Coding harnesses run the same script, from process start to a clean exit. The chart at the top is the whole bill, integrated over the run; under it the same harnesses at their worst moment — peak CPU and peak memory. Shorter is leaner in all three.
Click a bar to open that harness's repository.
The memory axis is inverted (heavier sits lower); dashed lines = first / last request.
One number for the whole run, area rather than peak. Beta, and not a scientific measurement: the coefficients are this project's own choice, so the bars order harnesses inside one batch and nothing more.
| harness | raw CU (local core·s + GB·s) |
standard core·s (× cpuScale) |
GB·s (not scaled) |
cpuScale | CU (scaled) |
|---|
Both forms name the same thing — the repository, with the long script of the benchmark as the measurement.
KonghaYao (2026). harness-perf-benchmark: long-script resource benchmark for coding
harnesses. https://github.com/KonghaYao/harness-perf-benchmark
@misc{harnessperfbenchmark,
title = {harness-perf-benchmark: long-script resource benchmark for coding harnesses},
author = {KonghaYao},
year = {2026},
howpublished = {\url{https://github.com/KonghaYao/harness-perf-benchmark}},
note = {Unified score CU = 1 core-second + 1 GB-second of process-tree CPU and memory,
integrated over the run (Beta)}
}