Review Git Diffs and Apply Safe Simplifications

Reviews changed code for reuse, quality, efficiency, and clarity issues, using parallel read-only sub-agents for inspection while reserving all actual edits for the main agent. It runs in one of three modes - review-only (report findings, default for "review"/"audit"/"check"), safe-fixes (also apply high-confidence fixes, default for "simplify"/"clean up"/"refactor"), or fix-and-validate (safe-fixes plus running the smallest relevant validation afterward). Step 1 determines scope in priority order - files explicitly named, current git changes, files edited earlier in the turn, or (only if reviewing with no diff available) the most recently modified tracked files - and picks the smallest correct diff command for the repo state (git diff for unstaged, git diff --cached for staged, both if mixed, or an explicit user-requested comparison), reading local instruction files (AGENTS.md, workflow docs, module style docs) first to distinguish real issues from intentional local patterns. Step 2 launches four parallel, strictly read-only review sub-agents when the scope justifies it (a tiny diff can be reviewed locally instead) - none may edit files, apply patches, stage, or commit, only report structured findings back to the main agent. Sub-agent 1 reviews for reuse opportunities: existing helpers that already solve the problem, duplicated or near-duplicate logic, inline logic that should call an existing helper. Sub-agent 2 reviews code quality: redundant or unnecessarily cached state, parameter sprawl from threading new arguments through call chains, copy-paste variations that should become a shared abstraction, leaky abstractions across module boundaries, and stringly-typed values where a typed contract already exists. Sub-agent 3 reviews efficiency: repeated work or duplicate API calls, sequential work that could run concurrently, new work added to hot paths without clear need, existence pre-checks where the operation itself could just be attempted, memory/listener leaks, and overly broad reads. Sub-agent 4 reviews clarity and standards: violations of local conventions, unnecessary complexity or deep nesting, overly clever code that hurts readability, over-simplification that collapses distinct concerns, and dead code or valueless indirection - flagging only issues that materially improve maintainability, correctness, or cost, not churn for its own sake. Step 3 has the main agent aggregate all findings into a normalized shape (file/line, category, why it's a problem, recommended fix, confidence), discarding weak or duplicative findings before any editing. Step 4 stops after reporting in review-only mode; in the fix modes, only the main agent applies edits, and only high-confidence, behavior-preserving ones - duplicated code replaced with an existing helper, redundant state or dead code removed, control flow simplified without behavior change, overly broad operations narrowed, unclear locals renamed in contained scope - explicitly skipping subjective refactors needing product or architectural judgment, and never staging, committing, or pushing. Step 5, only in fix-and-validate mode, runs the smallest relevant validation (targeted tests, typecheck, or lint/format) for the touched scope, preferring scoped checks over a full suite. Step 6 closes with a summary of what was reviewed, what was fixed, what was intentionally left alone, and whether validation ran - stating plainly if the code was already clean rather than manufacturing edits.

Use it when someone asks to review a diff or explicit file scope for reuse, code quality, efficiency, or clarity issues, or asks to simplify, clean up, or refactor changed code. It's specifically for reviewing/simplifying existing changes, not for implementing new features or fixing unrelated bugs.

Input is an explicit file scope or the current git diff (staged, unstaged, or a specified comparison). Output is, depending on mode, a structured findings report (file/line, category, problem, fix, confidence) or that report plus applied high-confidence fixes and optional validation results, closed with a summary of what changed and what was left alone.

Uses parallel read-only Codex sub-agents (explorer/reviewer roles) for the four review passes, reads local AGENTS.md/workflow/style docs to distinguish real issues from intentional patterns, and runs the project's own test/typecheck/lint tooling when validation mode is active.

Developers who want a structured, multi-lens (reuse/quality/efficiency/clarity) review of recently changed code, with the option to auto-apply only the safe, behavior-preserving fixes and leave subjective judgment calls for a human.