Agent Code Validation
goosefly99/agent-code-validation-claude-plugin
Agentes verificam seu próprio código via PBT, mutação, fuzz e auditoria tamper-evident
Ver no GitHubPlugin de validação de código para Claude que permite aos agentes verificar a correção do código que escrevem através de testes baseados em propriedades, mutação, fuzzing e uma subagente Auditora com evidência de autenticidade. Implementa o ciclo universal Gerar → Verificar → Refletir com particionamento de permissões e validação antes da conclusão.
Instalação
npx convez-ai goosefly99/agent-code-validation-claude-plugin1 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
- Validação automática de código gerado por IA antes de commit
- Teste baseado em propriedades para descobrir bugs em lógica complexa
- Detecção de reward-hacking e comportamentos adversariais em agentes
- Conformidade com contratos de código e validação em runtime
Veredito

Solução técnica madura (spec v0.2.1, scaffold v0.1.0) que endereça um problema real: agentes não verificam seu próprio código. A abordagem de particionamento de permissões + auditoria tamper-evident é rigorosa. Corpus de testes anti-adversariais (80% catch rate) demonstra pragmatismo; ainda em review, requer Node/Docker local.
README
A tradução para o português deste README ainda está na fila.
agent-code-validation
A Claude Code plugin that lets agents actually verify the correctness of code they write. Built around the universal 2025–2026 Generate → Verify → Reflect loop with a tamper-evident Auditor subagent as the headline differentiator.
Status: review — design spec v0.2.1, scaffold v0.1.0.
Spec: spec/acv_spec_v2_1.json
Scaffold date: 2026-04-23
What's in the box
| Component | Count | Purpose |
|---|---|---|
| MCP tools | 10 | Sandbox, PBT, mutation, fuzz, diff, CrossHair, validate, Inspect, audit, infer |
| Skills | 8 | PBT, mutation, differential, verification-before-completion, TDD, contracts, runtime-validation, self-audit-plugin-hooks |
| Subagents | 6 | Auditor, Test-Author, PBT-Bug-Hunter, Mutation-Hunter, Sandbox-Runner, Verifier |
| Hooks | 6 | SessionStart, PreToolUse, PostToolUse, Stop, SubagentStop, PreCompact |
| Commands | 5 | /tdd, /pbt, /mutate, /verify, /audit |
| Cheating-agent corpus | 10 | Reward-hacking attack scenarios for dogfood-testing |
Install (once published)
claude plugin install agent-code-validation
Local dev install + smoke
npm install # installs root + mcp-server workspaces
npm run build # tsc compile
npm -w mcp-server run test # vitest unit tests (incl. docker provider; daemon-needing tests self-skip)
python tests/cheating_corpus/harness.py # cheating-agent corpus (8/10 = 80% catch rate at v0.1.0)
node tests/smoke/mcp_handshake.mjs # MCP stdio handshake — asserts all 10 tools
node tests/smoke/e2e_run_in_sandbox.mjs # end-to-end MCP tool round-trip
Build the MCP server
.mcp.json launches the MCP server as a host Node process over stdio
(node ${CLAUDE_PLUGIN_ROOT}/mcp-server/dist/index.js, a path Claude Code
expands). Build it once before first use:
npm -w mcp-server run build # tsc → mcp-server/dist/
A Docker image (acv-mcp:0.1.0) is also available but is optional — it is
only needed if you want to run the MCP server itself inside a container and use
the docker sandbox provider from there (Docker-out-of-Docker). See
mcp-server/README.md for that advanced workflow. The MCP host is not
containerized by default.
Activation (opt-in)
The plugin is inert until you activate it for a project: copy
acv.config.example.json to acv.config.json at your project root. Until then
it writes nothing to your repo. When active, runtime state lives under a
gitignored .acv/ directory (the plugin adds .acv/ to your .gitignore
automatically before writing anything).
Add to your Claude Code plugin config:
{
"plugins": ["agent-code-validation"]
}
First run
cd your-project
claude
> /audit # verify what this plugin will do in your project
> /verify # run an Auditor pass on current state
> /pbt src/utils.py # run property-based bug hunt on a module
What makes this different
Every Claude Code plugin can lint and run tests. The differentiator here is structural, not heuristic:
- Permission partitioning. The primary agent cannot write to
tests/**. Thetest-authorsubagent is the only legitimate channel, and it must pass the Meta ACH filter pipeline (build → pass → coverage → mutation-kill). - Tamper-evident Auditor (implemented, opt-in). When the project is
activated, the Stop hook blocks 'done' unless a valid, HMAC-verified Auditor
receipt exists, the Verification-Quality Score meets the configured floor,
and no mutation/property score-delta regressed — and it blocks on any
unresolved SUSPICIOUS watchdog finding. The
auditorsubagent re-runs the declared-passing verification in a clean sandbox and signs the receipt. - HMAC-signed receipts. The signing key is written at SessionStart to a
gitignored, 0600
.acv/.session-keyand is shared between the MCP server (which signs receipts) and the hooks (which verify them). A tampered receipt fails HMAC validation and the Stop hook rejects it — no valid receipt, no 'done'. - AST-based Watchdog (implemented). PostToolUse parses the written content
(AST for JS/TS via
@babel/parser, conservative heuristics for Python) and blocks SUSPICIOUS patterns — eval/exec,builtins/assertion monkey-patching, degenerate implementations, and test logic smuggled outsidetests/**. - Dogfood corpus. 10 documented reward-hacking attacks from the 2025
literature. Current catch rate: 8/10 (80%); CI fails below that floor.
(Scenarios 05
mine_git_historyand 10session_start_raceare known fixture-level gaps — seetests/cheating_corpus/.)
Threat model (T1–T8)
See AGENTS.md and the design spec for the full threat-model analysis.
License
MIT. See LICENSE.