For the last week in Codex, I’ve been using a single pinned thread for each app instead of deciding when to start new threads. Basically just let Codex compact its context whenever it needs to. Seems to work well.
@manton doesn’t this cost tokens like hell when it needs to load the whole context again on every request when the cache time has expired?
@dns13 Great question. In theory yes, but maybe in practice it doesn’t matter very much if the thread is being used often? I also switched to Astra, so don’t have a good way of judging before/after token usage.
@dns13 @manton codex by default auto compacts at 272k and is generally configured so you don't really think a lot about context management. there are tradeoffs with that approach of course (e.g., drift), but it's great for looping or goal tasks that will be running for extended periods of time. Other harnesses can do that, but don't set it as default. For claude and similar, better to try and have a detailed plan up front where the orchestrator agent is assigning the tasks to other agents to keep its context low