| name | shell-safety-review |
| description | Sanity-check a shell command for destructive or irreversible side effects before running it, especially in an autonomous/unattended run. |
Before executing a shell command that was generated (by yourself or another
model) rather than typed by a human in the moment, scan it for these
red flags and stop to confirm with the user if any are present:
- Recursive/forced deletion:
rm -rf, rm -f *, find ... -delete,
emptying a trash/cache directory that wasn't explicitly named by the user.
- Overwrite-in-place without a check:
> file (truncates silently),
mv over an existing path, git checkout -- . / git reset --hard /
git clean -fd (discards uncommitted work) — verify there's nothing
uncommitted worth keeping first (git status, git stash if unsure).
- Wide-scope operations: a glob or wildcard that could match more than
the intended target (
chmod -R, chown -R, rm dir/* from the wrong
cwd) — print pwd and ls the target first if there's any ambiguity.
- Credentials/secrets in plain sight: a command that echoes, logs, or
commits an API key, token, or password — redact or route through a secret
manager instead.
- Network calls to non-obvious hosts:
curl/wget piped straight into
sh/bash (curl ... | sh), or a request to a host the user never
mentioned — read the script first instead of blind-piping.
- Force-push or history rewrite:
git push --force (especially to
main/master), git rebase on a shared branch, --no-verify to skip
hooks — these need explicit human sign-off, not a "looked right" pass.
- Background/detached processes that outlive the session:
nohup ... &,
disown — confirm the user wants a long-running process left behind.
If none of these apply, proceed normally. If one does, don't silently
"fix" the command yourself — explain the specific risk in plain language
and ask whether to proceed, adjust the scope, or skip it.
