omamori
Pre-1.0 — Breaking changes may occur between minor versions.
AI Agent's Omamori — protect your system from dangerous commands executed via AI CLI tools.
Supported AI CLI Tools
| Tool | Detection | Status |
|---|---|---|
| Claude Code | CLAUDECODE=1 |
Verified |
| Codex CLI | CODEX_CI=1 |
Verified |
| Cursor | CURSOR_AGENT=1 |
Provisional |
| Gemini CLI | GEMINI_CLI=1 |
Provisional |
| Cline | CLINE_ACTIVE=true |
Provisional |
Any tool setting AI_GUARD=1 |
AI_GUARD=1 |
Fallback |
Detection uses exact value matching (e.g. CLAUDECODE=1 only, not CLAUDECODE=true). Tools that set these environment variables when executing shell commands are automatically detected.
What It Does
When an AI CLI tool (Claude Code, Codex, Cursor, etc.) runs a shell command, omamori intercepts dangerous operations and replaces them with safe alternatives. Terminal direct execution is not affected.
[AI CLI Tool] → CLAUDECODE=1 → rm -rf target/
↓
[omamori shim]
↓
moved to Trash instead
[Terminal] → → rm -rf target/
↓
[/usr/bin/rm]
↓
deleted normally
Quick Start
Install via Homebrew (macOS)
Or build from source
Setup
# 1. Install shims + hooks + config (all in one command)
# 2. Add shim directory to PATH (add to .zshrc / .bashrc)
That's it. install --hooks auto-generates config.toml, runs verification, and shows a checklist:
omamori setup complete:
Shims:
[done] rm, git, chmod, find, rsync
Hooks:
[done] Claude Code hook script
[done] Cursor hook snippet
Config:
[done] Created: ~/.config/omamori/config.toml
[done] 7 rules verified, 10 detection tests passed
Next steps:
[todo] Add to your shell profile:
export PATH="$HOME/.omamori/shim:$PATH"
[todo] Merge Cursor hook into .cursor/hooks.json
How It Works
Layer 1 — PATH shim: Symlinks for rm, git, chmod, find, rsync point to the omamori binary. When invoked, omamori checks for AI tool environment variables (e.g. CLAUDECODE=1) and applies rules only if an AI tool is detected.
Layer 2 — Hooks (optional):
- Claude Code: A
PreToolUsehook script catches bypass attempts like/bin/rmdirect paths,unset CLAUDECODE, and warns on interpreter commands (python -c "shutil.rmtree(...)"). - Cursor: A Rust-native
beforeShellExecutionhandler (omamori cursor-hook) provides the same protection via Cursor's hook protocol.
Default Rules
| Command | Pattern | Action |
|---|---|---|
rm |
-r, -rf, -fr, --recursive |
trash — move to macOS Trash |
git |
reset --hard |
stash-then-exec — git stash first, then execute |
git |
push --force, push -f |
block |
git |
clean -fd, clean -fdx |
block |
chmod |
777 |
block |
find |
-delete, --delete |
block |
rsync |
--delete and 7 variants |
block |
Combined short flags are normalized: rm -rfv expands to match -r and -rf rules. The POSIX -- separator is respected for target extraction.
Configuration (v0.2+)
Built-in rules are always inherited. Config is auto-created by install --hooks. To regenerate manually:
Disable a rule via CLI (v0.3+):
Or edit config.toml directly:
[[]]
= "git-push-force-block"
= false
Move files to a custom directory instead of Trash:
[[]]
= "rm-to-backup"
= "rm"
= "move-to"
= "/tmp/omamori-quarantine/"
= ["-r", "-rf", "-fr", "--recursive"]
= "omamori moved targets to quarantine instead of deleting"
Override an existing rule's action:
[[]]
= "rm-recursive-to-trash"
= "move-to"
= "/tmp/omamori-quarantine/"
After editing, run omamori test to verify. Disabled rules show as SKIP:
Rules:
PASS rm-recursive-to-trash rm -r|-rf|-fr|--recursive -> trash
SKIP git-push-force-block (disabled by user config)
...
Summary: 5 rules (4 active, 1 disabled), 4 detection tests passed
Configuration notes
- Config file requires
chmod 600(permissions check enforced) - Only write rules you want to change — everything else is inherited
destinationmust be an absolute path on the same volume- System directories (
/usr,/etc,/System,/Library,/bin,/sbin,/var,/private) are blocked as destinations - Symlinks are rejected as destinations
destinationdirectory must exist before use (omamori will not create it)
Available Actions
| Action | Behavior |
|---|---|
trash |
Move targets to macOS Trash |
move-to |
Move targets to a user-specified directory (requires destination) |
stash-then-exec |
Run git stash first, then execute the original command |
block |
Refuse to execute |
log-only |
Log the event, then execute normally |
Safe Defaults
| Scenario | Behavior |
|---|---|
| No AI env var detected | Pass through to real command (no interference) |
| Config file missing | Fail-close: built-in default rules apply |
| Config file broken | Fail-close: built-in default rules apply + warning |
| Trash / move-to fails | Fail-close: refuse to run the original command |
sudo detected |
Block the command |
| Blocked destination | Fail-close: rule is disabled at config load time |
| Shim binary crashes | Fail-open: real command runs |
CLI
omamori test [--config PATH] # Verify policy rules
omamori exec [--config PATH] -- <command> [args...] # Run through policy engine
omamori install [--base-dir PATH] [--hooks] # Create shims + hooks + config
omamori uninstall [--base-dir PATH] # Remove shims + hook files
omamori init [--force] [--stdout] # Create/reset config file
omamori config list # Show all rules with status
omamori config disable <rule> # Disable a built-in rule
omamori config enable <rule> # Re-enable a disabled rule
omamori cursor-hook # Cursor beforeShellExecution handler
Structural Limitations
These are inherent to the PATH shim approach and documented honestly:
- Full-path execution (
/bin/rm,/usr/bin/git) bypasses the shim — mitigated by Layer 2 hooks (Claude Code + Cursor) sudochanges PATH before the shim runs — omamori blocks when it detects elevated execution in-process- Interpreter commands (
python -c "shutil.rmtree(...)") — Layer 2 hooks warn on known destructive patterns, but obfuscated code (base64, heredoc, variable indirection) cannot be detected find -exec /bin/rmbypasses the find shim because rm is invoked via absolute path — partially mitigated by Layer 2 hooks- Cross-device moves are not supported for
move-to(use a destination on the same volume)
For the full security model, see SECURITY.md.
Related
- nanika — explains what AI commands will do (detect + translate). Complementary to omamori (detect + replace).
License
Licensed under either of Apache License, Version 2.0 or MIT license at your option.