token truncators are simple by their nature, while their core value is more in the snowballed combined knowledge of different bugs and quirks community had to resolve along the way, and adding such tools is a pretty unsafe action, wonder how szr guarantees that its actions won’t make llms inputs worse

Unlike tools that mainly shorten responses or truncate command output, szr checks each reduced result against the original. If it drops an error, failing test, diagnostic code, or file:line reference, szr restores it.

It also handles repeated workflows well. Identical reruns become short references, changed reruns can be shown as deltas, and the full original output remains recoverable.

The goal is simple: reduce token use without removing the evidence an agent needs to debug and verify code. Check it out and leave some feedback. :)