Keeping the pod alive for an hour left two idle demo agents holding slots in a
pool of five. That starves real CI, and would eventually starve this demo too
by leaving nothing free when a build actually needs to run.
Fifteen minutes covers a demo session - a warm-up run and then the real one -
without parking capacity overnight for nobody, and instanceCap stops a second
agent existing at all.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
The tests run in 0.62 seconds; the build took 191. Agent scheduling and image
pull were ~54s, a full clone of a five-commit repo ~25s, and pip install
pytest ~55s. None of that is the thing being demonstrated.
A fixed label plus idleMinutes keeps the pod alive between builds. The default
inline template mints a random POD_LABEL per build, so a fresh pod was
guaranteed every time and the pip cache in the workspace volume was always
cold. With reuse, a second run skips scheduling, skips the image pull, and
finds pytest already cached.
The first run after an idle period still pays full price, so a throwaway run
before a demo is worth doing.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
