BlackBox

README excerpt

# BlackBox

A hybrid QA platform where **AI agents and human testers share one board.**

Point BlackBox at a product URL. Opus 4.8 explores the app, discovers its features, and generates a
test plan. Cases land on a board where a lead assigns each to an **AI agent** or a **human tester**.
AI agents execute cases in a real browser and record pass/fail with evidence — and when an agent
**can't confidently finish a case, it escalates and relays it to a human** with a full handoff brief,
instead of guessing. Humans resolve escalations from the same board.

## How it works
**Discover → Plan → Test**
1. Create a project and point at a target URL.
2. Opus discovers the app's features/use-cases; you review and edit the list.
3. Opus generates a test plan + structured cases (graded for quality).
4. Assign each case to an AI agent or a human.
5. AI agents run cases in a real browser; statuses update live with screenshots.
6. Low-confidence or blocked cases relay to a human with context; the human resolves them.
7. A report + audit trail summarize the run.

## Stack
- Cloudflare Workers (Hono API) + a React SPA (Vite + Tailwind + shadcn/ui)
- Cloudflare Agents SDK (Durable Objects) for per-run orchestration + live WebSocket updates
- Cloudflare Browser Rendering + Playwright for agent browser execution
- Cloudflare D1 (SQLite) for persistence
- Anthropic Claude Opus 4.8 for discovery, generation, in-loop action selection, and adjudication

## How Opus 4.8 is used
Discovery (understandin