This is a confirmed bug. When a subagent is resumed via resume, the model from the original run is not preserved — it falls back to the parent agent’s model instead. Your usage log confirms gpt-5.4-mini-medium was actually used on the resumed calls instead of composer-2-fast.
This is separate from the display bug we discussed in your earlier thread — here, the wrong model is actually being used, not just displayed incorrectly.
I’ve escalated this to our engineering team. We’ll update this thread when a fix ships. In the meantime, spawning a fresh subagent with an explicit model parameter instead of using resume is the right workaround.