# Harness follow-ups

- `pi_idle.py`: Keep `openai-codex` push-only. Pi 1.1.0's
  `openai-codex-responses.js` sends `prompt_cache_key` but does not apply
  `PI_CACHE_RETENTION`, so scheduling warning and compaction would target a cold
  cache. Revisit when Pi applies `PI_CACHE_RETENTION` in its `openai-codex`
  adapter; check `openai-codex-responses.js` in the installed package.
