Pixel vs shunt

Spotify's plugin that hands an agent's large file reads to a worker model, a second model on the bill.

Updated

Short answer

Both cut what the agent reads when it opens a large file. shunt hands the read to a second model; Pixel answers from its own index, with no second model on the bill: 79.7 to 97.2% less read across 8 well-known files, median 94.5%. shunt needs a Spotify Portal instance with AiKA; Pixel is one binary.

Side by side

Published figures, different samples. Pixel's own measurement set beside shunt's published figure: not a head-to-head, and the samples differ.

MeasurePixelshunt
Less read on large files79.7 to 97.2% measuredA similar ratio, by its own benchmark
Who reads the fileThe indexA worker model
Second model on the billNoneYes
NeedsOne binarySpotify Portal with AiKA
Writes boilerplate codeNoYes

Every figure above is on the benchmarks page, with its sample size, its method and the cases where Pixel loses.

Where shunt wins

Its worker also writes boilerplate code, which Pixel does not.

How they differ

shunt keeps large reads out of the agent’s context by handing them to a worker model, whose tokens are billed too. Pixel answers the same question, what does this file contain, with pixel list-signatures: the file’s definitions from its local index, with no model involved.

Choosing

Pick shunt if your team already runs Spotify Portal with AiKA and wants a worker that also drafts boilerplate. Pick Pixel for a single binary, no second model on the bill, and the rest of its index: callers, blast radius and Git history.