review-and-simplify-changes（审查并简化变更）

Review and Simplify Changes

Review changed code for reuse, quality, efficiency, and clarity issues. Use Codex sub-agents to review in parallel, but keep those sub-agents read-only: they should only inspect code and send findings back to the main agent. Only the main agent may apply high-confidence, behavior-preserving fixes.

Modes

Choose the mode from the user's request:

• review-only: user asks to review, audit, or check the changes
• safe-fixes: user asks to simplify, clean up, or refactor the changes
• fix-and-validate: same as safe-fixes, but also run the smallest relevant validation after edits

If the user does not specify, default to:

• review-only for "review", "audit", or "check"
• safe-fixes for "simplify", "clean up", or "refactor"

Step 1: Determine the Scope and Diff Command

Prefer this scope order:

1. Files or paths explicitly named by the user
2. Current git changes
3. Files edited earlier in the current Codex turn
4. Most recently modified tracked files, only if the user asked for a review but there is no diff

If there is no clear scope, stop and say so briefly.

When using git changes, determine the smallest correct diff command based on the repo state:

• unstaged work: git diff
• staged work: git diff --cached
• branch or commit comparison explicitly requested by the user: use that exact diff target
• mixed staged and unstaged work: review both

Do not assume git diff HEAD is the right default when a smaller diff is available.

Before reviewing standards or applying fixes, read the repo's local instruction files and relevant project docs for the touched area. Prefer the closest applicable guidance, such as:

• AGENTS.md
• repo workflow docs
• architecture or style docs for the touched module

Use those instructions to distinguish real issues from intentional local patterns.

Step 2: Launch Four Read-Only Review Sub-Agents in Parallel

Use Codex sub-agents when the scope is large enough for parallel review to help. For a tiny diff or one very small file, it is acceptable to review locally instead.

When spawning sub-agents:

• give each sub-agent the same scope
• tell each sub-agent to inspect only its assigned review role
• tell each sub-agent it is operating in a read-only review pass
• do not let sub-agents edit files, run apply_patch, stage changes, commit, or perform other state-mutating actions
• ask for concise, structured findings only
• ask each sub-agent to report file, line or symbol, problem, recommended fix, and confidence
• ask each sub-agent to return findings to the main agent only; they must not implement fixes themselves

Use four review roles.

Sub-Agent 1: Code Reuse Review

Review the changes for reuse opportunities:

1. Search for existing helpers, utilities, or shared abstractions that already solve the same problem.
2. Flag duplicated functions or near-duplicate logic introduced in the change.
3. Flag inline logic that should call an existing helper instead of re-implementing it.

This sub-agent is read-only. It must not edit files, apply patches, or make any other workspace changes.

Recommended sub-agent role: explorer for broad codebase lookup, or reviewer if a stronger review pass is more useful than wide search.

Sub-Agent 2: Code Quality Review

Review the same changes for code quality issues:

1. Redundant state, cached values, or derived values stored unnecessarily
2. Parameter sprawl caused by threading new arguments through existing call chains
3. Copy-paste with slight variation that should become a shared abstraction
4. Leaky abstractions or ownership violations across module boundaries
5. Stringly-typed values where existing typed contracts, enums, or constants already exist

This sub-agent is read-only. It must not edit files, apply patches, or make any other workspace changes.

Recommended sub-agent role: reviewer

Sub-Agent 3: Efficiency Review

Review the same changes for efficiency issues:

1. Repeated work, duplicate reads, duplicate API calls, or unnecessary recomputation
2. Sequential work that could safely run concurrently
3. New work added to startup, render, request, or other hot paths without clear need
4. Pre-checks for existence when the operation itself can be attempted directly and errors handled
5. Memory growth, missing cleanup, or listener/subscription leaks
6. Overly broad reads or scans when the code only needs a subset

This sub-agent is read-only. It must not edit files, apply patches, or make any other workspace changes.

Recommended sub-agent role: reviewer

Sub-Agent 4: Clarity and Standards Review

Review the same changes for clarity, local standards, and balance:

1. Violations of local project conventions or module patterns
2. Unnecessary complexity, deep nesting, weak names, or redundant comments
3. Overly compact or clever code that reduces readability
4. Over-simplification that collapses separate concerns into one unclear unit
5. Dead code, dead abstractions, or indirection without value

This sub-agent is read-only. It must not edit files, apply patches, or make any other workspace changes.

Recommended sub-agent role: reviewer

Only report issues that materially improve maintainability, correctness, or cost. Do not churn code just to make it look different.

Step 3: Aggregate Findings

Wait for all review sub-agents to complete, then merge their findings.

The main agent owns this step. Treat sub-agent output as review input only, not as permission to delegate code changes back out.

Normalize findings into this shape:

1. File and line or nearest symbol
2. Category: reuse, quality, efficiency, or clarity
3. Why it is a problem
4. Recommended fix
5. Confidence: high, medium, or low

Discard weak, duplicative, or instruction-conflicting findings before editing.

Step 4: Fix Issues Carefully

In review-only mode, stop after reporting findings.

In safe-fixes or fix-and-validate mode:

• Only the main agent applies fixes for this skill
• Apply only high-confidence, behavior-preserving fixes
• Skip subjective refactors that need product or architectural judgment
• Preserve local patterns when they are intentional or instruction-backed
• Keep edits scoped to the reviewed files unless a small adjacent change is required to complete the fix correctly

Prefer fixes like:

• replacing duplicated code with an existing helper
• removing redundant state or dead code
• simplifying control flow without changing behavior
• narrowing overly broad operations
• renaming unclear locals when the scope is contained

Do not stage, commit, or push changes as part of this skill.

Step 5: Validate When Required

In fix-and-validate mode, after the main agent finishes edits, run the smallest relevant validation for the touched scope.

Examples:

• targeted tests for the touched module
• typecheck or compile for the touched target
• formatter or lint check if that is the project's real safety gate

Prefer fast, scoped validation over full-suite runs unless the change breadth justifies more.

If validation is skipped because the user asked not to run it, say so explicitly.

Step 6: Summarize Outcome

Close with a brief result:

• what was reviewed
• what was fixed, if anything
• what was intentionally left alone
• whether validation ran

If the code is already clean for this rubric, say that directly instead of manufacturing edits.