Comment by Kim_Bruning

Ok, I have part of the answer at least.

Claude recently improved Opus 4.6 to have a 1Mtoken context. Cache normally invalidates after 5 minutes.

If you come back or --continue after a break (or 5 minutes), that's a MASSIVE hit to your session limit. 250000 tokens at Max x5 will ding you 10% of your session for "Hi, I'm back".

So say you don't typically do /compact very often. And say you're not very chatty and "do the right thing" by only asking a question once in a while? You'll burn through context like crazy.

Meanwhile if you have ADHD and anthropomorphize the bleep out of your claude and chat with them all day long? Hardly a dent!

This trick seems to work for now (ymmv)

Tell your system to

   CronCreate
      cron: "*/4 * * * *"
      prompt: "heartbeat — no action needed"

And turn it back off at end of day.

I'm sure anthropic will be thrilled by this, but I don't have a better solve at this time yet.

Context management is a thing. Unfortunately you're not allowed to use any tool other than claude code with the Anthropic Subscription, so I guess this is the solve they asked for. Allowing people to write their own tools with superior context management would seem to be a no-brainer to me, but what do I know?