tak.toml

# Instruction-counted benchmarks, measured by tak (https://github.com/jdx/tak).
#
# Pure CPU against fnox's own definitions: no network, no secrets, no provider,
# no repository state. Hermeticity was checked by running each with the proxy
# pointed at a dead port; both succeed unchanged and both outputs are
# byte-identical run to run.
#
# `--version` is deliberately absent. It measures clap rather than fnox, and a
# subject that can reach the network retires a different number of instructions
# depending on conditions unrelated to the code under test.

# A provider intentionally expands the generated configuration schema. Keep
# enough room for that expected work while still catching larger regressions.
[gate]
pct = 1.25

# The whole command tree, walked and serialised to ~12KB of KDL. This is also
# the closest thing fnox has to a pure startup measurement: a subcommand added
# with an expensive default shows up here first, and every invocation pays the
# clap construction it measures.
[bench.usage]
cmd = ["./target/release/fnox", "usage"]

# The settings schema, ~43KB of JSON. A different serialisation path over a
# different tree, and the larger of the two — reading them together separates a
# change in the command definitions from one in the settings.
[bench.schema]
cmd = ["./target/release/fnox", "schema"]
runs = 10