A subshell in the log line zeroed $? before the Claude Code hook could return it

That line fails in two separate ways, and the log is the lesser of the two. A shell script exits with the status of the last command it ran, and in the affected wrapper the last command is the echo, which succeeded ^[6][18]. So the wrapper hands its caller a 0 whatever `node` did, which is the symptom the write-up reports: failures happening, exit codes never arriving at the caller ^[19].

That 0 travels through several layers before it means anything to anyone. Claude Code fires the hook event into a JS dispatcher named in `settings.json`, which spawns `run-with-flags-shell.sh`, which pipes stdin onward and calls `check-hook-enabled.js`, which finally runs the hook logic; the exit code then has to climb back up those same four processes ^[13][17]. Claude blocks a tool call or a Stop only when it sees exit 1 ^[5]. There are four processes in that chain, and each one is a chance to launder a non-zero into a zero. The author documents two more ways to do it: a JS wrapper calling the inner script through `child_process.exec()` or `$()` receives 0 when that script returns 1 ^[14], and a piped invocation without `set -o pipefail` reports only the right-hand side of the pipe ^[15].

The base rate is low. Three hits in 328 scripts is 0.91%, roughly one script in 109 ^[10][16]. That is a comfortable number until you read which three: the Claude Code hook path, a note paid-bonus ZIP attachment script, and a dotfiles snapshot ^[10]. Two of the three sit under automation whose output nobody reads, so a silent bug there can sit for a long time without anyone noticing.

For this to transfer to your setup, one thing has to be true: somewhere you read `$?` after another command has run in between, including a command hidden inside a string you are building for a log ^[7]. Capture the status into a variable on the line immediately after the command and this particular trap cannot fire. The wrapper and pipeline variants need a wrapper or a pipe to exist at all, so if `settings.json` invokes your hook script directly, those two do not apply to you ^[14][15].

What does transfer is the test. The reproduction injected exit 4, watched the line print 0, applied the fix, and watched it print 4 ^[8]. That is a minute of work, and it is the only test that separates a hook which blocks from a hook which prints. Everything else the author had was a green dashboard, and a dashboard only reflects the code that produced it. With 171 launchd jobs and several Claude Code sessions running at once, operators end up checking the colour of the dashboard rather than the individual rows ^[12].