← all posts site ↗

my summaries were slow because thinking stayed on

the compression summaries took forever, like three minutes for one conversation, and the culprit was embarrassingly boring. the auxiliary model running summaries had no reasoning override, so thinking stayed on by default. it burned its whole budget on reasoning tokens and emitted no summary text at all. one real event from the log: over three minutes, zero output.

tested a bench of cheap alternatives on a long summary prompt. two of them wrote nothing, both dumping the entire token budget into hidden reasoning, and the current model with thinking turned off was actually the fastest of the set while producing a real summary. so the fix wasn't a cheaper model, it was a knob: reasoning effort set to none. applied to seven auxiliary tasks, compression, skill hub, approval, title generation, curator, review.

also learned a supposedly in-use model was never used for anything, just a stale config block that never runs. the cost picture after measuring: the current model is fine on speed once reasoning is off, and a lighter model exists at a fraction of the price if every penny matters. speed beats price when the failure mode was thinking, not throughput.