Skip to content

Claude Code Critical-Path rm Check for Unattended Runs

Claude Code asks before an rm or rmdir aims at a critical path, and denies it where nobody can answer.

Claude Code treats the filesystem root, your home directory and your working directory as critical paths. An rm or rmdir that targets one gets a prompt or a denial, and no allow rule or PreToolUse hook that returns "allow" approves it (permission modes). Release 2.1.330 (November 20, 2026) "Changed more rm commands aimed at the root, home, a top-level folder or the working folder to ask for approval; where nobody can answer, they're refused" (changelog). The entry does not list the added commands, so this page follows the current permission modes reference.

What counts as a critical path

Claude Code treats a target as critical when it is the filesystem root, a direct child of the root such as /usr or /etc, your home directory, or your working directory and its parents. Your additional working directories count only for a glob under them, such as rm -rf <dir>/* (permission modes).

The check also flags targets it cannot resolve in advance (permission modes):

  • A glob or trailing slash under a shell variable, such as rm -rf "$DIR"/*. An empty variable turns it into a removal from the root.
  • A recursive rm whose target is only command substitution output, such as rm -rf "$(pwd)".

What happens when nobody can answer

The outcome depends on the permission mode (permission modes).

Mode Outcome
default, acceptEdits Asks you to approve it
plan Asks you to approve it. If the classifier reviews commands during planning and no bypass permissions are available, it follows auto
auto Asks in the terminal with a time limit. Elsewhere, denies it
dontAsk Denies it
bypassPermissions Asks in the terminal before a time limit runs out. Elsewhere, see below

For a -p run or an Agent SDK session, auto denies the removal. In bypassPermissions, your --permission-prompt-tool or canUseTool callback decides. With neither, or under --permission-prompts none, Claude Code denies it unless a PermissionRequest hook allows it (permission modes).

In the terminal, the auto and bypassPermissions prompt counts down two minutes, then denies the command. After three unanswered prompts, Claude Code denies further removals at once. This applies from v2.1.281 (permission modes).

How it layers with rules and hooks

  • An explicit ask rule that matches the command asks you even in auto mode, with no time limit (permission modes).
  • A deny rule that matches the command blocks it, though a mod you installed can still approve some of those calls. A deny rule matches command text, so Bash(rm *) stops rm -rf build/ but not /bin/rm -rf build/ (permissions).
  • A PermissionRequest hook can answer the prompt in modes that ask (permission modes).

What the check does not cover

Anthropic states the limit: "The check can miss some commands, so it doesn't replace an operating-system control such as sandboxing" (permission modes). Three gaps follow:

  • A command the check misses gets no special handling, so an allow rule, bypassPermissions or the sandbox's auto-allow mode can approve it with no prompt.
  • The check applies to rm and rmdir. Other approved commands can delete critical paths too.
  • Sandboxed commands can still delete your working directory.

Four environment variables weaken the check. CLAUDE_CODE_EVEN_MORE_DANGEROUSLY_SKIP_PERMISSIONS=1 (v2.1.311 or later) runs these removals with no prompt in bypassPermissions mode (permission modes). CLAUDE_CODE_DISABLE_DANGEROUS_RM_TIMEOUT=1 hands auto mode's decision to the classifier in -p runs, instead of the denial (permission modes). CLAUDE_CODE_DISABLE_SUBSTITUTION_RM_PROMPT=1 turns off the check on a target that is only command substitution output (permission modes). CLAUDE_CODE_DISABLE_INLINE_SHELL_RM_PROMPT=1 turns it off on a critical path typed directly in a -c script, such as ~ (permission modes).

Rewrite a flagged command

Since v2.1.277 the permission prompt names the flagged rm and suggests a ${VAR:?} guard, "so headless runs can recover" (changelog). The reference gives four rewrites (permission modes):

  • A glob or trailing slash under $DIR: guard the expansion, as in rm -rf "${DIR:?}"/*, or use a literal path.
  • A glob or trailing slash under a variable that is normally set, such as $HOME: use a literal path.
  • A variable assigned from a directory-printing substitution, such as D=$(pwd); rm -rf "$D": use a literal path. A "${D:?}" guard does not clear this check, because the variable is not empty.
  • A target that is only command substitution output: run the substitution first, then remove the literal paths it prints.

Test the guard before an unattended run

Use a throwaway directory or container, never a checkout you care about. Confirm that claude --version reports 2.1.330 or later (changelog). Then, with CLAUDE_CODE_DISABLE_SUBSTITUTION_RM_PROMPT and CLAUDE_CODE_DISABLE_DANGEROUS_RM_TIMEOUT unset, run a -p session in auto mode with --output-format stream-json and ask Claude to run rm -rf "$(pwd)", an example from the reference. The table above predicts a denial. Denials appear as permission_denied system messages, and the final result message lists them in permission_denials (headless guide). The guide does not say whether a critical-path denial appears there, so read the whole output. If Claude declines before it tries the command, the run says nothing about the check.

Key Takeaways

  • The check covers rm and rmdir only, and it can miss commands.
  • In a -p run, auto denies the removal. bypassPermissions denies it unless your permission host or a PermissionRequest hook decides.
  • Pair the check with a sandbox. A deny rule matches command text, so /bin/rm slips past Bash(rm *).
  • Do not set CLAUDE_CODE_DISABLE_DANGEROUS_RM_TIMEOUT, CLAUDE_CODE_EVEN_MORE_DANGEROUSLY_SKIP_PERMISSIONS, CLAUDE_CODE_DISABLE_SUBSTITUTION_RM_PROMPT or CLAUDE_CODE_DISABLE_INLINE_SHELL_RM_PROMPT on a scheduled run.