Benchmarks

The repository contains two kinds of benchmark:

go test -bench=. -benchmem ./benchmark/
./compare/bench.sh

The Go benchmarks measure allocation and message-growth characteristics without requiring a model API. The comparison script measures process RSS against an installed Claude Code command.

Results are sensitive to operating system, Go version, model client behavior, task shape and measurement method. Published numbers must include those inputs and must distinguish runtime-only measurements from end-to-end model calls.

The benchmark suite is evidence for optimization decisions, not a compatibility test or a guarantee for every workload.