Shell Performance¶

Shell startup time was reduced from ~2.37s to ~0.08s — a 97% improvement — through a series of targeted changes.

What changed¶

Replaced chruby + nvm with mise — the biggest win. The previous setup sourced two chruby scripts, ran brew --prefix at startup, and used lazy-loader stub functions for nvm/node/npm/npx to avoid nvm's ~500ms cold start. All of that is now one line: eval "$(mise activate zsh)", which adds ~5ms and handles Ruby, Node, Python, Java, and Go with automatic per-project version switching.

Removed dynamic PATH calls — a $(ruby -e 'puts Gem.bindir') subshell that spawned a full Ruby process on every new shell has been removed.

Starship Ruby module disabled — starship's Ruby module executed Ruby on every prompt render, causing intermittent timeout warnings. Disabled in config/starship.toml; command_timeout = 2000 is set as a safety net for other modules.

Cached compinit — zsh rebuilds its completion dump (~/.zcompdump) on every shell start by default. A 24-hour freshness check now skips the rebuild (compinit -C) unless the dump is older than a day. Saves ~30–50ms per shell start with no visible downside.

Benchmark (MacBook Pro, Apple Silicon)¶

Measurement	Time
Original (chruby + nvm + Starship Ruby warnings)	~2.37s
After lazy NVM + removing Ruby PATH call	~0.58s
After replacing chruby + nvm with mise	~0.11s
After compinit caching	~0.08s
Total improvement	~97% faster

Optimization methodology¶

Every change above came from the same loop. Use it whenever startup feels slow again (a new plugin, a heavy eval in ~/.zshrc.local, etc.):

Measure a baseline — get a repeatable number before touching anything (see Measuring startup time). Always compare against the same machine and an otherwise-idle system.
Profile to find the cost — don't guess. Use zprof (below) to rank what actually consumes time during an interactive start.
Attack the biggest cost first, using one of three levers in order of preference:
- Eliminate — remove work that runs on every shell. The single biggest win here was deleting a $(ruby -e 'puts Gem.bindir') subshell that spawned a full Ruby process per shell.
- Replace — swap a slow tool for a faster one. chruby + nvm (two sourced scripts, a brew --prefix call, and nvm lazy-loaders) became one line: eval "$(mise activate zsh)".
- Cache / defer — pay a cost once and reuse it, or push it to first use. compinit -C skips rebuilding ~/.zcompdump unless it is older than a day; disabling starship's Ruby module avoids running Ruby on every prompt.
Re-measure — confirm the change helped and didn't regress correctness (completions still work, runtimes still switch per project).
Repeat until the next-biggest cost isn't worth chasing.

Keep machine-specific weight out of the tracked rc

Slow, machine-only setup (corporate tooling, one-off evals) belongs in ~/.zshrc.local, not the tracked home/zshrc. That keeps the shared baseline fast for every profile and machine.

Measuring startup time¶

Time a full interactive start, averaged over several runs:

# Average of 5 runs
for i in $(seq 1 5); do /usr/bin/time zsh -i -c exit 2>&1; done

For a tighter, statistically warmed-up number, use hyperfine (brew install hyperfine):

hyperfine --warmup 3 'zsh -i -c exit'

Profiling with `zprof`¶

Timing tells you how slow; zprof (a zsh builtin module) tells you what is slow. Enable it at the very top of your rc, start a shell, then read the report:

# Temporarily add as the FIRST line of ~/.zshrc (or ~/.zshrc.local):
zmodload zsh/zprof
# ...rest of your rc...

# Then open a new shell and run:
zprof   # prints functions ranked by cumulative time

The top entries are where to focus step 3 above. Remove the zmodload line once you're done. For a line-by-line view of what runs at startup, you can also trace it:

zsh -i -x -c exit 2>&1 | less   # every command sourced during an interactive start

The runtime manager (mise) and plugin manager (sheldon) are the two heaviest startup contributors by design — see the Glossary for what each does and the Architecture page for how they're wired in.