Feature request: directional external-directory policy with post-classifier confirmation

[BaseBlank]

Thanks for your great work.

Can consider adding a first-class external-directory permission policy to pi-automode, with separate handling for read and write/mutation access outside the current working directory?

The main use case is to allow agents to inspect files outside the workspace freely, while protecting those files from accidental modification or deletion.

Ideally, external writes could follow this flow:

1. deterministic safety checks;
2. model/classifier review;
3. if the classifier considers the action acceptable, require explicit human confirmation;
4. only then execute.

This would provide a stronger workflow for high-risk external modifications than either classifier-only approval or the current permissions.ask ordering.

Motivation

My Pi workflow is intentionally permissive for reads but conservative for mutations.

I often need the agent to:

• inspect source code in other repositories;
• read configuration files elsewhere on the machine;
• inspect installed packages and system configuration;
• occasionally modify configuration outside the current workspace.

Reading these files is usually low risk from an integrity perspective, so requiring a classifier call for every external read adds unnecessary latency and review overhead.

Writing outside the workspace is different. A mistaken edit to a user configuration, another repository, or a system-related file may cause substantial damage.

At the same time, I do not want to rely only on manual approval for those commands. Many shell commands are difficult for a user to review accurately, and one of the main benefits of auto mode is having a model perform semantic risk analysis.

For external writes, the workflow I would prefer is therefore:

external write/edit
        |
        v
deterministic hard-deny checks
        |
        v
classifier review
   |             |
 block          allow
   |             |
   v             v
 stop      human confirmation
                 |
              yes / no