Danger wall
The rules on your computer that stop risky actions.
The danger wall runs inside the desktop app on your computer. Every command the AI suggests is checked before it runs. It is code, not an instruction to the AI, so the AI can't talk its way past it.
Never allowed
These are refused, whatever anyone says, and can't be allowed once:
- Deleting or moving files with commands (
rm,del,mv...), formatting disks, changing permissions or owners. - Starting a shell or running arbitrary code (
bash,cmd,powershell,eval...). - Downloading or connecting elsewhere (
curl,wget,ssh,scp,nc...). - Running as administrator (
sudo,runas), changing system settings, services or scheduled tasks, stopping programs. git push,reset,clean,rebase,config,remoteand other history-changing Git commands.- Any command with shell symbols such as
; & | < > ` $ %, programs called by a path, or arguments pointing outside the folder or at secret files.
Asks you first
| Kind | Examples | Setting |
|---|---|---|
| Database | psql, mysql, mongosh, migrations, db:reset | Block database commands |
| Deploy and publish | docker, kubectl, terraform, vercel, npm publish | Block deploys |
| Installing packages | npm install x, pip install x, go get | Block installing packages |
| Anything unknown | Commands that are not known test or build commands | Always asks |
Runs without asking
Known test, lint and build commands, for example npm test, npm run build, pytest, python -m unittest, go test, cargo test, dotnet test, mvn test, php artisan test, and commands you list in Commands that may always run. Choose Ask before every command in Permissions to be asked for these too.
Extra checks before tests run
Tests run your project's own code. So CodeStilt asks you first when a task:
- changed a file that controls what the test command runs, such as
package.json,Makefile,pyproject.toml,conftest.pyor a test runner config; or - added code that goes online, starts programs, or reads environment variables, keys or
.env.
You see the file and the reason, and choose Allow once or Block. This stops a hidden instruction in a README or issue from getting the AI to write code that sends your secrets away. See Projects you don't trust.