PR Auto-Improvement Skill

Analyzes pull requests after creation and suggests targeted improvements. Strictly opt-in — never runs automatically. User must explicitly request PR improvement.

Advisory-only — suggests improvements, never force-pushes or modifies PRs without approval (R010).

Activation

Trigger	Behavior
User says "improve this PR"	Activate analysis
User says "review PR #N"	Activate analysis
PR created automatically	Do NOT activate (opt-in only)
CI fails on PR	Suggest activation, do not auto-run

Analysis Pipeline

1. Fetch PR diff (gh pr diff)
2. Categorize changes by type:
   - New code → check patterns, naming, structure
   - Modified code → check consistency, regression risk
   - Deleted code → check for orphaned references
3. Run improvement checks (see checklist below)
4. Generate improvement report
5. User approves → create follow-up commit(s)

Improvement Checklist

Category	Checks
Code Quality	Naming consistency, dead code, duplication, complexity
Type Safety	Missing types, any usage, assertion safety
Error Handling	Unhandled promises, missing try-catch, error propagation
Testing	Missing test coverage for new functions, edge cases
Documentation	Missing JSDoc for public APIs, outdated README refs
Security	Hardcoded values, injection risks, permission checks
Performance	Unnecessary re-renders, N+1 queries, missing indexes

Report Format

[PR Auto-Improve] PR #{number} — {title}
├── Files analyzed: {count}
├── Improvements found: {count}
│
├── [Code Quality] ({count} items)
│   ├── {file:line} — {description}
│   └── {file:line} — {description}
│
├── [Testing] ({count} items)
│   └── {file} — Missing test for {function}
│
├── [Documentation] ({count} items)
│   └── {file:line} — {description}
│
└── Estimated effort: {low|medium|high}

Apply improvements? [Y/n/select]

Improvement Modes

Mode	Behavior
all	Apply all suggested improvements
select	User picks which improvements to apply
report	Report only, no changes (default)

Implementation Flow

User: "improve PR #215"
  → Orchestrator activates pr-auto-improve
  → Fetch PR diff via mgr-gitnerd
  → Analyze with appropriate expert agent(s)
  → Generate report
  → User selects improvements
  → Delegate fixes to specialist agents (R010)
  → mgr-gitnerd creates follow-up commit

Agent Selection for Fixes

File Type	Agent
*.ts, *.tsx	lang-typescript-expert
*.py	lang-python-expert
*.go	lang-golang-expert
*.kt	lang-kotlin-expert
*.java	lang-java-expert
*.rs	lang-rust-expert
Test files	qa-engineer
Docs, README	arch-documenter
Mixed	Multiple agents in parallel (R009)

Integration

Rule	Integration
R009	Multiple file fixes execute in parallel
R010	Orchestrator coordinates analysis and fix delegation
R015	Full transparency on what improvements are suggested and why
R018	3+ fix agents → Agent Teams for coordination
worker-reviewer-pipeline	Can chain: auto-improve → worker-reviewer for critical fixes
pipeline-guards	Improvement count capped by guard limits

Opt-In Safeguards

• NEVER auto-activates on PR creation
• NEVER pushes changes without user approval
• NEVER modifies PR description or labels without approval
• Report mode is default; changes require explicit "apply" command
• All git operations go through mgr-gitnerd (R010)

Limitations

• Analyzes only the PR diff, not the entire codebase
• Cannot detect architectural issues (use dev-review for that)
• Max 50 files per analysis (skip larger PRs with warning)
• Does not run tests (delegates to qa-engineer if needed)