Tribunal Kit
Harmitx7/tribunal-kit
Agentes IA com garantia de correção — pipeline de revisão Tribunal integrado
Ver no GitHubKit de arquitetura anti-alucinação para agentes IA com 40+ agentes especializados, pipeline de revisão Tribunal com 16 revisores paralelos e motor de autonomia de longa duração. Previne erros recorrentes através de engine de jurisprudência e garante execução correta do código gerado.
Instalação
npx convez-ai Harmitx7/tribunal-kit1 comando no terminal: detecta o formato do repositório (marketplace ou plugin avulso) e instala via Claude Code. Requer Node 18+ e o Claude Code instalado.
Métricas
Casos de uso
- Geração de código com validação de correção antes de commit
- Automação de desenvolvimento com agentes especializados em domínios
- Projetos long-running sem perda de contexto ou looping infinito
Veredito

Solução madura e bem arquitetada para reduzir alucinações em agentes IA através de pipeline rigoroso de revisão. Ideal para equipes que precisam de garantias de qualidade em geração de código, mas exige investimento em configuração e compreensão da metodologia Tribunal.
README
A tradução para o português deste README ainda está na fila.
[!IMPORTANT] AI GENERATES CODE. TRIBUNAL ENSURES IT WORKS.
A zero-bloat.agent/intelligence payload that upgrades your IDE with 41 specialist agents, 31 workflows, and a 16-reviewer Tribunal pipeline. Zero hallucinations. Absolute execution certainty.
▓▒░ QUICK START
Drop Tribunal into any existing project to instantly weaponize your IDE.
# Pull the intelligence payload into your project directory
npx tribunal-kit init
[!NOTE]
initautomatically generates bridge rules for Cursor, Windsurf, Gemini, Copilot, and Claude. No configuration required.
▓▒░ THE MARATHON HARNESS (v4.4.4)
The v4.4 update introduces the Marathon Harness—an engine designed to keep autonomous agents on track during long-running, multi-session projects without looping or losing context.
⛓️ Directed Acyclic Graph (DAG) Support
Cascade failures are obsolete. Features can now be declared with dependencies (--deps=1,2). If a database schema task fails, the API route task is automatically flagged as Deadlocked and bypassed until the root issue is resolved.
🧠 Failure Context & Attempt Tracking
Agents no longer blindly retry failed approaches. When a feature fails, the reason and attempt count are permanently logged into the state matrix. The next agent to attempt the feature receives the exact failure history to course-correct immediately.
🔮 Memory Distillation
Context windows dilute over time. The new distill command allows agents to forge crucial architectural decisions into a permanent distilled_context.md memory matrix, bridging the amnesia gap between long work sessions.
▓▒░ THE PIPELINE // EVIDENCE-BASED CLOSEOUT
Code generation is solved. Code correctness is the frontier.
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 -->|16 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:#00e5ff,stroke:#fff,stroke-width:2px,color:#1a1a1a;
class D critical;
class G success;
▓▒░ THE SUPREME COURT (CASE LAW ENGINE)
The Tribunal Kit features persistent memory. The AI never makes the same mistake twice and auto-learns your engineering culture.
[!WARNING] 1. The Case Law Engine Record mistakes as legal precedent. The
precedence-reviewerchecks this database locally to forcefully block the AI from repeating banned patterns.
npx tribunal-kit case add(Record an AI hallucination)
[!TIP] 2. Skill Evolution Forge Stop writing manual rules. The system reads your Git diffs, strips token bloat, and auto-extracts your project's architectural idioms.
npx tribunal-kit learn(Digest staged files)
▓▒░ COMMAND ARSENAL
| Slash Command | Operational Scope |
|---|---|
/generate | Full Tribunal sequence: Generate → Audit → Human Gate. |
/create | Scaffold major applications via App Builder routing. |
/enhance | Safely extend existing codebases with zero regression. |
/swarm | Fan-out orchestrator. Dispatch isolated workers, synthesize output. |
/tribunal-full | Unleash ALL 16 domain reviewers simultaneously for maximum scrutiny. |
/debug | Systematic 4-phase root-cause investigation. No guessing. |
/ui-ux-pro-max | Advanced visual aesthetic engine. No generic AI slop. |