff available) the most recently modified tracked files - and the smallest correct git diff command is chosen based on repo state (git diff for unstaged work, git diff --cached for staged work, an explicit branch/commit comparison if requested, or both if work is mixed); it does not default to git diff HEAD when a smaller diff exists. Before reviewing or fixing, it reads the repo's own local instruction files (AGENTS.md, workflow docs, architecture or style docs for the touched module) to distinguish real issues from intentional local patterns. It never stages, commits, or pushes changes, and if the code is already clean by this rubric it says so directly rather than manufacturing edits.

Input is a scoped code diff (explicit files, git changes, or recently edited files) plus the repo's own local instruction files. Each of the four review roles inspects only its assigned dimension: reuse (existing helpers or abstractions duplicated by the change), quality (redundant state, parameter sprawl, copy-paste variation, leaky abstractions, stringly-typed values where typed contracts already exist), efficiency (duplicate reads or API calls, sequential work that could run concurrently, unneeded hot-path work, existence pre-checks where a direct attempt with error handling would do, memory or listener leaks, overly broad reads), and clarity (convention violations, deep nesting, weak names, redundant comments, over-compact or over-simplified code, dead code or indirection without value). Output in review-only mode is the normalized findings list (file/symbol, category, why it's a problem, recommended fix, confidence). In safe-fixes or fix-and-validate mode, output also includes the applied high-confidence fixes, plus, in fix-and-validate, the result of the smallest relevant validation run, and a closing summary of what was reviewed, fixed, intentionally left alone, and whether validation ran.

Uses Codex sub-agents (explorer for broad codebase lookup, reviewer for the review passes) run in parallel when the scope is large enough to benefit; for a tiny diff or single small file, reviewing locally instead is acceptable. Reads git diff / git diff --cached and repo instruction files (AGENTS.md, workflow and architecture docs) as its only other integrations.

Developers who want an automated, multi-angle code review of a specific diff - reuse, quality, efficiency, and clarity checked in parallel - with the option to have only the safe, behavior-preserving fixes applied automatically rather than a full rewrite.