TOFIX

Findings from a code scan on 2026-10-04.

Medium

src/plugins/minikube.sh:10 - the if ! cmd; then __var=$? pattern stores the status of the negated test, which is always 0 inside the then branch (if ! false; then echo $?; fi prints 0), so a failed completion/source is recorded as success and bashy_errors (src/bashy.sh:271, result = 0 means ok) never shows it. The same bug is in bash_completions_system.sh:9,18, bash_it.sh:56, eksctl.sh:14, fzf.sh:13, google_cloud_sdk.sh:10,17, kurtosis.sh:9, ng.sh:8, nvm.sh:11,18, oc.sh:10, powerline.sh:18, pureline.sh:14, python_venv.sh:15, rvm.sh:15, starship.sh:10, system_default_bashrc.sh:10 and virtualenvwrapper.sh:33; set __var=1 as bash_completions_prog.sh:11 does, and add a test for a failing activation.
src/plugins/rvm.sh:16 - assigns the message to _error (single underscore), a stray global, instead of the __error nameref, so the reason is lost; use __error.

src/plugins/rvm.sh:10 - exports a hardcoded RUBY_VERSION="2.3.3" (a long-EOL Ruby) on every activation with no comment, against the "never hardcode a version" rule in CLAUDE.md:20; drop it and let rvm's own default/.ruby-version decide, or document why it is pinned.
src/plugins/eksctl.sh:15 - error message says eskctl; fix the typo.
src/plugins/powerline-shell.sh:14 - uses the external which (also pypowerline.sh:16), forking a process and depending on a non-POSIX tool; use the command -v builtin instead.
doc/design.txt:28 - says "An installer is a plain bash function with no parameters", but every installer takes --force through bashy_install_args "$@" (src/core/install.sh:68, README "Writing an installer"); update the sentence.