Shell Security Ultimate

Your agent has root access. Every command it runs is one bad inference away from rm -rf / or curl | bash from a stranger's repo.

This skill makes it say what it is about to do, and how bad that could be, first.

How It Works

Every shell command gets classified before execution:

• 🟢 SAFE — Read-only, harmless. Runs without friction.
• 🟡 WARN — Could modify state. Logged, flagged, your call.
• 🔴 CRIT — Destructive or irreversible. Blocked until you say so.

No command runs unclassified. No silent chmod 777. No quiet dd if=/dev/zero. Your agent won't accidentally email your SSH keys, won't helpfully format a disk, and won't DROP TABLE users because it misread the task.

What You Get

• Pre-execution classification for every command, every time
• A four-line report per command — level, result, purpose, and any warning — printed to your terminal as it happens. It is a display, not a log file: nothing is written to disk and nothing is kept between calls
• Full override control — approve, deny, or escalate at any level

Who It's For

Anyone giving an AI agent shell access and wanting to sleep at night.

One caveat up front, because a security skill is the wrong place to bury it: by default this is prompt-level guidance, not a kernel-level block. The agent classifies and stops because it was told to, and an instruction can be talked around. Real code-level blocking needs the optional patch below plus a hook plugin. See What's Enforced by Code, and What Isn't.

Clone it. Fork it. Break it. Make it yours.

👉 Explore the full project: github.com/globalcaos/clawdbot-moltbot-openclaw

What's Actually in the Box

Four working files, two of which are inert until you run them on purpose, plus a tests/ directory that runs only when you run it. Anything else in the download (skill-card.md, _meta.json, .clawhub/) is registry metadata and does nothing.

Installing this skill patches nothing, writes nothing, and starts nothing.

Permissions, Data Flow & Consent

Short version: the classification is prompt text and touches no files. One optional script rewrites source code in a project that is not this one, and it asks first. Longer version, because a security skill is the last place you should take that on trust:

What data it touches. The text of the shell commands your agent is about to run, and whatever you write as their purpose. That is it. cmd_display.py receives those as command-line arguments, formats them, and writes to stdout. It opens no files, keeps no history, and stores nothing between calls. One caveat worth stating plainly: anything passed as a command-line argument is visible in the process table to other users on the same machine and lands in your shell history — so do not pipe secrets through the result argument.

What goes over the network. Nothing. There is no network code in this package.

What credentials it reads. None. No tokens, no keys, no auth files, and no environment variables other than OPENCLAW_DIR — a path you set, used only by the two patch scripts.

What gets written to disk. Nothing at all, unless you run patch-openclaw.sh. When you do, and only after you confirm, it writes exactly two things, both inside the checkout you named:

• a timestamped backup, pi-tool-definition-adapter.ts.backup.<date>
• the modified src/agents/pi-tool-definition-adapter.ts itself

unpatch-openclaw.sh writes one more backup (.unpatch-backup.<date>) before restoring.

Capabilities, and why each exists.

The Optional Patch — Consent and the Off Switch

patch-openclaw.sh is the one genuinely high-impact thing here, so it is worth being blunt: it rewrites source code in a codebase that is not this skill. It exists because OpenClaw's tool dispatcher does not, by default, give plugins a chance to veto a tool call. The patch adds that hook. Read the script before you run it — it is short, and it prints the exact change it will make before making it.

What guards it:

• --dry-run shows the target file and the change, then exits without touching anything.
• Explicit consent. It refuses to proceed until you type yes, or pass --yes. In a non-interactive shell without --yes it refuses outright rather than assuming.
• Target validation. It refuses to edit a tree whose package.json does not name an OpenClaw-family project, so a typo in OPENCLAW_DIR cannot point sed at your other work. Override with --allow-any-repo if you know better.
• Verified before anything is written. The patched file is built in a temp file and checked there: exactly one hook block, exactly one import, the import above the hook, the hook immediately above the call it guards — and then the patch is reversed on the candidate and the result must match the original byte for byte. If any of that fails the real file is never opened for writing. A half-applied patch (hook inserted, import anchor missed) leaves source that will not compile, and a check that only looks for the marker would call it a success. This one does not.
• A backup first — written after verification passes, so a refused patch leaves no litter — and an automatic restore if the file on disk somehow does not match what was verified.
• No build unless you ask. pnpm build runs only with --rebuild; otherwise the script prints the command and leaves the decision to you.

./scripts/patch-openclaw.sh --dry-run     # see the change, touch nothing
./scripts/patch-openclaw.sh               # asks before editing; does not rebuild
./scripts/unpatch-openclaw.sh             # OFF SWITCH — restores the file

The off switch is held to the same standard. unpatch-openclaw.sh resolves the block to line numbers and validates them before deleting anything: exactly one start marker, exactly one end marker, in that order, and a block the expected length. The naive version of this is a marker-to-marker sed range, which silently deletes to end of file when the end marker is missing — and then passes a "the marker is gone" check. If the block is not the one this skill wrote, the script prints it and stops; --force-range removes it anyway, once you have looked.

To undo it by hand instead, restore the .backup.<date> file it left beside the original, or run git checkout src/agents/pi-tool-definition-adapter.ts in that checkout.

What the patch does not do. It does not block anything on its own. It only lets a plugin that registers a before_tool_call hook return a block decision. This package does not ship such a plugin. Patching and then assuming you are protected is worse than not patching — so after patching, nothing changes until you install a plugin that implements the hook.

What's Enforced by Code, and What Isn't

The honest boundary, because the difference matters more here than anywhere else:

• Instruction-level (the default). The SAFE / WARN / CRIT gate is guidance the agent follows. It is only as strong as the model's instruction-following — a determined jailbreak, a prompt injection inside a file the agent reads, or a plain misclassification can get past it. Treat it as a seatbelt, not a vault.
• Code-level (only if you patch, and only with a hook plugin). A before_tool_call hook can genuinely refuse to execute a tool. That is real enforcement, and it needs both the patch above and a plugin that uses it.

If you need a hard guarantee, run the agent in a container or a VM with credentials it cannot escalate. This skill lowers the odds of a bad command; it does not make one impossible.

Turning it off. Deactivate or uninstall the skill and the classification behaviour stops — it is prompt text, so nothing lingers. Telling the agent to stand down for a session works the same way. If you ran the patch, run ./scripts/unpatch-openclaw.sh (then rebuild) to return the checkout to its original behaviour.

cmd_display.py Reference

python3 scripts/cmd_display.py <level> "<command>" "<purpose>" "<result>" [warning] [action]

Every argument is stripped before it is printed. The result you pass is usually the output of a command, which means it is text someone else may control — a filename in a cloned repo, an error from a remote server. Raw, a string carrying an erase-line escape can repaint the banner above it and turn a red CRITICAL header green. So ANSI/OSC/DCS escape sequences and control characters are removed from every argument, and each is length-capped so a huge one cannot scroll the real banner off screen. tests/test_cmd_display.py covers it.

Levels: safe 🟢, low 🔵, medium 🟡, high 🟠, critical 🔴. Output is at most four lines — level and command, result, purpose, and an optional warning or next action. It prints; it does not run the command for you, and it decides nothing.

Running the tests

python3 tests/test_cmd_display.py      # display cannot be forged by its own input
bash    tests/test_patch_roundtrip.sh  # patch/unpatch against a throwaway fake checkout

No dependencies, no network, and the patch tests build their own disposable checkout in a temp directory — they never touch a real one.

Everything this document describes is in this package. If you find a claim here that the code does not do, that is a bug — open an issue on the repo.