feat(code-review): optimize cost and improve review quality

[pokhiii]

Summary

• Cost optimization: Single consolidated comment instead of multiple inline comments (~50% cost reduction)
• Project-specific guidelines: Claude reads all .md files from docs/code-review/ directory
• Scope awareness: Reviews only what's within the PR's stated intent
• Streamlined prompt: Prioritized focus areas, skip list for generated files

Problem

Issue	Impact
$50 credits exhausted in 2 days	Projected $1000/month vs $400 target
Single PR review costs $0.89	Too expensive
Multiple scattered comments	Cluttered PR threads
Generic feedback	Not aligned with project standards
Reviews out-of-scope items	Wastes developer time

Changes

1. Single Comment Output

• Removed mcp__github_inline_comment__create_inline_comment from allowed tools
• Added explicit instruction for ONE summary comment with structured format

2. Project Guidelines Support

• New guidelines_dir input (default: docs/code-review)
• Claude reads ALL .md files in the directory
• Projects can split guidelines: security.md, style.md, api.md, etc.

3. Scope Awareness

• Explicit instruction to read PR title/description for intent
• Only review changes within stated scope
• Out-of-scope items noted briefly, not reviewed

4. Prompt Optimization

• Removed generic "Learning" section
• Prioritized focus: Security → Bugs → Performance → Maintainability
• Skip list: lock files, minified files, generated files

Expected Impact

Metric	Before	After
Cost per review	$0.89	~$0.40
Monthly projected	$1000	~$400
Comments per review	Multiple	ONE
Guidelines support	None	Directory-based

Test plan

• Create test PR in a project using this workflow
• Add "ready for review" label
• Verify Claude creates ONE comment (not multiple)
• Verify Claude reads guidelines if docs/code-review/ exists
• Verify review stays within PR scope
• Monitor API cost in Anthropic dashboard

🤖 Generated with Claude Code