Can I vibe code Kameleoon?

Build a lightweight, zero-flicker A/B testing system using TypeScript and Cloudflare Workers.

1. DATA ARCHITECTURE & STORAGE:
- Use Cloudflare KV to store experiment configurations. An experiment config schema must include: experimentId (string), status ('draft' | 'active' | 'paused'), variations (array of { id: string, weight: number, payload: JSON }), and targetingRules (array of { attribute: string, operator: 'equals' | 'contains' | 'in', value: string[] }).
- User Assignment Storage: Store assignment in an HTTP-only cookie named `ab_exp_{experimentId}` containing `{ variationId, timestamp }` signed with HMAC-SHA256.

2. EDGE ROUTER (Cloudflare Worker):
- Intercept GET requests for page URLs.
- Read request headers: User-Agent, Country, Cookie, URL path.
- Evaluate targeting rules against request attributes.
- If matched, check for existing assignment cookie. If absent, execute deterministic weighted sampling based on experiment configuration weights.
- Set assignment cookie on the response.
- Perform HTML transformation using HTMLRewriter: Inject variant payload (CSS/JS modifications or text replacements) directly into the stream before returning to client to guarantee ZERO layout flicker.

3. ANALYTICS & EVENT LOGGING:
- Send an asynchronous payload to Cloudflare Analytics Engine or custom HTTP log endpoint whenever a variation is served. Payload: experimentId, variationId, visitorId (hashed IP + UA), timestamp, URL, country.
- Provide a client-side JavaScript snippet (`window.trackConversion(eventName, value)`) that fires an event beacon to `/api/v1/event` containing current experiment cookies and event metadata.

4. STATISTICAL REPORTING ENDPOINT:
- Create a Node.js/Express admin endpoint `/api/v1/results/:experimentId` querying raw event records.
- Compute total impressions, conversions, conversion rate per variation.
- Compute Z-score and p-value using a standard two-proportions z-test formula to determine statistical significance.

OUT OF SCOPE:
- Do not build a visual drag-and-drop DOM editor.
- Do not implement machine learning propensity scoring models.

$ each button prefixes agent-specific run instructions · build your own product, never copy proprietary code, trademarks or designs