Code Jury [Vote]

4 independent AI judges evaluate your code — like America's Got Talent!

Works with Qwen Code, Claude Code, Gemini CLI, and any other AI coding assistant that supports custom commands/skills.

✨ Features

• 🎭 4 Dynamic Judges — randomly selected from 10 personality types each time
• 🌍 9 Languages — Ukrainian, English, Arabic, Russian, Polish, Spanish, German, French, Italian
• ⚖️ Weighted Voting — experts in security, architecture, testing carry more weight
• 📊 Score & Verdict — 1-10 scale with ✅/❌ verdict per judge
• 🔄 Branch Tracking — tracks progress across multiple attempts on the same branch
• 💾 Session Caching — remembers previous votes for the same diff
• 📈 Issue Persistence — shows what got fixed, what's still broken, what's new

🚀 Quick Install

One-line Installer (Recommended)

# Install for both Qwen & Claude
curl -fsSL https://raw.githubusercontent.com/thecapibara/vote-skill/main/install.sh | bash -s -- --remote

# Qwen Code only
curl -fsSL https://raw.githubusercontent.com/thecapibara/vote-skill/main/install.sh | bash -s -- --remote --qwen

# Claude Code only
curl -fsSL https://raw.githubusercontent.com/thecapibara/vote-skill/main/install.sh | bash -s -- --remote --claude

# Global installation
curl -fsSL https://raw.githubusercontent.com/thecapibara/vote-skill/main/install.sh | bash -s -- --remote --all --global

Clone & Install (Offline)

git clone https://github.com/thecapibara/vote-skill.git
cd vote-skill

# Both (default)
bash install.sh

# Qwen only
bash install.sh --qwen

# Claude only
bash install.sh --claude

# Global
bash install.sh --all --global

# Remote (download from GitHub)
bash install.sh --remote

Verify Installation

cd ~/.qwen/skills/vote  # or your project's .qwen/skills/vote
python3 scripts/select_judges.py --help

🎯 Usage

Qwen Code

Use the slash command:

/vote                     # Evaluate unstaged changes
/vote --mode lightning    # Quick check (fast, cheap)
/vote --mode thorough     # Deep review (slow, thorough)
/vote last commit         # Last commit
/vote last 3 commits      # Last 3 commits
/vote branch feature-auth # Branch changes

Or just ask naturally:

• "Evaluate my changes"
• "Check the last commit"
• "Vote on my feature branch"

Claude Code

Use the slash command:

/vote                     # Evaluate unstaged changes
/vote --mode lightning    # Quick check (fast, cheap)
/vote --mode thorough     # Deep review (slow, thorough)
/vote last commit         # Last commit
/vote last 3 commits      # Last 3 commits
/vote branch feature-auth # Branch changes
/vote file script.py      # Specific file
/vote history             # View history

Or directly run the selector script:

python3 .claude/skills/vote/scripts/select_judges.py --help

Gemini CLI

Use the skill command:

/skills code-jury            # Evaluate unstaged changes (balanced: 2× Pro + 2× Flash)

Quality modes:

# Lightning — 4× Flash-Lite (fastest & cheapest)
/skills code-jury            # Ask for lightning mode

# Flash — 4× Flash (fast but capable)
/skills code-jury            # Ask for flash mode

# Balanced — 2× Pro + 2× Flash (default)
/skills code-jury            # Default mode

# Thorough — 4× Pro (maximum quality)
/skills code-jury            # Ask for thorough mode

Or directly run the selector script:

python3 .gemini/skills/code-jury/scripts/select_judges.py --platform gemini --help

👥 The Judges

Each vote uses 4 judges randomly selected from 10 personality types:

Personality	Emoji	Weight	Claude/Qwen	Gemini	Expertise
Strict Critic	⚖️	1.7x	Sonnet	3.1 Pro	Architecture, errors, security
Supportive Mentor	🏫	1.0x	Haiku	3.1 Flash	Potential, best practices
Detail-Oriented Reviewer	🔍	1.2x	Haiku	3.1 Flash	Style, docs, tests, DRY
Creative Engineer	🎨	1.0x	Haiku	3.1 Flash	Creativity, elegance
Security Expert	🛡️	2.0x	Sonnet	3.1 Pro	Vulnerabilities, validation
Performance Optimizer	⚡	1.5x	Sonnet	3.1 Pro	Algorithms, memory, CPU
Testing Expert	🧪	1.6x	Sonnet	3.1 Pro	Unit tests, edge cases
Architecture Guru	🏛️	1.8x	Sonnet	3.1 Pro	SOLID, patterns, modularity
User Advocate	👤	1.1x	Haiku	3.1 Flash	API design, UX
Maintenance Focused	🔧	1.2x	Haiku	3.1 Flash	Readability, tech debt
Maintenance Focused	🔧	1.2x	Haiku	Readability, tech debt, legacy

Quality Modes

Each vote uses 4 judges. Choose the review depth. Modes vary by platform:

Claude / Qwen Code

Mode	Models	Use case
⚡ Lightning	4× Haiku	Quick checks, smallest diff
⚖️ Balanced	2× Sonnet + 2× Haiku	Default — optimal balance
🔍 Thorough	4× Sonnet	Important reviews, security

Gemini CLI

Mode	Models	Use case
⚡ Lightning	4× Flash-Lite	Fastest & cheapest
💨 Flash	4× Flash	Fast but capable
⚖️ Balanced	2× Pro + 2× Flash	Default — optimal balance
🔍 Thorough	4× Pro	Maximum quality

Usage: /vote --mode thorough or /vote --mode lightning

Weighted Voting Explained

• Judges with higher weight have more influence on the final verdict
• Consensus detection: Issues mentioned by 3+ judges get extra priority
• To pass: weighted "Yes" votes must exceed 60% of total weight
• Example: Security Expert (2.0x) saying ❌ hurts more than Mentor (1.0x) saying ❌

📊 Example Output

🎭 Vote Results
============================================================

👩‍⚖️ Оксана (Strict Critic) (weight: 1.7x)
  ✅ Good error handling, new functionality added
  ❌ Missing tests
  📊 6.5/10 | ❌ No

👨‍🏫 Тарас (Supportive Mentor) (weight: 1.0x)
  ✅ Clean architecture, good function names
  ❌ Needs more comments
  📊 8/10 | ✅ Yes

🔍 Ірина (Detail-Oriented Reviewer) (weight: 1.2x)
  ✅ Follows code style
  ❌ Magic numbers at line 45
  📊 7/10 | ✅ Yes

🎨 Богдан (Creative Engineer) (weight: 1.0x)
  ✅ Elegant caching solution
  ❌ Could optimize the loop
  📊 8/10 | ✅ Yes

============================================================
📊 Summary: 3.7 ✅ Yes | 1.5 ❌ No (weighted)
🎯 Average Score: 7.4/10

🎉 Great! Code passed the vote!

💡 Tips for improvement:
  • Add tests
  • Avoid magic numbers
  • Optimize the loop

🔄 Branch Tracking

When working on a feature branch, Vote skill tracks your progress:

1. First vote → ❌ Fails (finds: no tests, magic numbers)
2. Fix issues → Second vote → ✅ Shows what's fixed, what's still broken, what's new

📊 History on branch: feature-auth
============================================================

📋 Vote attempts:
  Attempt #1: 6.5/10 ❌
  Attempt #2: 7.5/10 ❌

📈 Trend: 6.5 → 7.5 (+1.0)

✅ Fixed issues:
  ✅ Added tests (was in attempt #1)

⚠️ Open issues:
  🔵 Deprecated function process()

============================================================
❌ Branch failing. Need to fix 1 issues.

🌍 Language Support

Judges names and personalities adapt to your language:

Language	Flag	Example Names
Ukrainian	🇺🇦	Оксана, Тарас, Ірина
English	🇬🇧	Alex, Maya, Chris
Arabic	🇸🇦	عمر, ليلى, حسن
Russian	🇷🇺	Алексей, Елена, Дмитрий
Polish	🇵🇱	Kasia, Tomek, Agnieszka
Spanish	🇪🇸	Carlos, Sofía, Miguel
German	🇩🇪	Hans, Greta, Fritz
French	🇫🇷	Pierre, Camille, Luc
Italian	🇮🇹	Marco, Giulia, Luca

📁 Project Structure

.qwen/skills/vote/           # Qwen Code skill
├── SKILL.md                 # Skill description
├── scripts/
│   ├── select_judges.py     # Main entry point — judge selection
│   ├── judge_profiles/      # 9 languages (names & personalities)
│   └── judge_review_guides/ # Detailed checklists for each judge type
├── sessions/                # Vote cache (gitignored)
└── branches/                # Branch history (gitignored)

.claude/
├── commands/
│   └── vote.md              # /vote slash command for Claude
├── agents/
│   ├── jury-judge-haiku.md  # Lightweight sub-agent
│   └── jury-judge-sonnet.md # Expert sub-agent
└── skills/vote/             # Claude Code skill
    └── scripts/             # Same scripts as Qwen Code

.gemini/skills/code-jury/    # Gemini CLI skill
├── SKILL.md                 # Skill description
└── scripts/                 # Same scripts as Qwen Code
.gemini/agents/
├── jury-judge-pro.md        # Expert sub-agent (Gemini 3.1 Pro)
└── jury-judge-flash.md      # Lightweight sub-agent (Gemini 3.1 Flash/Lite)

judges.json                  # Source of truth for judge weights, models, review guides
install.sh                   # Installer (--qwen, --claude, --gemini, --remote)
select_judges.py             # CLI: judge selection & localization

⚙️ Configuration

Key configuration files:

• judges.json — judge weights, models (sonnet/haiku + pro/flash/flash-lite), and focus areas
• select_judges.py — MODES_* dicts for quality mode behavior per platform
• scripts/judge_review_guides/ — detailed checklists for each judge type (10 files)
• Language profiles in scripts/judge_profiles/ — localized names & personalities (9 languages)

🔒 Privacy & Security

• Session data and branch history are NOT committed to git
• .gitignore files are included to protect sensitive information
• All data stays local — no external API calls (yet)

🛣️ Roadmap

• Real LLM API integration (Claude, GPT, Qwen)
• Custom judge personalities
• Vote report export (Markdown/PDF)
• Team sharing via git
• CI/CD integration
• Historical trends dashboard

🤝 Contributing

PRs welcome! Areas of interest:

• More language profiles
• New judge personality types
• LLM integration
• UI improvements

📝 License

MIT

👤 Authors

• 💡 Idea: thecapibara
• 🛠️ Implementation: Qwen Code + thecapibara