Adversarial Execution
blader/adversarial-execution
Revisão adversarial dupla em tempo de execução antes de marcar feito
Ver no GitHubSkill que implementa um portão de revisão em tempo de execução, exigindo que dois modelos independentes (Claude Opus e GPT-4) validem adversarialmente que o trabalho realmente funciona antes de marcar como concluído. Inclui captura de screenshots do produto em execução para verificar comportamento real, não apenas código.
Instalação
npx convez-ai blader/adversarial-execution1 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
- Validar que a implementação atende de fato a intenção original, não apenas passa em testes
- Verificar comportamento real de interfaces e produtos executando em navegador com screenshots
- Detectar desvios entre o plano inicial e o trabalho entregue durante execução
Veredito

Ferramenta sofisticada para quem precisa garantir qualidade além de testes verdes. Reduz significantly falsos positivos em validação de completude, mas requer setup em ambientes com acesso a múltiplos modelos e capacidade de execução em browsers. Ideal para projetos críticos onde a prova real de funcionamento importa.
README
A tradução para o português deste README ainda está na fila.
Adversarial Execution
An execution-time review gate: before any task or plan is marked done, two independent strong models review the actual work adversarially — gathering their own evidence, including browser-driven screenshots — and you loop until both converge on "intent achieved and the work actually works."
This is the execution-time counterpart to adversarial-planning. Where adversarial-planning bakes a review gate into the plan text, this skill runs that gate during execution.
Why
Green tests and a clean diff hide three failure modes. The gate sits between "I think I'm done" and "this is done" to catch them:
- Intent not achieved. It compiles and tests pass, but the work doesn't do what the task actually asked for.
- Not accurate in reality. The code reads right but the running product misbehaves — the button does nothing, the number is wrong, the layout breaks. Code reading can't see this.
- Plan-intent drift. The task closed green, but the work undermines the plan's overall goal.
How it works
- Run two independent reviewers as fresh processes — no shared context with the executor or each other:
- Claude Code — Opus 4.8, max thinking
- Codex — gpt-5.5, xhigh reasoning (read-only)
- Each reviewer gets the full original plan, the stated intent/acceptance criteria, and access to the real working state (repo, running app, logs, tests).
- For any user-facing surface, the review drives the product in a browser and captures screenshots as evidence — "looks right in the code" is not a pass.
- Each reviewer answers all three questions — intent achieved? actually works? advances the plan? — and returns either an explicit pass with evidence or a concrete gap list.
- Loop until convergence: close every gap, then re-review with fresh sessions for both. Done only when a fresh round from both reviewers surfaces no new gaps.
Both reviewers must pass. A disagreement is signal, not noise — investigate the disputed gap rather than averaging the verdicts.
Install
Claude Code
git clone https://github.com/blader/adversarial-execution.git ~/.claude/skills/adversarial-execution
Codex
git clone https://github.com/blader/adversarial-execution.git ~/.codex/skills/adversarial-execution
One-liner (auto-detect)
SKILLS_DIR="${CLAUDE_SKILLS_DIR:-${CODEX_SKILLS_DIR:-$HOME/.claude/skills}}" && \
git clone https://github.com/blader/adversarial-execution.git "$SKILLS_DIR/adversarial-execution"
Then restart your agent. The skill activates automatically whenever you are about to mark a task or plan done during execution. You can also invoke it explicitly:
Verify this is actually done before I call it. Run the adversarial execution gate on this work.
Scope
Use it when nearing completion of an implementation plan and the work has correctness or user-facing surface that must be proven, not assumed.
Not for autonomous loop skills ($turborazor, $ultrarazor, $ultrareview, $turbovac) — they define their own convergence criteria and forbid review gates.
License
MIT