Compression of the running architect prompt?

[Sarki]

Hello,

first of all, I must admit that late-cli is awesome, I'm truly impressed by what you can deliver with it and its potential, you can be proud of your achievement.

Now, concerning my feedback, I've faced a drop when running long conversations as you obviously exceed the available context size at some point.

I've worked around this in some of my longer running prompts and especially failing ones as:
"Resume your work following (context.md), your last thoughts are in (last_thoughts.md)"

where I copy/pasted the last 2-3 thoughts provided that they're referencing the steps in implementation_plan.md

Now, I value the fact that it may be tricky but while testing with Codex, I've seen that after a while runs a "conversation compression" to allow longer running sessions without loosing its trail (around 80-90k tokens apparently?).

Perhaps something to dig in there?

Anyway, thanks a lot for your time and efforts!

[Sarki]

FYI I've made the Pull-Request #69 for this.

[mlhher]

Thank you for your kind words. As already noted under the PR please also check out #63 to see the difference and compare these branches.

The architect summarization is something that has been brought up often enough that I am fine with adding it.