TRIBUNAL KIT

Your AI writes code that doesn't exist. Tribunal Kit stops it.

The governance layer for AI coding agents — 52 specialists, 28 reviewers, Rust core.

They import packages that don't exist. They call API methods that were deprecated two versions ago. They forget your project conventions mid-conversation. They generate code that looks right but silently breaks at runtime. No linter catches this. No type-checker catches this. The hallucination happens at the semantic layer — and until now, nothing governed it.

Tribunal Kit wraps your coding agents in a neurosymbolic verification envelope — 52 specialist agents, 28 parallel reviewers, 183 reusable skills, and a compiled Rust core — that intercepts AI output, validates it against your actual codebase, and blocks hallucinated code before it reaches disk. Works with Cursor, VSCode, Windsurf, Claude Code, and Aider.

npx tribunal-kit init    # Install the governance layer (< 60 seconds)
npx tribunal-kit sync    # Bridge with Cursor / Windsurf / VSCode
npx tribunal-kit status  # Verify everything is locked in

What's Inside

📈 Comparative Analysis

AI engineering requires more than static template rules or raw linters. See how Tribunal Kit stacks up against alternatives:

⚡ State-of-the-Art Performance (Rust Core)

Tribunal Kit v6 splits heavy computational tasks between a native Rust core and a flexible JS orchestrator:

• Compiled Rust Core (tribunal-core): Powers all deterministic operations, such as path-traversal sandboxing, Levenshtein distance calculations, text merges, and memory database reads to eliminate Node startup latency.
• Zero-Latency Hash Manifests: File synchronization and updates use SHA-256 incremental hash diffs, copying only modified assets and reducing CLI setup time by 95%.
• Semaphore-Bounded Parallelism: Fully concurrent operations with thread limits (64 in Rust, 32 in Node.js) to avoid resource starvation in complex monorepos.

🛠️ Advanced Setup

Already ran npx tribunal-kit init? Here's how to go further:

# Install Git pre-push governance hook
npx tribunal-kit hook

# Scan your workspace for AI hallucinations (phantom packages, unverified APIs)
npx tribunal-kit guardrail

# Record an AI coding mistake as permanent precedent (case law)
npx tribunal-kit case add

# Optimize a skill file using the SkillOpt evolution engine
tk optimize-skill --target ./skills/auth-security.md "npm run test:auth" --epochs 5

[!TIP] The tk hook command installs a Git pre-push hook that auto-runs tk guardrail before every push — catching hallucinated imports and unverified APIs before they reach your remote.

[!NOTE] MCP Server: To connect Tribunal Kit to Cursor, VSCode, or Claude Desktop as an MCP server, see the MCP Server Integration section below.

⚖️ The Tribunal Pipeline — Mitigating AI Code Hallucinations

Code generation is solved. Code correctness is the frontier.

The Tribunal Pipeline intercepts raw agent generation and routes it through a parallel suite of 28 domain-specific reviewers before presenting changes to the developer:

graph TD
    A[User Request] -->|Intent Parsed| B(Context Broker)
    B --> C{Inner-Loop Validator}

    C -->|Syntax & Secrets Check| D[Parallel Tribunal Review]
    C -.->|Failed| E[Maker Auto-Correction]
    E -.-> C

    D -->|28 Domain Reviewers| F[Human Gate]
    F -->|Approved| G((Committed to Disk))

    classDef default fill:#1a1a1a,stroke:#333,stroke-width:2px,color:#fff;
    classDef critical fill:#ff1637,stroke:#fff,stroke-width:2px,color:#fff;
    classDef success fill:#ccff00,stroke:#fff,stroke-width:2px,color:#1a1a1a;

    class D critical;
    class G success;

Reviewer Swarms Include:

• logic-reviewer · Semantic soundness & behavior checks.
• security-auditor · Payload boundaries, SQL injection & OWASP scanning.
• resilience-reviewer · Async error boundaries and retry logic.
• ui-ux-auditor · Structural accessibility (a11y) & premium animations.
• schema-reviewer · Type narrowing and database integrity checks.

🧠 Omniscience Cognitive Alignment Engine (OCAE)

The Omniscience Cognitive Alignment Engine (OCAE) aligns any LLM with the reasoning loops of a senior staff engineer:

• Step 0 Epistemic Loop: Always-on cognition loop enforcing confidence checking (L1–L5), knowledge freshness audits, and precision budgeting before any script execution.
• Dynamic API Trap Mitigation: Automatically guards code blocks against framework-specific compiler breakages (e.g., React 19 hook constraints, Drizzle ORM filtration issues, and Next.js 15 route headers).
• Prose Alignment Formatting: Collapses ugly lists and bullet points into highly readable, scannable documentation prose, stripping out typical AI conversational introduction/conclusion slop.

🧬 SkillOpt: Autonomous Self-Evolution Engine

Stop writing and tuning system prompts by hand. The SkillOpt Self-Evolution Engine automatically refines your prompt rules directly from test harness feedback:

# Optimize a custom skill against a test harness
tk optimize-skill --target ./skills/auth-security.md "npm run test:auth" --epochs 5 --candidates 3

1. Proposal Generation: The LLM suggests micro-patches to improve the target instruction file.
2. Rust Deduplication: Proposals are compiled, token-checked, and filtered using normalized Levenshtein similarity (default 0.85) to exclude redundant changes.
3. Harness Evaluation: The test command is executed. Successfully passing tests raise the candidate score.
4. Genetic Promotion: The highest-scoring candidate is promoted as the new baseline for the next epoch.

🏛️ Supreme Court Case Law & Memory

Tribunal Kit builds a permanent repository memory layer that spans across conversation sessions:

1. Supreme Court Case Law (tk case)

Record AI coding errors as permanent local precedents. The precedence-reviewer actively references this local database to block the AI from repeating past code defects or pattern bugs.

• Add precedence: tk case add
• Search case law: tk case search "postgres deadlock"

2. 4-Type Persistent Memory (tk memory)

Manages your project context utilizing a strict 4-category cognitive taxonomy:

• Semantic Memory — Project context (e.g., "Uses Drizzle with SQLite").
• Procedural Memory — Action guidelines (e.g., "Compile Rust binary before publishing").
• Episodic Memory — Development history and events.
• Working Memory — Current task scope.

Budget-gated recall ensures that agents only pull relevant memory segments, avoiding token window bloat and context dilution.

🏃 The Marathon Harness — Long-Running Autonomy

The Marathon Harness governs long-running multi-session tasks, keeping agents on track without looping or stalling:

• Feature DAG Graphing: Declare tasks with dependency bounds (e.g., --deps=1,2). If a core schema migration fails, dependent API route tasks are dynamically deadlocked and bypassed until fixed.
• ANSI TUI Swarm Dashboard: Intercepts verbose, noisy terminal output when running parallel swarms (tk /swarm), projecting agent research, coding, and review steps in real-time.
• Failure Context Recalls: Tracks failure histories, error stacks, and retry budgets. If a task is picked up by a new agent session, the agent receives the exact history of failed approaches to course-correct instantly.

🔌 Model Context Protocol (MCP) Server Integration

Tribunal Kit hosts an out-of-the-box Model Context Protocol (MCP) server via stdio. Connect it to Cursor, VSCode, Windsurf, or Claude Desktop to allow coding agents to query tools dynamically.

Config Example (mcp_config.json / Claude Desktop config)

{
  "mcpServers": {
    "tribunal-kit": {
      "command": "node",
      "args": ["C:/Users/sunrise/Desktop/pfojects/cli project/tribunal-kit/bin/wrapper.js"],
      "env": {
        "PROJECT_ROOT": "C:/Users/sunrise/Desktop/pfojects/cli project/tribunal-kit"
      }
    }
  }
}

Exposed MCP Tools

• verify_contracts — Proactively verifies proposed code against team contract rules before writing to disk.
• get_tribunal_skill / list_tribunal_agents — Dynamically injects skills/agent guidelines without overloading system prompts.

📜 Sovereign Covenant Protocol — AI Agent Behavioral Contract Testing (tk contract)

Design by Contract applied to AI coding agents. Declare invariants in declarative YAML rules stored in .tribunal/contracts/:

name: 'No console.log in production code'
scope: 'src/**/*.ts, src/**/*.tsx'
severity: block
must_not:
  - pattern: 'console.log'
    message: 'Use structured logger instead of console.log'

# Scaffold starter contracts
npx tribunal-kit contract init

# Run contract checks against target or modified files
npx tribunal-kit contract verify

# Replay a failure context trace snapshot
npx tribunal-kit contract replay <trace_id>

💻 CLI Command Reference

Below is the structured list of all core commands available via npx tribunal-kit <command> (or the alias tk):

🤝 Contributing & Security

We maintain high code standards and absolute runtime safety:

• Security Policy: Tribunal Kit operates with zero runtime network dependencies, zero dangerous eval executes, and sandboxed file paths to protect your codebase. Read more in SECURITY.md.
• Contributing Guide: We welcome community-authored agents, skills, and workflows! Please read CONTRIBUTING.md to set up your local development environment and run our verification tests.