Action Model

Forge’s input layer is action-first: raw events (KeyboardEvent.code, gamepad axes, mouse buttons) flow into an Input instance via an InputSource, get mapped through Bindings, and your game queries the resulting actions ("jump", "move.x").

[ DOM / Gamepad / scripted ]
        │  raw events
        ▼
    InputSource ─── drain() ──► Input.advance ─── refresh ──► action state
                                                                │
                                                                ▼
                                                          input.pressed("jump")
                                                          input.axis("move.x")
                                                          input.vector("move.x", "move.y")

The flow has three stable boundaries:

1. Source — a transport (browser_source, scripted, noop_source). Provides drain(): readonly RawInput[].
2. Bindings — a Record<Action, readonly Trigger[]> and a parallel Record<Action, readonly AxisBinding[]>. Pure data; safe to serialize.
3. Input — holds current key/axis/button state, resolves actions on every advance().

Why the indirection?

Replays record actions, not keystrokes. Same recording plays back identically whether the user uses WASD or a gamepad; whether the keyboard layout is QWERTY or Dvorak; whether the future-you renames KeyA to something else. Action streams are also far smaller (one event per state transition vs. one per keypress).