safety-net

          

<div align="center">


</div>

A Coding Agent CLI plugin that acts as a safety net, catching destructive git and filesystem commands before they execute.

Contents

• •Why This Exists
• •Why Use This Instead of Permission Deny Rules?
• •What About Sandboxing?
• •Prerequisites
• •Quick Start

- Codex Installation - Claude Code Installation - Gemini CLI Installation - GitHub Copilot CLI Installation - Kimi CLI Installation - OpenCode Installation - Pi Installation

• •Status Line Integration

- Emoji Mode Indicators

• •Diagnostics
• •Explain (Debug Analysis)
• •Commands Blocked
• •Commands Allowed
• •What Happens When Blocked
• •Testing the Hook
• •Breaking Change: Custom Rules Migration
• •Custom Rules

- Config File Location - Rule Schema - Matching Behavior - Rule Examples - Error Handling

• •Advanced Features

- Strict Mode - Paranoid Mode - Worktree Mode - Shell Wrapper Detection - Interpreter One-Liner Detection - Secret Redaction - Audit Logging

• •Development
• •License

Why This Exists

We learned the hard way that instructions aren't enough to keep AI agents in check. After Claude Code silently wiped out hours of progress with a single rm -rf ~/ or git checkout --, it became evident that soft rules in an CLAUDE.md or AGENTS.md file cannot replace hard technical constraints. The current approach is to use a dedicated hook to programmatically prevent agents from running destructive commands.

Why Use This Instead of Permission Deny Rules?

Claude Code's .claude/settings.json supports deny rules with wildcard matching (e.g., Bash(git reset --hard:*)). Here's how this plugin differs:

At a Glance

Permission Rules Have Known Bypass Vectors

Even with wildcard matching, Bash permission patterns are intentionally limited and can be bypassed in many ways:

CC Safety Net Handles What Patterns Can't

Defense in Depth

PreToolUse hooks run **before** the permission system. This means CC Safety Net inspects every command first, regardless of your permission configuration. Even if you misconfigure deny rules, CC Safety Net provides a fallback layer of protection.

Use both together: Permission deny rules for quick, user-configurable blocks; CC Safety Net for robust, bypass-resistant protection that works out of the box.

What About Sandboxing?

Claude Code offers native sandboxing that provides OS-level filesystem and network isolation. Here's how it compares to CC Safety Net:

Different Layers of Protection

Why Sandboxing Isn't Enough

Sandboxing restricts filesystem + network access, but it doesn't understand whether an operation is destructive within those boundaries. These commands are not blocked by the sandbox boundary:

[!NOTE]

Whether they're auto-run or require confirmation depends on your sandbox mode (auto-allow vs regular permissions), and network access still depends on your allowed-domain policy. Claude Code can also retry a command outside the sandbox via dangerouslyDisableSandbox (with user permission); this can be disabled with allowUnsandboxedCommands: false.

Sandboxing sees git reset --hard as a safe operation—it only modifies files within the current directory. But you just lost all uncommitted work.

When to Use Sandboxing Instead

Sandboxing is the better choice when your primary concern is:

• •Prompt injection attacks — Reduces exfiltration risk by restricting outbound domains (depends on your allowed-domain policy)
• •Malicious dependencies — Limits filesystem writes and network access by default (subject to your sandbox configuration)
• •Untrusted code execution — OS-level containment is stronger than pattern matching
• •Network control — CC Safety Net has no network protection

Recommended: Use Both

They protect against different threats:

• •Sandboxing contains blast radius — even if something goes wrong, damage is limited to cwd and approved network domains
• •CC Safety Net prevents footguns — catches git-specific mistakes that are technically "safe" from the sandbox's perspective

Running both together provides defense-in-depth. Sandboxing handles unknown threats; CC Safety Net handles known destructive patterns that sandboxing permits.

Prerequisites

• •Node.js: Version 18 or higher is required to run this plugin

Quick Start

Codex Installation

1. Enable Codex plugin hooks in ~/.codex/config.toml:

``toml [features] plugin_hooks = true ``

1. Add the marketplace:

``bash codex plugin marketplace add kenryu42/cc-marketplace ``

1. Start Codex.
2. In the TUI, run /plugins.
3. Use arrow keys to select [cc-marketplace].
4. Press Enter to install the plugin.
5. run /hooks and select the safety-net PreToolUse hook and press t to trust it.

Claude Code Installation

/plugin marketplace add kenryu42/cc-marketplace
/plugin install safety-net@cc-marketplace
/reload-plugins

Claude Code Auto-Update

1. Run /plugin → Select Marketplaces → Choose cc-marketplace → Enable auto-update

Gemini CLI Installation

gemini extensions install https://github.com/kenryu42/gemini-safety-net

GitHub Copilot CLI Installation

/plugin install kenryu42/copilot-safety-net

Kimi CLI Installation

Install CC Safety Net into your Kimi CLI config:

npx -y cc-safety-net hook install --kimi-cli

OpenCode Installation

Install CC Safety Net with OpenCode's native plugin command:

opencode plugin -g cc-safety-net

[!NOTE]

OpenCode can sometimes keep using a stale cached plugin version. See

anomalyco/opencode#25293 for the current tracking issue.

To force OpenCode to reinstall cc-safety-net, remove its cached package and

install the version you want:

```sh

rm -rf ~/.cache/opencode/packages/cc-safety-net@latest

opencode plugin -g -f cc-safety-net@latest

If you prefer pinning a specific version:

rm -rf ~/.cache/opencode/packages/cc-safety-net@latest

opencode plugin -g -f cc-safety-net@<version>

Restart OpenCode after updating so the plugin is loaded from the refreshed

cache.

Pi Installation

Install CC Safety Net with Pi's package installer:

pi install npm:cc-safety-net

Status Line Integration

CC Safety Net can display its status in Claude Code's status line, showing whether protection is active and which modes are enabled.

Add the following to your ~/.claude/settings.json:

Using Bun (recommended):

{
  "statusLine": {
    "type": "command",
    "command": "bunx cc-safety-net statusline --claude-code"
  }
}

Using Claude X:

{
  "statusLine": {
    "type": "command",
    "command": "BUN_BE_BUN=1 claude x cc-safety-net statusline --claude-code"
  }
}

[!NOTE]

The claude x command is only compatible with the native version of Claude Code. If you installed via npm, please use npx or bunx instead.

Using NPM:

{
  "statusLine": {
    "type": "command",
    "command": "npx -y cc-safety-net statusline --claude-code"
  }
}

Piping with existing status line:

If you already have a status line command, you can pipe CC Safety Net at the end:

{
  "statusLine": {
    "type": "command",
    "command": "your-existing-command | bunx cc-safety-net statusline --claude-code"
  }
}

Changes take effect immediately — no restart needed.

Emoji Mode Indicators

The status line displays different emojis based on the current configuration:

Multiple mode emojis are combined when multiple environment variables are set. Mode flags use CC_SAFETY_NET_* names; legacy SAFETY_NET_* names are still accepted.

Diagnostics

Run the diagnostic command to verify your installation and troubleshoot issues:

npx cc-safety-net doctor
# or with bun
bunx cc-safety-net doctor

The doctor command checks:

Options

Explain (Debug Analysis)

Trace how CC Safety Net analyzes a command step-by-step. Useful for debugging why a command is blocked or allowed, or when developing custom rules.

npx cc-safety-net explain "git reset --hard"
# or with bun
bunx cc-safety-net explain "git reset --hard"

Options

Examples

npx cc-safety-net explain "rm -rf /"
npx cc-safety-net explain --json "git checkout -- file.txt"
npx cc-safety-net explain --cwd /tmp "git status"

Commands Blocked

Commands Allowed

What Happens When Blocked

When a destructive command is detected, the plugin blocks the tool execution and provides a reason.

Example output:

BLOCKED by CC Safety Net

Reason: git checkout -- discards uncommitted changes permanently. Use 'git stash' first.

Command: git checkout -- src/main.py

If this operation is truly needed, ask the user for explicit permission and have them run the command manually.

Testing the Hook

You can manually test the hook by attempting to run blocked commands in Claude Code:

# This should be blocked
git checkout -- README.md

# This should be allowed
git checkout -b test-branch

Breaking Change: Custom Rules Migration

[!WARNING]

The custom rules system has moved from legacy inline config files to a rulebook-based layout. Legacy inline config files (.safety-net.json and ~/.cc-safety-net/config.json) are no longer loaded at runtime. If they contain rules, commands now fail closed (stay blocked) until you migrate.

Who Is Affected

• •Affected: users who previously defined custom rules in .safety-net.json (project scope) or ~/.cc-safety-net/config.json (user scope).
• •Not affected: users with no custom rules. All built-in destructive-command protections are unchanged and continue to work out of the box.

What Breaks

"Fail closed" means commands stay blocked until the legacy rules are migrated to the new layout.

How to Migrate

# Convert legacy inline rules into the new rulebook layout
npx -y cc-safety-net rule migrate

# Optionally delete verified legacy files after migration
npx -y cc-safety-net rule migrate --cleanup

# Validate the migrated rules
npx -y cc-safety-net rule verify
npx -y cc-safety-net rule test

Before / After

Before — a single inline config with rules embedded:

.safety-net.json          # project rules (inline)
~/.cc-safety-net/config.json   # user rules (inline)

After — rule migrate creates a rulebook-based layout automatically:

.cc-safety-net/rules/rule.json                    # project rulebook sources + overrides
.cc-safety-net/rules/project-rules/rulebook.json  # migrated project rules
~/.cc-safety-net/rules/rule.json                  # user rulebook sources + overrides
~/.cc-safety-net/rules/user-rules/rulebook.json   # migrated user rules

See Custom Rules for the full authoring guide and Error Handling for fail-closed details.

Custom Rules

Beyond the built-in protections, you can define your own blocking rules to enforce team conventions or project-specific safety policies.

[!TIP]

The best way to create custom rules is to use the /cc-safety-net skill to create custom rules interactively with natural language.

Examples

/cc-safety-net read my package.json and suggest blocking rules
/cc-safety-net set up rules to block all terraform destroy commands
/cc-safety-net verify my rules and fix any errors

[!NOTE]

If your agent does not support skills, prompt it with:

```

run npx -y cc-safety-net rule doc and help me set up custom rules

```

Create Rules Manually

Create a starter project rule config and rulebook:

npx -y cc-safety-net rule init

This creates .cc-safety-net/rules/rule.json:

{
  "version": 1,
  "rules": ["project-rules"],
  "overrides": {}
}

Rule definitions live in .cc-safety-net/rules/project-rules/rulebook.json:

{
  "rulebook_version": 1,
  "name": "project-rules",
  "version": "1.0.0",
  "description": "Project-specific CC Safety Net rules.",
  "author": "project",
  "allowed_commands": ["git"],
  "rules": [
    {
      "name": "block-git-add-all",
      "command": "git",
      "subcommand": "add",
      "block_args": ["-A", "--all", "."],
      "reason": "Use 'git add <specific-files>' instead of blanket add."
    }
  ],
  "tests": [
    {
      "command": "git add -A",
      "expect": "blocked",
      "rule": "block-git-add-all"
    },
    {
      "command": "git add README.md",
      "expect": "allowed"
    }
  ]
}

After editing rulebooks, run:

npx -y cc-safety-net rule sync
npx -y cc-safety-net rule verify
npx -y cc-safety-net rule test

Now git add -A, git add --all, and git add . will be blocked with your custom message.

Config File Location

Config files are loaded from two scopes and merged:

1. User scope: ~/.cc-safety-net/rules/rule.json (use rule init --global)
2. Project scope: .cc-safety-net/rules/rule.json in the current working directory

Local rulebook sources are bare names like project-rules. GitHub rulebook sources use owner/repo#ref/<rulebook-name> and point to .cc-safety-net/rules/<rulebook-name>/rulebook.json in that repository.

Legacy inline config files (.safety-net.json and ~/.cc-safety-net/config.json) are no longer loaded at runtime. Empty legacy files are ignored, but legacy files with rules and invalid legacy files fail closed until migrated or fixed. Convert existing legacy rules with npx -y cc-safety-net rule migrate; use npx -y cc-safety-net rule migrate --cleanup if you also want to delete verified legacy files after migration. See Breaking Change: Custom Rules Migration for the full upgrade guide.

Merging behavior:

• •Rulebooks from both scopes are combined
• •Duplicate active rulebook names are invalid
• •Project overrides win over user overrides for the same <rulebook-name>/<rule-name> key

This allows you to define personal defaults in user scope while letting projects disable or replace reasons for specific rules.

If no config file is found in either location, only built-in rules apply.

Config Schema

Override values are either "off" to disable a rule or { "reason": "..." } to replace the rule reason.

Rulebook Schema

Rule Schema

Fixture Schema

Every rule must have at least one blocked fixture. Add allowed fixtures for close-but-safe commands.

Matching Behavior

• •Commands are normalized to basename (/usr/bin/git → git)
• •Subcommand is the first non-option argument after the command
• •Arguments are matched literally (no regex, no glob), with short option expansion
• •A command is blocked if any argument in block_args is present
• •Short options are expanded: -Ap matches -A (bundled flags are unbundled)
• •Long options use exact match: --all-files does NOT match --all
• •Custom rules only add restrictions—they cannot bypass built-in protections

• •Short option expansion: -Cfoo is treated as -C -f -o -o, not -C foo. Blocking -f may false-positive on attached option values.

Rule Examples

.cc-safety-net/rules/rule.json:

{
  "version": 1,
  "rules": ["project-rules"],
  "overrides": {}
}

.cc-safety-net/rules/project-rules/rulebook.json:

{
  "rulebook_version": 1,
  "name": "project-rules",
  "version": "1.0.0",
  "allowed_commands": ["npm"],
  "rules": [
    {
      "name": "block-npm-global",
      "command": "npm",
      "subcommand": "install",
      "block_args": ["-g", "--global"],
      "reason": "Global npm installs can cause version conflicts. Use npx or local install."
    }
  ],
  "tests": [
    {
      "command": "npm install -g typescript",
      "expect": "blocked",
      "rule": "block-npm-global"
    },
    {
      "command": "npm install typescript",
      "expect": "allowed"
    }
  ]
}

{
  "rulebook_version": 1,
  "name": "project-rules",
  "version": "1.0.0",
  "allowed_commands": ["docker"],
  "rules": [
    {
      "name": "block-docker-system-prune",
      "command": "docker",
      "subcommand": "system",
      "block_args": ["prune"],
      "reason": "docker system prune removes all unused data. Use targeted cleanup instead."
    }
  ],
  "tests": [
    {
      "command": "docker system prune",
      "expect": "blocked",
      "rule": "block-docker-system-prune"
    },
    {
      "command": "docker ps",
      "expect": "allowed"
    }
  ]
}

{
  "rulebook_version": 1,
  "name": "project-rules",
  "version": "1.0.0",
  "allowed_commands": ["git", "npm"],
  "rules": [
    {
      "name": "block-git-add-all",
      "command": "git",
      "subcommand": "add",
      "block_args": ["-A", "--all", ".", "-u", "--update"],
      "reason": "Use 'git add <specific-files>' instead of blanket add."
    },
    {
      "name": "block-npm-global",
      "command": "npm",
      "subcommand": "install",
      "block_args": ["-g", "--global"],
      "reason": "Use npx or local install instead of global."
    }
  ],
  "tests": [
    {
      "command": "git add -A",
      "expect": "blocked",
      "rule": "block-git-add-all"
    },
    {
      "command": "npm install -g typescript",
      "expect": "blocked",
      "rule": "block-npm-global"
    }
  ]
}

Error Handling

Rulebook-backed custom rules fail closed when configured rulebooks cannot be loaded safely:

[!IMPORTANT]

If you add or modify custom rules manually, always validate them with npx -y cc-safety-net rule verify and npx -y cc-safety-net rule test.

Block Output Format

When a custom rule blocks a command, the output includes the rule name:

BLOCKED by CC Safety Net

Reason: [block-git-add-all] Use 'git add <specific-files>' instead of blanket add.

Command: git add -A

Advanced Features

Mode and debug flags use `CC_SAFETY_NET_*` environment variables. Older `SAFETY_NET_*` names (without the CC_ prefix) still work for strict, paranoid, and worktree toggles.

Strict Mode

Malformed or missing hook input JSON always fails closed. By default, ambiguous shell command parsing is allowed through. Enable strict mode to fail closed when a shell command cannot be safely analyzed (e.g., unterminated quotes or malformed bash -c wrappers):

export CC_SAFETY_NET_STRICT=1

Paranoid Mode

Paranoid mode enables stricter safety checks that may be disruptive to normal workflows. You can enable it globally or via focused toggles:

# Enable all paranoid checks
export CC_SAFETY_NET_PARANOID=1

# Or enable specific paranoid checks
export CC_SAFETY_NET_PARANOID_RM=1
export CC_SAFETY_NET_PARANOID_INTERPRETERS=1

Paranoid behavior:

• •rm: blocks non-temp rm -rf even within the current working directory.
• •interpreters: blocks interpreter one-liners like python -c, node -e, ruby -e,

and perl -e (these can hide destructive commands).

Worktree Mode

Linked git worktrees are designed as disposable, isolated workspaces — discarding changes inside one doesn't risk the main working tree. Worktree mode relaxes local-discard rules when (and only when) the command is proven to run inside a linked worktree:

export CC_SAFETY_NET_WORKTREE=1

When enabled, these commands are allowed inside a linked worktree:

• •git restore <file> and git restore --worktree <file>
• •git checkout -- <file>, git checkout <ref> -- <file>, git checkout --force,

and ambiguous multi-positional checkout forms

• •git switch --discard-changes and git switch -f / --force
• •git reset --hard and git reset --merge
• •git clean -f (and combined short flags like -fd)

These remain blocked even in linked worktrees because they reach beyond the local working tree:

• •git push --force (affects remote)
• •git branch -D (affects shared refs)
• •git stash drop / git stash clear (stash is shared across worktrees)
• •git worktree remove --force (could delete another worktree)

Detection is fail-closed and mostly filesystem-based:

• •A linked worktree is identified by a .git file containing gitdir: whose

resolved git directory contains a commondir file. Main worktrees and submodules don't satisfy this and are not relaxed.

• •The cwd walk uses realpath so symlinked paths resolve correctly.
• •git -C <path> (including chained -C and attached -Cpath) is honored;

unresolved targets keep the command blocked.

• •Relaxation is disabled if cwd becomes unknown (e.g., after cd/pushd),

if --git-dir / --work-tree is passed, or if GIT_DIR / GIT_WORK_TREE / GIT_COMMON_DIR is set in the environment.

• •Git may be invoked from a trusted system path to inspect effective config that

could make submodule operations recursive.

Shell Wrapper Detection

The guard recursively analyzes commands wrapped in shells:

bash -c 'git reset --hard'    # Blocked
sh -lc 'rm -rf /'             # Blocked
bash -c 'git stash drop'      # Blocked

Interpreter One-Liner Detection

Detects destructive commands hidden in Python/Node/Ruby/Perl one-liners:

python -c 'import os; os.system("rm -rf /")'  # Blocked
python -c 'import os; os.system("git stash drop")'  # Blocked
python -c 'import os; os.system("dd if=/dev/zero of=/dev/sda")'  # Blocked
python -c 'import os; os.system("mkfs.ext4 /dev/sda1")'  # Blocked
python -c 'import os; os.system("shred -u secret.txt")'  # Blocked

Secret Redaction

Block messages automatically redact sensitive data (tokens, passwords, API keys) to prevent leaking secrets in logs.

Audit Logging

All blocked commands are logged to ~/.cc-safety-net/logs/<session_id>.jsonl for audit purposes:

{"ts": "2025-01-15T10:30:00Z", "command": "git reset --hard", "segment": "git reset --hard", "reason": "...", "cwd": "/path/to/project"}

Sensitive data in log entries is automatically redacted.

Development

See CONTRIBUTING.md for details on how to contribute to this project.

License

MIT