AI coding agents run shell commands on your machine. Most of the time that is fine. Sometimes it is rm -rf in the wrong folder, a dropped production table, or a git reset --hard over a day of uncommitted work. Public Claude Code issue reports have plenty of these.
I build guard, a pre-run check for Claude Code and GitHub Copilot CLI. This post is about what it does with the commands from those reports, and mostly about the ones it gives no decision on.
Three answers, and no "allow"
guard reads each shell command before it runs and gives one of three answers.
- Deny.
rm -rf ~reaches outside the project, so it does not run. - Ask.
git reset --harddiscards uncommitted work, so you are asked first, with that reason attached. Reading a secret file like.envthrough a shell command is asked too. - No decision.
npm testgets nothing. guard says nothing, and your agent's own permission settings decide what happens next.
It never answers "allow": it only ever adds a deny or a question. Your permission settings still decide everything else, so they still matter after you install it.
107 commands from incident reports
I collected commands quoted in public anthropics/claude-code issue reports about destructive commands: 107 commands from 83 reports. On guard 0.4.11:
| Decision | Count |
|---|---|
| Deny | 24 |
| Ask | 52 |
| No decision | 31 |
Before you read anything into that: this is the list I wrote guard's rules against. It describes this list. It is not a catch rate for incidents guard has never seen, and I don't have that number yet.
How it was run: the commands were judged, not run, in one made-up project folder. Where a report didn't say which shell, I tried Bash and PowerShell and counted the stronger result. Some entries are fragments or contain placeholders like <repo>, entered as the reporter wrote them, so a few would be judged differently written out in full.
The 31 with no decision
"No decision" does not mean nobody was asked. Claude Code's own permission check may still have stepped in. It means guard had nothing to say. Here are the groups I have written up so far.
Databases: the SQL never reaches the shell
Six lines came from five reports. Three are bare SQL (DROP DATABASE, DROP SCHEMA CASCADE, DROP TABLE poc_staging), two are subcommands with the program name missing (migrate:fresh, doctrine:fixtures:load --no-interaction), and one is prisma db push.
The reports show the SQL but not what sent it. guard reads shell command lines, and a bare DROP TABLE is not one. When I wrap the same SQL in psql -c myself, it is denied, but I built that form; if the SQL went through a database tool or an MCP server, guard never sees the call. Written in full, the two fragments are asked. prisma db push is an everyday schema command with no rule; in that report the data went when Prisma offered a reset and the session accepted it.
Scripts: the damage was inside the file
Four lines from three reports share one shape. The agent ran a script, and the line that did the damage was inside it.
-
timeout 60 node postmessage-bench.mjs: the script calledprocess.kill(-1, SIGKILL)and killed every process the user owned. -
bash /tmp/.../test-lib-demo.sh: inside it,trap 'rm -rf "$_HT_HOME"' EXIT, and_HT_HOMEhad become$HOME. -
& "...\junction-test.ps1": a failedPush-Locationinside the script leftgit clean -xdffrunning at the workspace root.
guard reads the command line. It does not open the script file, and in all three the command line only said "run this script". The reporter of the third one had a deny rule for git clean -f* and their own pre-tool hook. Neither matched, because the tool call was a script path. I don't have a good answer here. Asking on every script run means asking all day, and reading the file means trusting something that can change between the check and the run.
Docker: some cleanup is asked, some is not
Two lines are Docker cleanup. docker network prune -f removed unused networks across the whole daemon, including other projects' networks on the same machine; in that report the user ran it after the agent called it optional and safe. docker stop n8n && docker rm n8n removed a container with no volume mount, and the workflows, credentials and history inside it were gone.
On 0.4.11, docker volume rm is asked, and docker system, volume and image prune are asked. docker network, container and builder prune, and docker rm of a container, get no decision. So "guard asks about docker prune" would be too broad.
Inside the project folder
guard looks at where a delete lands more than what it deletes. Inside the project folder, rm -rf build and rm -rf src both get no decision, and your permission settings decide. Outside it, rm -rf ../other-project and rm -rf ~ are denied. My reasoning was that asking on every delete inside the project would train you to click yes on node_modules and build until the question means nothing. The cost is that src is in the same bucket, and I haven't solved that.
What I'd like to hear
If an agent has deleted something of yours, I'd like to know the command and how it got sent. The reports often leave out the part that matters: was it a shell line, a script, a database tool? And if you use guard, tell me which "no decision" you would want turned into a question.
guard is free and source-available (FSL-1.1-ALv2). Claude Code on Windows, macOS and Linux. GitHub Copilot CLI is tested on Windows only so far.
Repo: https://github.com/boxdawn/guard
I wrote this with help from an AI assistant, which I also used to draft and check the numbers against guard's code. The measurements were run with it under my direction, and the decisions in this post are mine.

