Hello.
This week, I spent all my time cleaning up the mess that Claude made throughout Sunday until dawn.
Vibe coding seems to start with debugging and end with debugging.
In the process, I've used up almost all my tokens.
Token renewal is next Monday. Since there are no tokens for this week's holiday, I'll likely take an unintended break.
But then, suddenly while looking at it, I checked /status and...
100% of your usage came from subagent-heavy sessions Each subagent runs its own requests. Be deliberate about spawning them and consider configuring a cheaper model for simpler subagents.
100% of your usage came from sessions active for 8+ hours These are often background/loop sessions. Continuous usage can add up quickly so make sure it is intentional.
98% of your usage was at >150k context Longer sessions are more expensive even when cached. /compact mid-task /clear when switching to new tasks.
19% of your usage came from subagents under "fork" If this runs frequently, consider configuring its subagents with a cheaper model or tightening their prompts
It says heavy use of subagents is causing severe token consumption, but
this week I had it monitor builds and fix build errors a lot while sleeping.
Previously, with the pattern where I ran the build and asked Claude to look at errors, more than 40% of tokens remained, but
with the instruction to monitor and complete the build, Claude executes the /loop function to check itself again, and that's all there is to it. But no matter how much I think about it, I feel like it shouldn't be a task that uses that many tokens...
While thinking about ways to save tokens,
I suddenly remember the tunallama plugin by passerby.
So, similarly,
// My instruction > Claude build > loop monitoring >
// if(error occurs) { stop error > report to Claude > check log > diagnose > take action > rebuild with Claude } >
// else if(build complete) { proceed next }
From proceeding like this, to something like below
// My instruction > Claude perception > plugin sub (local) model instruction > build > loop monitoring >
// if(error occurs) { stop error > report to Claude > check log > diagnose > take action > plugin sub (local) model instruction again } >
// else if(build complete) { report to Claude > proceed next }
I'm asking if there's a way to do this.
