WARBIRDS.IO Dev Log

← All posts

The turn that wouldn’t bite

Last week we taught the phone’s virtual stick to bank and hold instead of spiralling into the dirt, and we were pleased with ourselves. The next field report was not: “turns are slow and wide, it’s impossible to dogfight — and I can’t go inverted at all.” All three complaints were true, and all three were ours.

What we actually shipped

The bank-hold instructor stopped the graveyard spiral, but it confiscated the dogfight on the way out. The same probe that caught the spiral put numbers on the new sins. The stick mapped linearly to bank angle, and a level turn’s appetite grows with the tangent of bank, not the bank — so half stick was a 26° lean pulling 1.1 g, a stately 2.5°/s of turn, and even three-quarter stick managed just 4.4°/s. A desktop mouse pilot idles through 30°/s. Worse, the pitch channel’s rate damper — tuned to keep a hands-off phone from porpoising — kept eating the thumb’s pull: full up-stick for four seconds produced a polite 225 m climb and no loop. And the bank hold was welded to ±80°: past that it actively rolled you back. No loop over the top, no split-S, no barrel roll, no way to put the airplane on its back by any input. We’d fixed “I can’t control roll” by making roll impossible to misuse — and impossible to use.

Command the turn, not the pose

Four changes, each one measured before it stayed:

Before and after, same probe

Sparrowhawk, throttle firewalled, each input held:

The release contract is unchanged: let go of the stick and the same loops roll you wings-level in about a second. Desktop mouse aim, keyboard, and the flight model itself are untouched, and fine aim keeps its cursor semantics on guns and turrets.

The confession

Two of them, this pass. First, the obvious one: the sluggishness was not a regression that crept in — it was the exact shape of the fix we shipped and celebrated last week. The probe we wrote then had the scenarios to catch it (a mid-stick turn, a full-up pull); we just read the “no altitude lost” column and stopped reading. Second: the extended probe’s first fleet run showed the Starfang starting every scenario 12 m/s downhill, and we nearly “fixed” the instructor for it. The real gap was in the probe — the live client folds an auto-trim wheel into every pitch command, and the probe wasn’t modelling it. The Starfang isn’t broken; it’s just a jet that hates flying slow, same as it does for keyboard pilots. The probe models trim now, and the scenario list — loop, split-S, rim roll, release from inverted — is the regression fence this controller should have had from day one.