A training-data factory from two services

You're training a computer-use model and need thousands of hours of screen-and-action pairs. On Hoody, the actions are already data: your task runner drives the desktop through /display/input/batch, so every click and keystroke is a JSON payload you composed. While it runs, ffmpeg captures the same display and streams it through a pipe to a collector container, which writes video shards next to the requests.

The alignment problem mostly disappears, because the action log isn't instrumentation bolted onto a GUI: it's what you sent. Twenty containers run twenty task scripts in parallel; twenty pipes deliver twenty screen recordings.

Teams build capture rigs, hook libraries, and replay harnesses for this. Here the rig is two services that already existed.

The action log is the request log you already sent.