Flash Khallas

A quirky Dubai highway driving game — powered by apra-fleet.

Race down Sheikh Zayed Highway, flash your headlights at slowpokes to make them move, but watch out for cops and juiced-up Nissan Patrols.

Quick Start

npm install
npm run dev        # starts dev server on :3000
npm test           # run tests
npm run test:coverage  # with coverage report

How It Works

The app has two parts:

• Wrapper (React + Vite) — sidebar with a chat panel and fleet status display
• Game (HTML5 Canvas) — runs in an iframe, pure vanilla JS

Type a feature request or bug report in the chat panel to trigger an apra-fleet doer-reviewer pair. Watch the fleet status as it plans, implements, reviews, and ships your request.

Game Controls

Key	Action
Arrow Left/Right	Switch lanes
Space / F	Flash headlights

Mobile: Swipe to switch lanes, tap to flash.

Car Types

Car	What happens when you flash
Regular (grey)	Moves aside — you score!
Cop (dark blue, flashing lights)	You get busted — KHALLAS!
Nissan Patrol (gold)	Just goes faster

Architecture

flash-khallas/
├── src/                  # React wrapper app
│   ├── components/
│   │   ├── ChatPanel     # Fleet command input
│   │   ├── GameFrame     # iframe host
│   │   └── FleetStatus   # Doer-reviewer progress
│   └── App.jsx
├── public/game/          # Standalone canvas game
│   ├── index.html
│   └── game.js
├── tests/                # Vitest unit tests
└── .github/workflows/    # CI with coverage

Contributing via Fleet

This project is designed as an apra-fleet demo. Features are tracked as GitHub Issues and implemented by fleet doer-reviewer pairs.

Chat input → apra-fleet dispatch → doer plans + implements →
reviewer reviews → commit + push → CI runs → PR raised → auto-reload

Fleet Pipeline API

The server exposes a pipeline API used by the fleet status sidebar.

Automatic response population

GET /api/pipeline automatically reads Claude session transcripts from ~/.claude/projects/ to populate the response for each completed task. The fleet member's registry entry maps to its project directory, and each session's last assistant message is matched to pipeline tasks by task ID (the [xxxxx] token in dispatch messages).

Posting full prompt + response manually

If the automatic transcript reader doesn't cover a case (e.g. the session was wiped), the PM can post the full data directly:

curl -X POST http://localhost:3001/api/dispatches \
  -H 'Content-Type: application/json' \
  -d '{
    "ts": "2026-06-19T14:00:00.000Z",
    "member": "flash-khallas-doer",
    "prompt": "Full task prompt text here...",
    "response": "Fleet member response here..."
  }'

Fields:

• ts — ISO timestamp matching the dispatch start time (used for task matching); defaults to now if omitted
• member — fleet member name (e.g. flash-khallas-doer, flash-khallas-reviewer)
• prompt — full prompt text (used as fullPrompt in the modal)
• response — response text shown in the modal Output section