unknown

Codex compaction endpoint is literally voodoo.

I used to reset my context window at 50% religiously because that's what I had to do with other agents. 

Until one day I heard @steipete say "no, what are you doing that for?" 

It was a hard habit to break at first because I had to learn to trust the model not to drift.  I wanted the best result, not the fastest one. 

But I'm so glad he said that to me because it changed the way I build with Codex.

Now, I spend all my time in Spec/Plan, paying attention to the details. Once that's done, I just let it ride. 

Let it compact.  I don't know how they do it but it's great.

I personally like to write the plan to a file, and then tell the agents to update the plan with a concise log of its work as it completes tasks.

This helps keep the agent on task,  especially through compaction as the state is also managed within the plan. 

I've never gone 12 hours though 😆

Peter Steinberger 🦞

A user shares their experience with Codex's compaction endpoint, noting it allows working beyond the previously recommended 50% context window reset threshold.