@zeeg 1. Does the ci workflow work locally/local container?
2. Consistent dns failure types? + at the same time?
3. “One per run” - does is self heal i.e via retry?
For sure it’s the small things that let Codex down; desktop+app feature inconsistencies, project selection on chats, chat/work friction, no access to cloud agents, plans disappearing from the side bar, chat organisation is real messy…
That being said, when it all works it’s great.
Interesting observation: a lot of skill-pilling going on claiming it’s revolutionised agent capability.
I don’t doubt it, but code cleanliness and improving the quality of comms to the agent will disproportionately improve output vs skills alone.