PreToolUse guard for Bash: blocks dangerous shell commands (rm -rf ~, fork bombs, curl|sh, force push to main) before they run.
A PreToolUse hook fires before every Bash call and tests the command against a curated pattern list: rm targeting home (~, $HOME, ${HOME}, /home/<user>, /Users/<user>), root or system directories, the cwd or its parent (..), dd to disk devices (including macOS /dev/disk*), mkfs, fork bombs, curl/wget piped to any shell or run through bash <(curl ...) / sh -c "$(curl ...)", force push to main/master (flag before or after the branch, or a +main refspec), git reset --hard, git clean -f, chmod 777, docker volume deletion, and more. On a match it denies the call and shows the pattern id and reason, so Claude knows exactly why. Three safety levels control how much is blocked, and per-level ask mode turns a hard deny into a prompt you approve or reject. Every block/ask is logged to ~/.claude/hooks-logs/<date>.jsonl.
/plugin marketplace add karanb192/claude-code-hooks # once per machine
/plugin install block-dangerous-commands@claude-code-hooks
Restart Claude Code, done. (Or from a shell: claude plugin install block-dangerous-commands@claude-code-hooks.)
codex plugin marketplace add karanb192/claude-code-hooks
codex plugin add block-dangerous-commands@claude-code-hooksOpen /hooks in Codex and approve this plugin. Re-approve after every update; Codex skips unapproved or changed hooks. Do not install it alongside guard-pack, which includes this guard.
This entry point checks Bash and native apply_patch calls. Patch files are checked individually; unreadable patches are blocked. Shell patches must use a plain command with a quoted heredoc. Literal searches and commit messages mentioning apply_patch pass. See the shared adapter's limits, including shell sessions and dynamically constructed commands.
Export configuration variables before launching Codex. An ask-mode match becomes a block because Codex hooks cannot ask. Logs stay in ~/.claude/hooks-logs/ with host: "codex", including on Codex-only machines.
| Event | Runs | What happens |
|---|---|---|
PreToolUse (Bash) |
sync (must finish before the command runs) | Tests the command against every pattern at or below the configured safety level; on a match returns permissionDecision: "deny" (or "ask" in ask mode) with the pattern id and reason, and logs the event. Safe commands pass through untouched. |
| Level | What's blocked | Use case |
|---|---|---|
critical |
Catastrophic only (rm -rf ~, fork bombs, dd to disk) | Maximum flexibility |
high |
+ Risky (force push main, git reset --hard, curl|sh) | Recommended, default |
strict |
+ Cautionary (any force push, sudo rm, docker prune) | Maximum safety |
Set the level via HOOK_SAFETY_LEVEL (see below). Don't edit the installed script: plugin updates overwrite it.
All optional, set via environment variables:
| Variable | Default | Purpose |
|---|---|---|
HOOK_SAFETY_LEVEL |
high |
Safety level: critical, high, or strict. Invalid values fall back to high. |
HOOK_ASK_CRITICAL |
unset (deny) | The literal string true makes critical-level matches prompt instead of deny. |
HOOK_ASK_HIGH |
unset (deny) | Same, for high-level matches. |
HOOK_ASK_STRICT |
unset (deny) | Same, for strict-level matches. |
HOOK_FAIL_OPEN |
unset (deny) | The literal string true lets the call through when the guard's own check throws, instead of denying it. See When the guard fails. |
Ask mode is strictly opt-in and per level: only the literal string true enables it ("1" does not), and enabling it for one level never softens another. A common setup: keep critical on deny, set HOOK_ASK_STRICT=true so cautionary patterns prompt instead of blocking.
Pair it with protect-secrets for coverage of secrets exposure (cat .env, printenv, ~/.ssh reads): that lives in a separate hook by design.
From 1.2.0, a guard that cannot finish its check refuses the call instead of allowing it. A check that throws is denied with a [guard-error] reason. A timeout (10 s) or a missing node is blocked by "onFailure": "block" on Claude Code v2.1.295 or later; older versions ignore that key and let the call run. On Codex a missing node is blocked by the command's || exit 2, but a timeout still lets the call run, because Codex has no setting for it. Set HOOK_FAIL_OPEN=true to let the call through when the check throws; When a guard fails covers the rest of the way back.
Blocked/asked commands are logged (with pattern id, level, decision, and session metadata) to ~/.claude/hooks-logs/<date>.jsonl. It makes no network calls (the script only uses fs and path), so everything stays on your local machine.
/plugin uninstall block-dangerous-commands@claude-code-hooks
Part of claude-code-hooks.