code-jury

name	code-jury
description	Evaluate code changes with 4 independent AI judges. Each judge has a unique personality, expertise, and weighted vote. Use PROACTIVELY when asked to review, evaluate, or vote on code changes.

🎭 Code Jury for Gemini

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

How It Works

1. Choose quality mode (4 options for Gemini)
2. Select 4 random judges from 10 personality types
3. Spawn 4 sub-agents (Pro for expert, Flash/Flash-Lite for lightweight)
4. Each judge evaluates independently
5. Aggregate results with weighted voting + consensus detection

Step 0: Choose quality mode

Ask the user to pick a review mode, or infer from context:

Mode	Models	Use case
⚡ Lightning	4× Flash-Lite	Quick checks, smallest diff
💨 Flash	4× Flash	Fast but capable
⚖️ Balanced	2× Pro + 2× Flash	Default — optimal balance
🔍 Thorough	4× Pro	Important reviews, security

Inference:

• If user said "quick check" or "fast" → default to Lightning
• If user said "deep" or "security" → default to Thorough
• If user said "flash" → use Flash mode (Gemini-only)
• Otherwise → default to Balanced

Step 1: Get the code diff

Fallback chain — try each until one works:

1. git diff HEAD — unstaged changes
2. git diff — changes vs index
3. No .git directory → list recent files, ask user which to review
4. User requests specific scope → git show HEAD, git diff main...feature, etc.

Step 2: Detect language & select judges

Detect language from conversation context (Ukrainian → uk, English → en, etc.)

Run the judge selector:

python3 .gemini/skills/code-jury/scripts/select_judges.py --platform gemini --lang <lang> --mode <mode>

This outputs 4 judge configs with:

• assigned_name, assigned_gender, localized titles/emojis
• gemini_model_full (gemini-3.1-pro or gemini-3.1-flash or gemini-3.1-flash-lite)
• vote_weight, focus, review_guide

Step 3: Spawn 4 sub-agents in parallel

For each judge, use the appropriate sub-agent:

• model = "pro" → use jury-judge-pro sub-agent
• model = "flash" or "flash-lite" → use jury-judge-flash sub-agent

Spawn ALL 4 in parallel. Each gets:

• The judge's personality JSON (with assigned_name, assigned_gender, localized titles, focus)
• The code diff
• Their review_guide filename so they can read the detailed checklist

Gender fields:

• If assigned_gender = "male" → use emoji_male + title_male
• If assigned_gender = "female" → use emoji_female + title_female

Step 4: Aggregate results

Step 4a — Extract issues: Parse each judge's "❌ What I disliked" section. Extract the core issue.

Step 4b — Detect consensus:

• Group issues by similarity (same file/function/concept)
• 3+ judges → consensus issue (high priority)
• 2 judges → shared issue (medium priority)

Step 4c — Calculate weighted score:

• Weighted average: sum(score_i × vote_weight_i) / sum(vote_weight_i)
• Consensus penalty: For each major issue mentioned by 3+ judges, you MAY subtract 0.1 to 0.3 from the total score (max -1.0). Minor/style issues should not trigger this penalty.
• Round to 1 decimal place

Pass threshold: weighted "Yes" votes >= 50% of total weight (was 60%). Good code should pass!

Output Format

🎭 Code Jury Results
============================================================

{For each judge:}
## {emoji} {assigned_name} — {localized_title}

**Focus:** {focus areas}
**Weight:** {vote_weight}x

### Verdict: {✅ Yes / ❌ No} — Score: {X}/10

#### ✅ What I liked
- [Specific positive observation with line references]

#### ❌ What I disliked
- [Specific negative observation with line references]

#### 💡 Advice
- [Actionable improvement suggestion]

============================================================
📊 Summary: {weighted_yes} ✅ Yes | {weighted_no} ❌ No (weighted: Yes {pct}% / No {pct}%)
🎯 Weighted Score: {score}/10
🔥 Consensus Issues (3+ judges): {list or "none"}

{🎉 PASSED or 😔 FAILED}

💡 Top improvement tips:
  1. {most common issue}
  2. {second most common issue}
  3. {third most common issue}

Judge Types

Type	Model	Weight	Focus
Strict Critic	Pro	1.7x	Architecture, errors, security
Supportive Mentor	Flash	1.0x	Potential, best practices
Detail-Oriented	Flash	1.2x	Style, docs, tests, DRY
Creative Engineer	Flash	1.0x	Creativity, elegance
Security Expert	Pro	2.0x	Vulnerabilities, validation
Performance Optimizer	Pro	1.5x	Algorithms, memory, CPU
Testing Expert	Pro	1.6x	Unit tests, edge cases
Architecture Guru	Pro	1.8x	SOLID, patterns, modularity
UX Advocate	Flash	1.1x	API design, error messages
Maintenance Focus	Flash	1.2x	Readability, tech debt