How are people managing context in Codex?

Codex
, , ,
1

I use PrimeRadiant’s Superpowers a LOT with Codex. I really like the structure it adds to the workflow and at this point I use it on most of my larger coding tasks.

One thing I’ve noticed, though, is that my context window can fill up pretty quickly on longer sessions.

I don’t know how much of that is actually Superpowers, how much is just the size of the work I’m throwing at Codex, or how much comes from the number of skills and agents involved. That’s kind of what I’m trying to figure out.

OpenAI recently published an article about building longer-running agent workflows, and one part that caught my attention was their discussion about long-running tasks filling the context window as the agent works with tools, skills, and accumulated information.

That immediately made me think about Superpowers.

So I’m curious if there are other people here using it heavily:

How do you manage your context window?

Do you tend to start new Codex threads between different stages of the work?

Have you changed anything about how you use Superpowers to keep context usage under control?

Are there certain parts of Superpowers you only use when a task is big enough to justify them?

Or do you mostly just let Codex compact the conversation and keep going?

I’d also be really interested if anyone has ever compared context usage with and without Superpowers enabled.

I’m not trying to squeeze every possible token out of Codex. If Superpowers uses more context but helps Codex make fewer mistakes and keeps larger tasks organized, that’s a perfectly reasonable tradeoff to me.

I’m mostly curious whether other people have run into the same thing and what they’ve found works well.

OpenAI article that got me thinking about this:

2

Once compacting begins, I consider a session lost. I start a new one.

Agents.md files are short. Readmes are compact. The models’ verbosity is counterproductive here. A lot of human time is spent separating words from information. The current generation of models (Sol) makes this worse, they want to over specify trivial things.

If a module implementation is running into compaction, the module’s probably getting too big.

Compact and coherent interfaces are critical. Requirements and design docs composed by AI are rarely useful.

3

Thanks for the insight. I appreciate it!

I have been wondering recently if all the ceremony is really worth it and does it actually pay off.

I think I’m going to give it a try without all the bells and whistles for a couple of days to feel out how context window consumption reacts.

Regarding starting a new session, do you branch from the chat you’re currently in to a new one, or start a completely fresh chat when you’re nearing compaction?

4

Every module’s a clean session. If I rebuild a module it’s a clean session. The compaction is just an indicator that things might be going off the rails.

So no, no branches, no history, no nothing. I have some agents files I recycle.

5

That makes sense. I feel like I bought into the idea that more tools was better without testing A/B enough to really make that decision insightfully.

I literally just turned off a bunch of what I was using before, and already see a context consumption difference and (I did expect this) an increase in speed.

One thing I’m noticing, and I’m still not sure what’s causing this, is an almost immediate jump to around 39K tokens upon the start of a new chat. Any idea where this may be coming from or how I could do some investigating?

6

dunno. :thinking: do you have some mcp thing connected?

i imagine it depends a lot on your environment and where you start codex from. I start the cli/extension sessions directly from the appropriate directory. If you start it from a superdirectory you might have a lot more potential stuff that might get loaded.

but this is just an armchair analysis on my end.

7

Yeah I’m going to have to do some poking around. I have Linear as a plugin which I use for issue tracking, but that’s an on demand use I believe. The only actual MCPs I see are codex_app and node_repl.

8

do report back what you find! I imagine it’s super valuable for the community!

9

Will do! Once I get some actual, empirical evidence, I’ll be sure to share.