An Apple TV game in a day: 31 commits, 131 tests and a check that could not fail

I broke one test on purpose and ran my check script. It printed "all good" and exited with code 0. It had behaved that way the whole time I was building the game.

The game is simple: two cats on Apple TV, a dog that chases them, and school questions (in Russian) at the doors of the houses. I started on Saturday at 12:49. By Sunday 11:41 the repo had 31 commits, six TestFlight builds and 131 tests. We built it together: agents wrote the code and checked each other's work, I set the tasks and played with the kids.

The script runs the tests and both builds, and before every TestFlight upload it was supposed to stop a bad one. It couldn't. A shell pipeline returns the exit code of its last command, not the first, and the grep at the end found a matching line even in the error message. A second agent read the scripts and pointed it out. I ran the old version against a deliberately red test: exit 0. I fixed it, and the same test now stops everything.

Smaller things turned up too. A question at the second door could cut off the first one while it was being read. A corrupt save was silently overwritten by a fresh one. On the Mac, a second window started a second game on top of the first.

The voice reading the questions is mine, cloned with Qwen3-TTS on a local GPU: 144 Russian lines, each one run back through speech recognition, 0 mismatches. The music is 16 loops from ACE-Step and Lyria, four per time of day. There are 12 meows, nine of them from Freesound; I picked seven by rating without listening, and the kids will do the listening. A one-word question, «Сколько?» ("How many?"), was too short for them to follow. It is now a full sentence: "How many fish are in the picture?"

A green light I have never seen go red tells me nothing.

Not verified: ten minutes on the actual TV without frame drops, and the stress in the voice by ear. Next: a menu and a single-player mode.

Related reading