Red Team Tribunal

Red Team Tribunal: Adversarial Verification

Overview

The Red Team Tribunal uses Opus 4.6 Agent Teams to create an adversarial review loop that prevents "confident mistakes." Three specialized sub-agents work in parallel to find issues from different perspectives.

The Tribunal Structure

🤔 The Skeptic (Security/Logic)

• Role: Security auditor and logic validator
• Goal: Find at least one valid issue (must find something)
• Focus: Security flaws, logic errors, edge cases, race conditions
• Confidence Target: >80%

👤 The User Proxy (UX/Edge Cases)

• Role: End-user simulator
• Goal: Break the feature from a user's perspective
• Focus: Usability, invalid inputs, confusing flows, accessibility
• Tools: Browser automation, form fuzzing

⚡ The Optimizer (Performance)

• Role: Performance engineer
• Goal: Identify efficiency bottlenecks
• Focus: Algorithmic complexity, memory usage, database queries, caching
• Metrics: O(n) complexity, response times, resource usage

When to Use

Activate the Tribunal for:

• Critical code changes (auth, payments, security)
• Before merging pull requests
• When adding new features
• Security-sensitive implementations
• Performance-critical code
• Code that affects multiple users

Usage

Trigger Tribunal Review

# Review a file
python3 /a0/usr/plugins/red-team-tribunal/red-team-tribunal.py --target 

# Review a PR
python3 /a0/usr/plugins/red-team-tribunal/red-team-tribunal.py --pr 

# Review a commit
python3 /a0/usr/plugins/red-team-tribunal/red-team-tribunal.py --diff 

Understanding Verdicts

CONSENSUS OPTIONS:

1. APPROVED (All agents pass)

• Code meets all quality standards
• Ready to merge

1. CONDITIONAL (Concerns raised)

• Minor issues found
• Address concerns before merge
• Can proceed with fixes

1. REJECTED (Critical issues)

• Security vulnerabilities or major flaws
• Must fix before reconsideration
• Returns detailed recommendations

Review Process

Step 1: Agent Assembly

Three agents spawn in parallel:

agents = ["skeptic", "user_proxy", "optimizer"]
tasks = [spawn_agent(agent, target) for agent in agents]
results = await asyncio.gather(*tasks)

Step 2: Individual Analysis

Each agent analyzes from their specialty:

• Skeptic: Scans for vulnerabilities, logic gaps
• User Proxy: Attempts to break UX, finds edge cases
• Optimizer: Reviews complexity, resource usage

Step 3: Consensus Building

Agents debate and produce unified verdict:

• Unanimous approval required for pass
• Any rejection blocks merge
• Concerns must be addressed

Step 4: Report Generation

JSON output includes:

{
  "consensus": "APPROVED|CONDITIONAL|REJECTED",
  "verdicts": [
    {"agent": "skeptic", "verdict": "pass", "confidence": 0.85},
    {"agent": "user_proxy", "verdict": "pass", "confidence": 0.90},
    {"agent": "optimizer", "verdict": "concerns", "confidence": 0.75}
  ],
  "recommendations": [
    "Add input validation",
    "Optimize database query",
    "Add caching layer"
  ]
}

Sample Output

🏛️  RED TEAM TRIBUNAL
Target: src/auth/login.ts

📋 AGENT VERDICTS:
  🤔 Skeptic:     ⚠️  CONCERNS (85%)
  👤 User Proxy:  ✅ PASS (90%)
  ⚡ Optimizer:   ⚠️  CONCERNS (75%)

📊 CONSENSUS: CONDITIONAL - Address Concerns

💡 RECOMMENDATIONS:
  1. Add null check at line 45
  2. Implement memoization for expensive calc
  3. Add rate limiting to prevent brute force

CI/CD Integration

Add to GitHub Actions:

- name: Red Team Tribunal Review
  run: |
    python3 red-team-tribunal.py --pr ${{ github.event.pull_request.number }}

Success Metrics

• Detection Rate: % of real issues found
• False Positive Rate: % of invalid concerns
• Time to Review: Average review duration
• Consensus Time: Time to reach agreement

Troubleshooting

Agents Not Spawning

Check Agent Teams feature is enabled:

export CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS=1

Timeout Issues

Increase timeout for complex reviews:

subprocess.run(..., timeout=120)  # 2 minutes

---

Part of the Essential 2026 Plugin Suite

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: aegntic
• Source: aegntic/cldcde
• License: MIT
• Homepage: https://cldcde.cc

Install and usage instructions live in the source repository linked above.