SUN — Synthetic User Network

Towards autonomous, continuous product improvement

SUN is a dockerized, prompt-driven browser evaluation tool powered by Claude. You describe what you want tested. SUN browses, captures evidence, and returns a single concrete recommendation — with reasoning, screenshots, and copy-paste implementation steps.

The vision: kick it off once, implement the recommendation, and let it run again. Evaluate → recommend → implement → repeat.

---

How it works

Write a testing prompt in plain English. SUN uses Claude to generate a structured execution plan before doing anything — so you can review the goal, steps, and constraints before approving.

Once approved, SUN launches a Playwright browser, navigates the target, captures screenshots at each step, and streams events live to the UI. The run history dropdown lets you reload any past run without re-running anything.

Every run produces a dedicated review page: one recommendation, the reasoning behind it, annotated screenshots, and a copy-paste AI prompt you can hand directly to an LLM to implement.

---

Quickstart

# 1. Copy environment template and add your Anthropic API key
cp .env.example .env
# edit .env → set CLAUDE_API_KEY

# 2. Start with Docker (preferred)
docker compose up --build

# 3. Open the UI
open http://localhost:3020

Or run directly on the host:

npm run mvp

---

Environment variables

Variable	Default	Purpose
CLAUDE_API_KEY	—	Anthropic API key (required)
CLAUDE_MODEL	claude-sonnet-4-6	Claude model to use
CLAUDE_MAX_ATTEMPTS	2	Retries on rate-limit / overload
PORT	3020	HTTP server port
SUN_ALLOWED_HOSTS	chirpper.com,...	Hosts SUN is permitted to navigate
SUN_MAX_EXECUTION_STEPS	8	Max browser steps per run
SUN_TIMEOUT_MS	15000	Per-action timeout (ms)

---

What gets stored

Each run is persisted under artifacts/runs/<uuid>/:

• run.json — full run record (plan, events, screenshots metadata, analysis)
• *.png — screenshot files captured during execution

Runs are browsable from the dropdown on the home page. No database required.

---

Tech

• Runtime: Node.js, TypeScript
• Browser: Playwright (Chromium)
• AI: Anthropic Claude API (claude-sonnet-4-6)
• Serving: Bare node:http — no framework
• Persistence: Filesystem (artifacts/runs/)
• Streaming: Server-Sent Events

---

Legacy smoke runners

Targeted Chirpper smoke runners remain available as secondary workflows:

npm run smoke
npm run smoke:multi-user-lineage

These are not the primary product story — they exist for historical coverage and targeted verification.