Debug API

window.__hover is the dev-only API for driving + inspecting the running game. Available on pnpm dev and during Playwright runs (MODE === 'test'); stripped from production builds.

// Type lives in src/debug.ts
declare global {
  interface Window {
    __hover?: HoverDebug
  }
}

Use it from the browser DevTools console, from a Playwright spec via page.evaluate(() => window.__hover!.player()), or from the Claude Preview MCP.

Lifecycle

Property	Notes
__hover.ready	true once boot completes (renderer up, world built, player spawned). Most other accessors return null / safe defaults until then.

Renderer / loop

Method	Returns	Notes
backend()	'webgpu' | 'webgl2'	Which renderer was selected at boot.
fps()	number	Smoothed framerate.
frame()	number	Total frames since boot. Useful as a deterministic time axis in tests.

Input

Method	Notes
gamepads()	Snapshot of every connected pad: { axes, buttons, mapping, … } per index.
intent()	The current merged player intent. { throttle, steer, brake, fire, boost, pitch }.
setIntentOverride(intent | null)	Pin the player's intent. Bypasses keyboard / gamepad until you pass null. The canonical way for tests to "drive" the bike.

Player state

Method	Returns	Notes
playerEid()	number | null	ECS entity id of the player bike.
player()	{ position, velocity, groundDistance, isGrounded, speed } | null	Per-frame snapshot.
race()	{ lap, lapsToFinish, nextCheckpoint, checkpointsCrossed, totalCheckpoints, finished, raceTime } | null	Race progress.
standings()	Standing[]	Rank ordering (player + AI).

All bikes

Method	Returns	Notes
bikes()	BikeDebugSnapshot[]	Every bike's pos, vel, rot (quaternion), angvel, intent, held pickup. Critical for AI debugging.

Pickups

Method	Notes
heldPickup()	The player's held pickup type, or null.
setHeldPickup(type | null)	Force the player's held pickup. For deterministic e2e tests.
setBikeHeldPickup(eid, type | null)	Same, but for any bike — for AI-fires-pickup tests.

Combat

Method	Notes
combat()	{ shieldRemaining, stunRemaining } for the player.
combatEntityCounts()	{ mines, missiles } — count of live mine + missile entities, regardless of render.

Auto-play

Method	Notes
toggleAutoPlay()	Flip auto-play on / off. Returns the new state.
isAutoPlay()	Current auto-play state.

Patterns

Probe the player from the console

window.__hover.player()
// → { eid: 42, position: { x: 0, y: 1.2, z: -120 }, velocity: { ... }, speed: 24.3, ... }

Drive the bike in a test

await page.evaluate(() => {
  window.__hover!.setIntentOverride({
    throttle: 1, steer: 0, brake: 0, fire: false, boost: false, pitch: 0,
  })
})
await page.waitForTimeout(2000)
const player = await page.evaluate(() => window.__hover!.player())
expect(player!.speed).toBeGreaterThan(20)

Deterministic pickup test

await page.evaluate(() => window.__hover!.setHeldPickup('missile'))
await page.evaluate(() => window.__hover!.setIntentOverride({
  throttle: 1, steer: 0, brake: 0, fire: true, boost: false, pitch: 0,
}))
await page.waitForTimeout(100)
const counts = await page.evaluate(() => window.__hover!.combatEntityCounts())
expect(counts.missiles).toBe(1)

Why this exists

Two reasons:

1. e2e tests need a deterministic surface. The intent override + entity inspection lets Playwright drive scenarios without timing flake.
2. Inspection from outside the page. Claude (via the Claude Preview MCP) can preview_eval against window.__hover to read state without screenshots, which is essential for debugging "why did the bike just do that" questions during dev.

Source: src/debug.ts.