QA Argus — Agent Autônomo de Testes
Luqman-Ud-Din/blackbox-qa-agent
Audita web apps em todos os viewports e navegadores, detecta bugs e abre tickets ADO automaticamente
Ver no GitHubAgent autônomo que audita automaticamente todas as rotas, viewports e navegadores de uma aplicação web, detecta defeitos visuais e funcionais, e cria tickets no Azure DevOps com screenshots anotados. Executa 74 skills especializadas para testes responsivos, testes cross-browser, detecção de defeitos DOM, validação de formulários, análise de conteúdo e revisão visual com IA.
Instalação
npx convez-ai Luqman-Ud-Din/blackbox-qa-agent1 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
- Testes responsivos automatizados (mobile, tablet, desktop)
- Testes cross-browser paralelos (Chromium, Firefox, WebKit)
- Detecção de defeitos visuais e funcionais com IA vision
- Validação de formulários e regras de negócio
- Auditoria de acessibilidade (dark mode, RTL, reduced motion)
- Integração com Azure DevOps para filing automático de bugs
Veredito

Solução robusta e operacional para QA end-to-end totalmente automatizado. Reduz drasticamente esforço manual em testes responsivos e cross-browser, com visão IA para detectar problemas visuais que testes DOM não alcançam. Excelente para equipes que usam Azure DevOps e precisam manter qualidade em múltiplos viewports e navegadores simultaneamente.
README
A tradução para o português deste README ainda está na fila.
QA Argus
Autonomous QA engineer for Claude Code. Crawls every route × viewport × browser of your app, detects defects, and files Azure DevOps bugs with screenshots — automatically.
What it does
Runs a single command — /argus-qa:qa-argus — and gets you:
- Responsive audit across mobile / tablet / laptop / desktop viewports
- Cross-browser testing in Chromium, Firefox, and WebKit (in parallel)
- 74 specialized skills — 35 DOM detectors, 28 functional tests, 2 content reviews, 1 AI vision review, 8 pipeline skills
- DOM defect detection — overflow, typography, touch targets, images, layout, forms, accessibility, RTL, dark mode, reduced motion, forced colors, mobile keyboard, safe-area, breakpoints, zoom 200%, and 20+ more
- Form-specific testing — 15 dedicated form skills covering boundaries, special chars, validation real-time, password rules, conditional logic, file upload, wizards, datetime, OTP, comboboxes, inline edit, tag input, input masks
- Runtime defect detection — JavaScript errors, unhandled rejections, failed HTTP requests, CORS errors, slow requests, stuck loading spinners
- Content quality review — spelling, grammar, word choice, lorem ipsum leaks, mojibake, untranslated keys (visible text + placeholders + alt + aria-label)
- AI vision review — Sonnet vision scans critical-route screenshots for visual bugs DOM probes cannot see (cropped UI, broken icons, modal positioning, color contrast)
- Per-issue annotated screenshots — each ADO ticket has a single red box around its specific defect
- ADO bug filing with deduplication — same defect across viewports = one ticket; regressions on closed bugs reopen them automatically
- Resilience built-in — per-cell timeouts, batched Sonnet dispatches, streaming writes to disk so no single hanging page kills the audit
- Auto-install on first run — Playwright MCP and node_modules install themselves if missing
Works on any frontend framework (Angular, React, Vue, Next.js, Nuxt, SvelteKit, Remix, Astro, plain HTML — anything Playwright can navigate to).
Quick start
After installing the plugin, in your project root:
/argus-qa:qa-argus
On first run, a setup wizard asks for:
- Your Azure DevOps organization + project
- Your ADO Personal Access Token
- Your app's local URL + login route + test credentials
The wizard writes .claude/automation.config.json (non-sensitive config) and .claude/secrets.json (PAT + passwords, gitignored). Then the audit runs.
On every subsequent run, the agent just runs the audit using your saved config.
Common commands
| Command | Purpose |
|---|---|
/argus-qa:qa-argus | Full audit — all apps, all routes, all viewports, all browsers |
/argus-qa:qa-argus --app <name> | Audit one app only |
/argus-qa:qa-argus --route <path> | Audit one specific route |
/argus-qa:qa-argus --device mobile | Audit one viewport only |
/argus-qa:qa-argus --browser chromium | Use only one browser |
/argus-qa:qa-argus --dry-run | Detect issues but do not file ADO bugs |
/argus-qa:qa-argus --headless | Run browsers invisibly (faster) |
/argus-qa:qa-argus --no-vision | Skip the AI vision review |
/argus-qa:qa-argus --resume <runId> | Resume an audit that crashed mid-run (skips completed cells) |
/argus-qa:qa-argus-setup | Re-run the setup wizard |
How it works
0. Step 0 auto-invokes qa-argus-setup if .claude/automation.config.json is missing
1. /argus-qa:qa-argus-setup wizard → writes .claude/automation.config.json on first run
2. /argus-qa:qa-preflight → URL health, auto-installs Playwright MCP if missing,
auto-installs node_modules, verifies server, secrets
3. /argus-qa:qa-route-discovery → crawls routes via Playwright MCP (Sonnet for SPA judgment)
4. /argus-qa:qa-phase-strategy → orders routes into 3 phases (public → login → auth-gated)
5. Per-cell execution → orchestrator calls Playwright MCP per cell, dispatches each
enabled skill on its declared model (Haiku for mechanical
probes, Sonnet for judgment, vision review, content review)
6. /argus-qa:qa-vision-review → AI vision pass on critical-route screenshots (Sonnet vision)
7. /argus-qa:qa-bug-filer → deduplicates and files ADO bugs with screenshots (Sonnet)
8. /argus-qa:qa-coverage-report → prints a summary
9. qa-state.json updated → tracks chronic issues across runs
Each skill is self-contained. Browser operations happen through Playwright MCP tools (or inline-Bash fallback). No permanent Playwright runner script exists.
What you'll see
🔍 QA Agent — Run qa-20260514-rp02
Phase 1 — Public routes (4 cells) ✓
Phase 2 — Login route (1 cell) ✓ session established
Phase 3 — Auth-gated routes (7 cells) ✓
Filed 11 bugs:
✓ Created: 11
↻ Reopened: 0
✏ Commented: 0
Screenshots:
✓ Annotated: 11 (per-issue, ideal)
⚠ Clean fb: 0
✗ Missing: 0
Each ADO bug gets:
- Title:
[QA] <issueType> on <route> — <selector> - HTML description with reproduction steps, viewport details, and the failing selector
- Annotated screenshot attached, with a red box around the exact element
Configuration
Everything project-specific lives in .claude/automation.config.json in your own project (created by the wizard). The plugin itself ships only generic defaults.
The config file controls:
- ADO org + project + API version
- Repos and base branches
- Apps to audit (URL, login route, test credentials, source/assets paths)
- Viewports (default mobile / tablet / laptop / desktop)
- Cross-browser settings (which browsers to run)
- Headed vs headless mode
- Per-route priority overrides
- Detection rule thresholds
- Vision review settings
A template with placeholders ships at templates/.claude/automation.config.template.json. The setup wizard copies it to .claude/automation.config.json on first run.
Customization
Agent-wide settings (workers, headed/headless, browser list, model routing, vision scoping, grammar scoping, resilience timeouts) all live in skills/qa-argus/customize.toml. Each detector and test skill optionally has a config.json you can edit to change thresholds, disable rules, or extend selectors.
Prerequisites
Before installing, your machine needs Node.js, Playwright (with browser binaries), jq, and a few other tools. Your ADO account needs a Personal Access Token with the right scopes.
See INSTALL.md for the full prerequisite checklist and setup walkthrough.
Architecture
| Layer | What it does | Default model |
|---|---|---|
Orchestrator (qa-argus) | Loads config, sequences the pipeline, dispatches per-cell work, applies resilience gates (timeouts, batching, streaming writes) | Sonnet (session) |
Setup / preflight (qa-argus-setup, qa-argus-ready, qa-preflight) | First-run wizard (auto-invoked) + URL health + auto-installs Playwright MCP and node_modules | Haiku |
Discovery (qa-route-discovery, qa-phase-strategy) | Finds routes via MCP + orders them into phases | Sonnet (discovery), Haiku (strategy) |
| Execution | Orchestrator calls Playwright MCP per cell; runs each enabled skill's probe | Per-skill |
Detection (qa-detect-* × 35) | Layout, typography, touch, images, forms, a11y, loading, console, network, RTL, dark mode, reduced-motion, forced-colors, mobile keyboard, safe-area, orientation, breakpoints, zoom-200, sticky scroll, modal viewport fit, hover-touch, word break, form-csrf, form-a11y, form-captcha, form-error-summary, dropdown-viewport-clip, overflow-controls, responsive-images, viewport-meta, form-validation, form-autocomplete | Haiku |
Functional tests (qa-test-* × 28) | Navigation, mobile-nav, auth-flow, data-controls, widgets, states, history, idempotency, keyboard, dragdrop, i18n, theme, test-cases, plus 15 form skills (boundaries, special-chars, realtime, password-rules, conditional, file-upload, wizard, submit-state, datetime, formatted-inputs, OTP, combobox, inline-edit, tag-input, input-mask) | Haiku (Sonnet for auth-flow + test-cases) |
Content review (qa-review-content, qa-review-hidden-text × 2) | Spelling, grammar, word choice, lorem ipsum, mojibake, untranslated keys in visible text and hidden text (placeholders, alt, title, aria-label) | Sonnet |
Vision review (qa-vision-review × 1) | Multimodal screenshot review for cropped UI, broken icons, modal positioning, dropdown clipping, color contrast — visual bugs DOM probes cannot see | Sonnet vision |
Reporting (qa-bug-filer, qa-coverage-report) | Files ADO bugs + writes summary | Sonnet (filer for prose), Haiku (report) |
All detectors and functional tests are framework-agnostic. They read the rendered DOM / page state via Playwright MCP, not source code.
Prerequisites
In addition to Node.js, Playwright, and jq, install Playwright MCP (once per machine):
npm install -g @playwright/mcp@latest
Add it to your Claude Code MCP servers config so tools like browser_navigate, browser_evaluate, browser_click are available. If MCP isn't installed, the orchestrator falls back to inline Bash + Playwright snippets — slower but functional.
License
MIT — see LICENSE for details.
Contributing
Pull requests welcome. To add a new detector:
- Create a folder under
skills/qa-detect-<your-name>/ - Add
SKILL.mddescribing what it detects (instructions, not code) - Add
config.jsonwithdetectType,issueTypearray, and rule thresholds - Add
detect.js(orsetup.js+collect.jsfor Playwright-type) - Register the detector in
skills/qa-argus/customize.tomlunder[detectors]
The orchestrator picks up new detectors automatically. No core changes needed.