diff-review（差异审查）

Diff Review - Multi-Agent Code Review

Review code diffs using multiple AI reviewers in parallel, then merge findings into a unified report.

Usage

/diff-review [<commit-hash>] [--reviewer=<gemini|codex|claude|auto|all>]

Parameters

• <commit-hash> (optional): Review a specific commit instead of uncommitted changes. Accepts full or short hash.
• --reviewer (optional): Select reviewer mode

- gemini: Use Gemini CLI only - codex: Use Codex CLI only - claude: Use Claude only (spawns new session) - auto: Auto-select best reviewer based on code (reads reviewers/auto-select.md) - all or omitted: Multi-Agent Mode - use all three reviewers in parallel

When NOT to Use This Skill

• For non-code reviews (documentation proofreading, prose editing)
• When reviewing external/third-party code you don't control
• For commit message generation (use a dedicated commit skill)

Workflow

Step 0: Check Prerequisites

Verify CLI availability:

# Check Gemini CLI
which gemini || echo "GEMINI_NOT_FOUND"

# Check Codex CLI
which codex || echo "CODEX_NOT_FOUND"

If neither CLI is found for multi-agent mode: Warn user and suggest installation.

If only some CLIs available: Proceed with available reviewers only.

Step 1: Parse Parameters

IF <commit-hash> is provided:
    Set REVIEW_MODE = "commit"
    Set COMMIT_HASH = <commit-hash>
ELSE:
    Set REVIEW_MODE = "uncommitted"

IF --reviewer is specified:
    IF --reviewer=auto:
        Read reviewers/auto-select.md and follow its logic
    ELSE IF --reviewer=gemini|codex|claude:
        Set SINGLE_REVIEWER_MODE = true
        Set REVIEWER = specified value
    ELSE IF --reviewer=all:
        Set MULTI_AGENT_MODE = true
ELSE:
    Set MULTI_AGENT_MODE = true (default)

Step 2: Get Git Diff

If REVIEW_MODE = "commit":

# Verify the commit exists
git rev-parse --verify $COMMIT_HASH

# Get commit metadata
git --no-pager log -1 --format="%H%n%s%n%an%n%ai" $COMMIT_HASH

# Get diff stats for the commit
git --no-pager diff --stat ${COMMIT_HASH}~1 $COMMIT_HASH

# Get full diff for analysis
git --no-pager diff ${COMMIT_HASH}~1 $COMMIT_HASH

# Count changed files
git --no-pager diff --name-only ${COMMIT_HASH}~1 $COMMIT_HASH | wc -l

# Count total changed lines
git --no-pager diff --numstat ${COMMIT_HASH}~1 $COMMIT_HASH | awk '{added+=$1; removed+=$2} END {print added+removed}'

If commit hash is invalid: Report "❌ Invalid commit hash: $COMMIT_HASH" and stop.

If REVIEW_MODE = "uncommitted":

# Get diff stats (staged + unstaged)
git --no-pager diff --stat HEAD 2>/dev/null || git --no-pager diff --stat

# Get full diff for analysis
git --no-pager diff HEAD 2>/dev/null || git --no-pager diff

# Count changed files
git --no-pager diff --name-only HEAD 2>/dev/null | wc -l

# Count total changed lines
git --no-pager diff --numstat HEAD 2>/dev/null | awk '{added+=$1; removed+=$2} END {print added+removed}'

If no changes detected: Report "Nothing to review - no uncommitted changes found." and stop.

Step 3: Auto-detect Expertise

Read expertise/_index.md to get detection rules.

For each expertise defined:

1. Check if diff content matches trigger_patterns
2. Check if changed files match file_patterns
3. If matched, load the corresponding expertise/*.md file

Store matched expertise prompts for injection into reviewer prompts.

Step 4: Execute Review

Multi-Agent Mode (default)

Execute all three reviewers in parallel using the exact commands below.

⚠️ CRITICAL: Each CLI has different syntax. Do NOT mix them up!

• Gemini: git diff | gemini -p "prompt" (accepts stdin pipe ✅)
• Codex: codex exec "prompt" (non-interactive headless mode, reads codebase directly ✅)
• Claude: Spawn subagent via Task tool

Construct [FULL_PROMPT] for each reviewer by combining:

1. The reviewer's role file (e.g., reviewers/gemini-role.md)
2. Any matched expertise prompts from Step 3
3. The review target context (commit hash, diff stats)

All three reviewers MUST receive the same review instructions — only the delivery mechanism differs.

PARALLEL:
    Task 1: Run Gemini Review
        - Load reviewers/gemini-role.md
        - Inject matched expertise prompts
        - Execute:
            # For specific commit (pipe the diff as context):
            git --no-pager diff ${COMMIT_HASH}~1 $COMMIT_HASH | gemini -p "[FULL_PROMPT]"
            # For uncommitted changes:
            git --no-pager diff HEAD | gemini -p "[FULL_PROMPT]"
        - Store result as GEMINI_RESULT

    Task 2: Run Codex Review
        - Load reviewers/codex-role.md
        - Inject matched expertise prompts
        - Execute using `codex exec` (NOT `codex review`):
            # For specific commit:
            codex exec --sandbox read-only --ephemeral "Review commit ${COMMIT_HASH}. [FULL_PROMPT]"
            # For uncommitted changes:
            codex exec --sandbox read-only --ephemeral "Review all uncommitted changes. [FULL_PROMPT]"
        - ⚠️ Do NOT use: `codex -p` (doesn't exist), `codex review` (flag conflicts with prompt)
        - `codex exec` reads the codebase directly — no need to pipe diff
        - `--sandbox read-only` = safe, `--ephemeral` = don't save session
        - Store result as CODEX_RESULT

    Task 3: Run Claude Review
        - Load reviewers/claude-role.md
        - Inject matched expertise prompts
        - In Claude Code: use Task tool to spawn a subagent with diff + prompt
        - Store result as CLAUDE_RESULT
END PARALLEL

Then proceed to Step 5: Coordinate & Merge.

Single Reviewer Mode

Execute only the selected reviewer using the commands above for the chosen reviewer. Output result directly to terminal, then STOP (skip Step 5).

Step 5: Coordinate & Merge (Multi-Agent Mode Only)

Load reviewers/coordinator.md and provide it with:

• GEMINI_RESULT
• CODEX_RESULT
• CLAUDE_RESULT

The Coordinator will:

1. Parse all three reviews into structured format
2. Identify similar issues across reviewers
3. Merge similar issues while preserving source attribution
4. Preserve unique issues found by only one reviewer
5. Sort by severity: Critical > High > Medium > Low
6. Generate unified report using templates/report.md

⚠️ All emoji in the report template (🔵🟢🟣🚨⚠️📝💡📊📋📎 etc.) MUST be preserved in the final output. Do not strip or replace them.

Step 6: Save Report (MANDATORY)

⚠️ This step is NOT optional. You MUST save the report to a file.

# Generate filename
# For commit review:
REPORT_FILE="diff-review-$(echo $COMMIT_HASH | cut -c1-7)-$(TZ='America/New_York' date +%Y%m%d-%H%M%S).md"
# For uncommitted changes:
REPORT_FILE="diff-review-$(TZ='America/New_York' date +%Y%m%d-%H%M%S).md"

You MUST do BOTH of the following:

1. Output the full report to terminal
2. Write the report to $REPORT_FILE in the current working directory

# Write report to file — THIS IS REQUIRED, DO NOT SKIP
cat << 'EOF' > "$REPORT_FILE"
[full report content here]
EOF

# Confirm to user
echo "📄 Report saved to: $REPORT_FILE"

Report includes:

• Review target (commit hash + message, or "uncommitted changes")
• Summary table (severity counts)
• Merged issues with source attribution
• Raw outputs in collapsible sections

File naming convention:

• Commit review: diff-review-<short-hash>-YYYYMMDD-HHMMSS.md
• Uncommitted: diff-review-YYYYMMDD-HHMMSS.md

General Review Guidelines

These guidelines apply to all reviewers:

Code Quality

• Logic correctness
• Error handling completeness
• Edge case coverage
• Code readability

Security

• Input validation
• Authentication/authorization issues
• Sensitive data exposure
• Injection vulnerabilities

Performance

• Algorithmic efficiency
• Resource management (memory, connections)
• Unnecessary computations
• Caching opportunities

Best Practices

• DRY principle adherence
• SOLID principles
• Proper abstractions
• Documentation completeness

Fallback Handling

If a reviewer fails:

IF GEMINI fails:
    Log: "⚠️ Gemini review failed: [error]"
    Continue with other reviewers

IF CODEX fails:
    Log: "⚠️ Codex review failed: [error]"
    Continue with other reviewers

IF CLAUDE fails:
    Log: "⚠️ Claude review failed: [error]"
    Continue with other reviewers

IF ALL reviewers fail:
    Output: "❌ All reviewers failed. Please check CLI installations."
    Provide installation links

If only one reviewer succeeds:

Output that reviewer's result directly without coordination step.

If two reviewers succeed:

Proceed with coordination using available results.