tak.toml

# Instruction-counted benchmarks, measured by tak (https://github.com/jdx/tak).
#
# Every subject here is hermetic, and each was checked by running it with the
# proxy pointed at a dead port rather than assumed to be.
#
# The three config-resolving benchmarks run with MISE_OFFLINE=1, set by the
# `perf` task. Without it `ls`, `env` and `hook-env` reach the network and log
# HTTP warnings; with it they do not, and their output is byte-identical run to
# run. Offline covers less of mise than a real invocation does — but the part it
# leaves out is precisely the part no benchmark can measure repeatably.
#
# They read `benchmarks/perf-fixture`, a committed config pinning exact
# versions. A floating version would make the measurement depend on what the
# registry said that day.
#
# `mise --version` and `mise x -- echo` are deliberately absent: the first
# checks for updates, and the second cannot run offline because it has to
# resolve a tool.

# The fixed cost every invocation pays: process start, dynamic linking, and
# building the clap command tree. Reading the others against this separates
# "mise got slower to start" from "this command's path got slower".
[bench.startup]
cmd = ["./target/release/mise", "--help"]

# Config resolution: parse the fixture, resolve three tools against the local
# store, and report. This is the path a shell pays on every prompt.
[bench.ls]
cmd = ["./target/release/mise", "-C", "benchmarks/perf-fixture", "ls"]

# The same resolution, rendered as an environment. `env` and `hook-env` diverge
# in how they serialise, so a change in one and not the other is informative.
[bench.env]
cmd = ["./target/release/mise", "-C", "benchmarks/perf-fixture", "env"]

# What `mise activate` runs on every directory change, which makes it the
# command most sensitive to a startup regression.
[bench.hook-env]
cmd = ["./target/release/mise", "-C", "benchmarks/perf-fixture", "hook-env"]

# The compiled-in tool registry — 983 tools, ~75KB of output — walked and
# rendered. The largest pure-CPU path that needs neither network nor installed
# tools, and it grows every time a tool is added.
[bench.registry]
cmd = ["./target/release/mise", "registry"]
runs = 10