Auto mode using a safety classifier for tool calls

Feature Description

Instead of just having edit and plan modes, introduce a new mode called "Auto Mode". This would use an automatic safety classifier for things like bash commands and such, similar to what Claude Code has, so that users don't have to constantly babysit the harness when it's running commands. It would likely be powered by a cheap open source model, like DeepSeek V4 Flash or Qwen 3.8 Flash to keep costs low, but the classifier model should be able to be changed by the user at any point.

Use Case

Allows for users who are working to step back while the agent does its own thing, instead of having to manually accept bash commands and the like. There is already functionality for allowing the agent to run specific commands on its own, after it's been allowed, but I find that often doesn't work very well for one reason or another. In workflows where the user may be doing multiple things at once, it can get annoying to have to go back and allow things in order to stop the agent from hanging on a single bash command request for potentially minutes on end while the user isn't aware it's even there.

Additional Context

No response

How important is this to you?

Important for my workflow

No files, tests, or concrete entry points are named. Start by locating the existing edit and plan modes and the functionality that allows approved commands to run automatically. Done would require a defined Auto Mode that evaluates tool calls with a user-configurable safety classifier and avoids unnecessary manual approval.