@_Garg_Sarthak It was a good experiment thanks and keep pushing!
(I started my saas 1 years ago, 500mrr today, not the financial liberty but still magic internet money!)
I feel like I can kill my weekly Claude Ultra session in like 18 hours, so I tried to understand why, and how to mitigate this situation.
Found it: issue #108124 on anthropics/claude-code π§΅
Since v2.1.317, every background subagent that finishes wakes the parent session, and each wake is a full model call that re-reads the whole context.
Run 10 agents from a 400k-token session and you pay for ~4M tokens of re-reads alone.
What I did:
β’ claude install 2.1.296 (last version without the bug)
β’ DISABLE_AUTOUPDATER=1 so it stays there
β’ CLAUDE_CODE_AUTO_COMPACT_WINDOW=500000 (compact at 500k, not 1M)
β’ fewer, bigger subagents
β’ wide sweeps via parallel `claude -p` in one bash call: zero wake-ups
let see where it go, open to discuss or any tips :D