Pi's dev note roasts rivals' context pruning: prune+spill cuts prefill up to 88%

teortaxesTex · x · 2026-08-22

Pi (@pidotdev) compared agent frameworks' context management, asking: after 50 hours of running, does an agent still know what it did?

It called DeepSeek Harness's tool result pruner permanently-lossy: keeping only head and tail of long outputs means a key error (e.g., at char 12000 of a 30k-char log) is gone forever.
Pi's approach: prune + spill — keep a small slice in context, spill full tool results to disk with file path and offset, so the model can grep/sed/read them back on demand.
Results: 19 real sessions on GLM-5.3 and DeepSeek V4 Flash showed 26–35% lower context usage and 72–88% lower uncached prefill, with full recoverability. The PR literally says "strictly better than dsh's permanently-lossy pruner".

More from coding & agent