Build log Part 2

We installed OLP on our own PC, and our own AI session stopped first

FeedPia

On the morning of Nov 9, we installed OLP 0.1.1 on our founder’s PC. A few minutes later, it stopped two commands from the AI session we were using, before they ran.

What the installer told us

The installer added OLP as a hook in our Claude Code settings, then printed a few lines. One said that OLP runs only on this computer, and that the hook itself sends nothing over the network.

Another line stood out. It said our settings had 32 “always allow” rules for commands that can delete or overwrite files. OLP didn’t change those rules. It only pointed them out.

What stopped

  • A Python command that runs a test script
  • An MCP tool call that reads the terminal screen

OLP marked both as “not classified yet for this project.” It treats unclassified commands as writes, so it stopped both before they ran. It told us that a person has to classify each one, once, in their own terminal. Read commands like grep and ls went straight through.

Why we built it this way

If OLP let an unknown command through on the guess that it only reads, and that command deleted files, nobody would find out. So when OLP doesn’t know, it asks. We listed this as a known limit from the start: OLP asks even about read-only commands it doesn’t recognize.

What’s still open

That session was running in a mode with no permission prompt. With nowhere to ask, the work stopped there. We haven’t decided yet how to handle this case. When we do, and it goes into a new version, we’ll say so.

Our mistake

That day, we said where the previous version was installed, without checking, and we were wrong. So OLP was installed fresh in a different folder, the same folder where the session stopped. We’ve left this record as it happened.

Our founder put it this way: “My own tool stopped my own work first. I built it to stop and ask before running a command it hasn’t seen, and on my own PC it did exactly that.”

We’ve written down what confused us and what stopped.

Installing OLP takes one line, and there’s no signup.

npx @feedpia/olp install

Repository: https://github.com/feedpia/outcome-lookahead-protocol

If something stops or looks wrong while you try it, please tell us in an issue.

At your next permission prompt, look before you choose. It’s one line to install.

See the one-line install