Hacker News Reader

Comment by comradesmith | Hacker News Reader

1. Make tests 2. Commit them 3. Proceed with implementation and tell agent to use the tests but not modify them

It will probably comply, and at least if it does change the tests you can always revert those files to where you committed them

View on HN

Are there really no ways to control read/write permissions in a smart way? I've not had to do this yet, but is it really only capable of either being advisory with you implementing all the code, or it having full control over the repo where you just hope nothing important is changed?

You could probably make a system-level restriction so the software physically can't modify certain files, but I'm not sure how well that's going to fly if the program fails to edit it and there's no feedback of the failure.

You can use a Claude PreToolUse command hook to prevent write (or even read) access to specific files.

With this approach you can enforce that Claude cannot access to specific files. It’s a guarantee and will always work, unlike a prompt or Claude.md which is just a suggestion that can be forgotten or ignored.

This post has an example hook for blocking access to sensitive files:

https://aiorg.dev/blog/claude-code-hooks#:~:text=Protect%20s...

No. I don't want the mental burden of auditing whether it modified the tests.

Then, run the agent vm-sandboxed, with tests mounted as a read-only directory, if your directory structure allows it.

Or, less securely, hash the tests and check the hash with a hook, post tool use. Or a commit hook.

Reply on news.ycombinator.com