Review and simplify code changes with parallel analysis

Simplify Code reviews a diff for reuse, quality, efficiency, and clarity issues, then optionally applies only high-confidence, behavior-preserving fixes. It runs in one of three modes: review-only (report findings, stop), safe-fixes (also apply the safe fixes), or fix-and-validate (apply fixes, then run the smallest relevant validation), defaulting to review-only for "review/audit/check" requests and safe-fixes for "simplify/clean up/refactor" requests. It first determines the diff scope - explicitly named files first, then current git changes (git diff unstaged, git diff --cached staged, or an explicit branch/commit target the user named, reviewing both if work is mixed), then files edited earlier in the turn, then most-recently-modified tracked files only as a last resort - and reads the repo's own AGENTS.md and nearby style or architecture docs before judging anything, so an intentional local pattern is not flagged as a defect.

Use it when a user asks to simplify, clean up, refactor, or review changed code and wants high-confidence, behavior-preserving improvements on a scoped diff rather than a subjective rewrite. It explicitly skips subjective refactors needing product or architectural judgment, never stages, commits, or pushes changes itself, and stops to say so if the scope is unclear rather than guessing at one.

Input is a git diff or an explicit set of files. Four parallel sub-agent reviews cover reuse (duplicated logic, missed existing helpers), quality (redundant state, parameter sprawl, leaky abstractions, stringly-typed values), efficiency (repeated work, needless sequential steps, hot-path additions, memory/listener leaks), and clarity (convention violations, deep nesting, over-simplification, dead code) - each reporting file, line or symbol, problem, recommended fix, and a confidence level of high, medium, or low. Findings below high confidence, or that conflict with local instructions, are discarded before any edit. Output in safe-fixes/fix-and-validate mode is the applied fixes plus a closing summary: what was reviewed, what was fixed, what was intentionally left alone, and whether validation ran - or an explicit statement that the code was already clean if no edit was warranted.

Codex sub-agents for the four parallel review roles (or a local review pass instead, for a tiny diff); the smallest relevant validation for the touched scope in fix-and-validate mode - targeted tests, a typecheck or compile, or a formatter/lint check, preferring scoped runs over the full suite.

Developers who want a scoped diff cleaned of duplication, dead code, and inefficiency without a rewrite, and reviewers who want a structured, confidence-ranked findings list before deciding what to actually change.