Long Qwen sessions on build-*/archive-* tasks grew past 460 KB and the service answered
400 straight away; batch resampled them as model_http and paused after three in a row,
so those tasks could never finish their trials. context_limit pays the line like
output_limit and does not count toward the environment pause.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>