`write_file` overwrites existing files without confirm

[wary hedge]

https://github.com/NousResearch/hermes-agent/issues/510#issuecomment-4303266210

Do you have the experience that your agent using write_file tool to overwriting existing files?

I have a suggestion to add a confirm warning to make the agent think twice when it doing such a tool call, what is your thoughts?

@bright hedge

If you dont want it to be able to I'd suggest just using docker backend

But the problem is I'm requesting agent to write a worklog for what we done, I should definitely need to allowed it to able doing this

Using Docker backed won't help this, because it will doing absolutely same thing when inside sandboxed environments.

[bright hedge]

There is no foolproof way to prevent it from writing to the code

or the files

If you dont approve it it can just find a bunch of ways to write scripts to edit a file

[wary hedge]

Even I added the iron rules to its SOUL.md that "don't use write tool when there's a existing file, use patch tool instead"

But it will still forgot

I've setup restic backup every hour and git so I don't really need to worry about data loss, but it is still making me frustrating

[wary hedge]

Clear message to the agent that the "deny" is requested/triggered by user/security scan/pre-define rules is required, and ask agent to think twice or ask the user for further instructions

[wary hedge]

And don't messup the system prompt for a such warning in the tool description, only shows when it is doing such things, so the system prompt stays clean

[bright hedge]

That implies it will still ignore your steering