BrotherEye.

The banner stopped shelling out to Python~110ms

The login glitch effect ran a small inline Python script. It fires twice on every login shell, and almost all of the cost was interpreter startup rather than the animation itself.

Rewritten in pure zsh: the same six-step reveal, none of the launch overhead. Two calls per login shell, so roughly a fifth of a second back on every new tab.

The opening animation was starting a second language to draw itselfabout 110ms

Almost all of its cost was starting that language, not animating. Rewritten to use only what the terminal has, it gives a fifth of a second back per tab.

compinit trusts its cache for a day24h

compinit re-scans fpath and security-checks every directory on each startup, which costs well over 100ms on every single shell.

-C reuses the cached dump and skips the check. A glob qualifier forces one full rebuild a day, so completions for newly installed formulae still appear without paying the cost every time.

The tab-completion cache is trusted for a dayonce a day

The machinery behind pressing Tab rechecked every folder it knows on every startup. It now trusts yesterday’s answer and rebuilds daily.

if [[ -n ${_be_zcompdump}(#qN.mh+24) ]]; then
  compinit -d "$_be_zcompdump"      # older than 24h — full rebuild
else
  compinit -C -d "$_be_zcompdump"   # fresh — trust the cache
fi

The banner greets you once, not everywhere-o login

Gated on a TTY alone, the figlet + fastfetch + glitch sequence fired in every subshell, every nested zsh, every split pane, at roughly 0.6s each time.

Gating on a login shell keeps it to the places you actually open a terminal. Terminal.app and iTerm2 start each new tab as a login shell, so it still greets you exactly where you expect.

The greeting was firing everywhere, not just where you open a terminal

It was running in every nested shell and split pane, at over half a second each. It now fires only where you open a window.

Your own aliases ambush your functions

ls is aliased to eza, ping to gping, du to ncdu, cat to bat. An alias is expanded when a function is PARSED, not when it runs, so every one of those aliases silently rewrites the internals of any function that uses the bare command.

command ls, command ping, command du, command cat wherever the real tool is the one that matters. Forensic metadata should be genuine stat output, not a themed listing with icons and git columns.

Its own shortcuts were ambushing its own commands

The file swaps several standard commands for nicer ones, and those swaps were silently rewriting the insides of its own tools.

Re-sourcing over a stale alias aborts the parse

weather, vsim and ip all started life as aliases and became functions. Sourcing the new file into a shell that still holds the old alias fails outright with "defining function based on alias", which breaks the rest of the file rather than just that line.

unalias <name> 2>/dev/null immediately before each definition, so reload is always safe.

Reloading the file could abort halfway through

Three shortcuts grew into real functions. Reloading into a terminal still holding the old one failed outright, and took the rest of the file with it.

unalias vsim 2>/dev/null
vsim() {
  [[ $# -eq 0 ]] && { echo "Usage: vsim <file.v> [more.v...]"; return 1 }
  iverilog -o sim.out "$@" && vvp sim.out
}

An alias appends arguments at the end

As an alias, vsim adder.v would expand to iverilog -o sim.out && vvp sim.out adder.v, handing the input file to the simulator instead of the compiler, and leaving iverilog with nothing to compile.

It has to be a function so "$@" lands in the middle of the pipeline where the compiler expects it.

A shortcut puts your arguments in the wrong place

A shortcut can only bolt what you type onto the end. This one needed it in the middle, so it had to become a real function.

A lexical sort hides the busiest process on the machine

Sorting ps output by CPU without -n compares "9.5" against "36.2" as text and ranks the wrong one first. Leaving the header in before sorting lets it resurface as a fake row mid-table.

tail -n +2 drops the header before sorting, and -rnk3 compares numerically.

Sorting numbers as text ranks the wrong one first

The command listing the busiest programs compared its numbers character by character, which puts 9.5 above 36.2.

ps aux | tail -n +2 | sort -rnk3 | head -15

read folds consecutive tabs together

Tab is an IFS whitespace character, so reading a tab-separated row with read -r collapses empty fields. A device with no label would shift its STATUS flag left into the NAME column.

Split with a quoted array expansion instead, ${(@ps:\t:)line}, which preserves empty fields.

Reading a table collapsed its empty columns

The ordinary way of reading tab-separated rows treats several tabs as one, so a device with no label had its columns shift left.

macOS arp drops leading zeros from MAC octets

arp prints c4:79:5:43:77:13, not c4:79:05:43:77:13. A raw prefix taken from that never matches the IEEE OUI table, so every vendor lookup came back empty.

Re-pad every octet to two hex digits and upper-case it before the lookup.

Hardware addresses arrive missing their leading zeros

The machine reports them shortened. The manufacturer lookup expects the full form, so every lookup came back empty.

No git integration, deliberately

Prompts that show branch and dirty state run git status before every single prompt, which is exactly where a large repository makes the shell feel slow.

The prompt costs nothing inside a repo, however large. sitrep reports branch and dirty count on demand instead.

No repository information in the prompt, deliberately

Prompts that show your branch interrogate the repository before drawing every prompt, which is what makes a terminal feel slow in a large project. This one reports it on request.

The reveal animates as one block, not one line

A single carriage return only rewinds the last visual line, so a MISSION banner long enough to wrap animated its tail while its head sat frozen. That is what forced an earlier 60-character truncation.

Draw every line, then walk the cursor back up with an explicit escape before the next frame, so a wrapped banner is shown whole.

The reveal animates as a block, not a line

Redrawing an animating line only rewinds the last line on screen, so a banner long enough to wrap animated its tail while its head sat frozen.

Syntax highlighting must be sourced last

zsh-syntax-highlighting wraps every ZLE widget that exists at the moment it is sourced. Anything binding widgets afterwards goes un-highlighted.

It is the final statement in the file, with a comment saying so. New integrations go above it, never below.

The syntax colouring has to load last

The component that colours your typing wraps whatever exists when it loads, so anything added after is never coloured. It is the last line of the file.

History depth is a feature, not housekeeping

zsh defaults to 1000 entries. Autosuggestions draw purely from history, so a shallow history directly weakens every suggestion the shell can offer.

100,000 entries, shared across tabs, deduplicated. Every entry lost is a suggestion that never appears, and it costs a few megabytes.

How much history it keeps is a feature100,000 entries

The suggestions the terminal offers come entirely from past commands, so keeping only the default thousand weakens every one of them.