convez.ai
MCP ServerAutomaçãoAvançado
piyushpathakqa

Vigilis — Framework QA com Agentes

piyushpathakqa/argus

Testes Playwright gerados e auto-curados por agentes Claude

Ver no GitHub
InicialPluginScore · de 0 a 100

MCP server + CLI que coloca um agente Claude em loop para autoria, geração, triagem e auto-cura de testes Playwright. Explora apps web, escreve testes reais com assertions, diagnostica falhas (real bug vs DOM drift) e abre PRs de correção automaticamente.

Instalação

$npx convez-ai piyushpathakqa/argus

1 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

Estrelas
5
Forks
1
Acompanhando
0
Saúde
42%
Último commit
há 3 meses
Versão
—
Issues abertas
1
Contribuidores
2

Casos de uso

  • Gerar testes end-to-end desde a URL sem escrever código
  • Diagnosticar e corrigir quebras de teste por drift de DOM
  • Usar como gate de QA em GitHub Actions
  • Integrar QA agentic no fluxo de Claude Desktop/Code

Veredito

Ferramenta madura e sofisticada para QA agentic em produção. Ciclo completo (author → generate → triage → heal) com suporte duplo (MCP + CLI) e refusa genuína de papelaria sobre bugs reais — ideal para times que querem custo e fricção baixos em suítes frágeis.

README

A tradução para o português deste README ainda está na fila.

👁️ Vigilis — Agentic QA Framework

An agentic QA framework that authors, generates, triages, and self-heals Playwright tests for any web app — usable from Claude Code/Desktop as an MCP server or from CI as a CLI — with the whole loop running as a deployment gate in GitHub Actions.

CI QA Gate

🌐 Live site → vigilis.dev · the landing page (apps/web), built with Next.js, Framer Motion & a pre-rendered Remotion hero.

[!NOTE] M0–M3 complete, M4 in progress. The full loop runs as code: generate a test from a URL → gate it in CI → triage a failure → self-heal DOM drift with a PR (while refusing real bugs). See Roadmap.


Why

Test suites are expensive to write and brittle to maintain. Vigilis puts a Claude agent in the loop to do the slow parts: explore an app and write real Playwright tests, then — when the UI drifts — diagnose the failure and open a fix PR, while still refusing to paper over genuine bugs.

The idea: one core, two consumers

Vigilis defines its QA tools once and exposes them twice.

                     ┌──────────────────────────────┐
                     │   @argus/core                │
                     │   Agent loop (Claude)        │   Anthropic Messages API + tool use
                     │   + single Tool Registry     │   browser · dom · fs · playwright · git
                     └───────┬───────────────┬──────┘
                             │               │
              ┌──────────────▼──┐         ┌──▼───────────────┐
              │ @argus/mcp      │         │ @argus/cli       │
              │ MCP server      │         │ npx argus ...    │
              │ (Claude Desktop)│         │ (used in CI)     │
              └─────────────────┘         └──────────────────┘

The loop: four behaviors

StageInputThe agent…Output
AuthorPlain-English intentcompiles intent into a structured test plan*.plan.json
GenerateA URLexplores the app, writes specs with assertionstests/*.spec.ts
TriageA failed runclassifies real bug vs DOM drift vs flakeroot-cause report
HealA drift verdictrewrites the locator, verifies green, opens a PRa pull request

Quickstart

pnpm install
cp .env.example .env   # add your ANTHROPIC_API_KEY
pnpm build

Watch the agent run (E2E)

The agent loop is live. Point it at the bundled demo app and watch it explore — navigate, snapshot the DOM, read data-testids, and click through the login → cart flow:

pnpm --filter @argus/core exec playwright install chromium   # one-time
pnpm --filter @argus/sample-shop dev                          # terminal 1 → http://localhost:3100
node --env-file=.env packages/cli/dist/index.js smoke http://localhost:3100/login   # terminal 2

It prints a step-by-step trace and a token/cost line (~$0.05–0.15 per run on the fast model). Requires a real Anthropic API key (a Max subscription doesn't fund the API).

Or have it write a test and run it green:

node --env-file=.env packages/cli/dist/index.js generate http://localhost:3100/login --run

It explores the app, writes tests/generated/login.spec.ts, and runs it against sample-shop (3 passed, 0 failed). Defaults to Opus for quality; add --model claude-haiku-4-5 for ~10¢ runs.

See it self-heal

Seed a DOM drift, then watch Vigilis triage and fix it (full runbook: docs/DEMO.md):

NEXT_PUBLIC_ARGUS_DEMO_DRIFT=1 pnpm --filter @argus/sample-shop dev   # renames a data-testid
node --env-file=.env packages/cli/dist/index.js heal \
  http://localhost:3100/login --spec tests/generated/login.spec.ts    # → dom-drift → fix → PR

On dom-drift it rewrites the locator, re-runs to verify green, and opens a PR. On a real bug (NEXT_PUBLIC_ARGUS_DEMO_BUG=1) it refuses and blocks the gate — Vigilis improves signal, it doesn't hide failures.

Use it in your own Playwright project

Install the vigilis package, then vigilis init scaffolds a vigilis.config.json so generate/triage/heal pick up your project's defaults (baseUrl, testDir, model) — explicit flags always override it:

npm i -D vigilis
export ANTHROPIC_API_KEY=sk-...
vigilis init                                       # detects playwright.config.*, writes vigilis.config.json
vigilis generate https://your-app.com/login --run  # explore → write + run a real spec

Or point any command at any URL directly; --base-url runs generated specs against any host:

node --env-file=.env packages/cli/dist/index.js generate https://your-app.com/login \
  --run --base-url https://your-app.com

Generated files are plain Playwright specs — copy them into your repo's test folder and own them like any other test. (An npm install-able argus package is on the roadmap.)

Drive it from Claude Desktop / Code (MCP)

@argus/mcp exposes the QA tools over MCP. Setup + config snippets: docs/MCP.md.

Provenance receipts

When the Treeship CLI is installed, every argus heal run is sealed into a signed, offline-verifiable receipt automatically — self-healing QA you can audit. No hard dependency; --no-receipt to opt out. See docs/TREESHIP.md.

Repo layout

argus/
├─ packages/
│  ├─ core/   # agent loop, tool registry, Claude client, prompts
│  ├─ mcp/    # MCP server wrapping the registry
│  └─ cli/    # `argus author|generate|triage|heal`
├─ apps/
│  ├─ sample-shop/   # Next.js demo target (login + products + cart)
│  └─ web/           # landing page (Next 15 · motion · Remotion hero) → Vercel
└─ tests/     # generated Playwright specs land here

Roadmap

  • ✅ M0 — Foundations · monorepo, tooling, CI
  • ✅ M1 — Core + sample-shop + Generate · argus generate <url> --run writes + runs real tests
  • ✅ M2 — CLI + GitHub Actions gate · the QA Gate check runs generated specs (red/green)
  • ✅ M3 — Triage + Heal · argus heal fixes DOM drift with a PR, refuses real bugs
  • 🚧 M4 — MCP server + polish · argus-mcp is live (docs/MCP.md); demo GIFs next
  • ✅ Landing page · apps/web deployed to Vercel → vigilis.dev

License

MIT © Piyush Pathak

Ver o original no GitHub

Bom para

Relacionados