On Windows, each program a Claude Code hook starts costs 40 to 200 ms

A hook on Edit and Write runs before every edit Claude makes, and Claude waits for it. On Windows that hook runs in Git Bash, and in Git Bash starting a program is slow: on AO's PC each one cost 43 to 203 ms, depending on how busy the PC was. The same program on Linux, on the same PC, cost 1.4 ms. Twenty small programs in one bash took 0.9 to 6.6 s on Windows and 28 to 38 ms on Linux, so a hook that feels free when you write it on a Mac or Linux can make every edit slow on Windows.

What it did to AO's own hook. AO's AIsle plugin has a hook on every edit, in every project on the computer, though it only has work to do in folders that joined a room. Version 0.3.4 found its version, its folder and a hash of the path first, with about twenty programs (grep, cut, tr, sed, sha256sum) before it checked whether the folder had joined at all. Measured on 2026-10-08, the way Claude Code starts it:

• 0.3.4 on Windows: median 3.0 to 5.3 s per edit, up to 9.7 s, at 63 to 87% CPU.
• The current version on Windows: median 0.09 to 0.17 s.
• 0.3.4 on Linux in WSL: 60 to 80 ms. The current version: 11 to 21 ms.

The plugin gives the hook a 5 s limit, so on a busy Windows PC the old version could run out of time on an ordinary edit. On Linux nobody would have noticed.

The fix is bash builtins on the common path. The current version reads the hook's JSON with read, finds the fields with [[ =~ ]] and BASH_REMATCH, reads its list of folders with read -d '', and leaves at once when the folder is not on it. Only then does it start any program. These swaps were measured in Git Bash on the same PC:

x=$(date +%s)          277 ms      x=$EPOCHSECONDS       0.8 ms
x=$(basename "$p")     264 ms      x=${p##*/}            1.3 ms
x=$(cat "$f")           65 ms      x=$(<"$f")            1.0 ms
x=$(printf '%s-%s' a b) 34 ms     printf -v x '%s-%s' a b  0.7 ms

A $(...) costs time even with only a builtin inside, because bash starts a copy of itself to run it: 26 to 81 ms in Git Bash, 1 ms on Linux. $(<file) was the exception in bash 5.3: 1 ms.

Shell form starts bash twice. A hook written as "command": "bash script.sh" goes to Git Bash, which then starts a second bash for the script. With "args" set, Claude Code starts the program directly. In our runs that saved a median of about 70 ms per call on Windows. One caution we found: on this PC, Windows' own search of the PATH finds WSL's bash.exe in C:\Windows\System32 before Git's. We did not test which one exec form starts, so give the full path to Git's bash.exe if you use it.

Node.js is not a way out. An empty node -e "" took a median of 77 ms in the first run and 392 ms when the PC was busy, against 53 ms for bash in both.

How to measure your own machine (in Git Bash, then in WSL or on Linux for comparison):

n=50; t0=$EPOCHREALTIME; for ((i=0; i<n; i++)); do /usr/bin/true; done; echo "$(( (${EPOCHREALTIME/./} - ${t0/./}) / n / 1000 )) ms per program"

Not measured. Other Windows PCs and Git for Windows versions, a PC with no other load, and the effect of antivirus scanning on each program start. The busy runs shared the PC with other work, so the higher numbers say how bad it gets, not what it always is.