Four approval modes and command-level security restrictions for the Pi coding agent.

Story

I use the Pi coding agent a lot, and I want to control what the agent is allowed to do: sometimes full permissions, sometimes read-only planning, and always a hard block for dangerous commands.

pi-permission-suite is a fork with this lineage: @gotgenes/pi-permission-system → Eddie0521/pi-permission-suite → ansidev/pi-permission-suite. My fork adds a pluggable Decision API, wildcard and argument-based tool rules, and a debug command.

Technologies

pi-permission-suite is built with:

• TypeScript.
• tree-sitter-bash and web-tree-sitter to parse shell commands.

Features

• Four approval modes: Act, Auto, Ask and Plan.
• /approval-mode command and Ctrl+Q shortcut to switch modes.
• set_approval_mode tool, so the agent can switch modes by itself.
• Rule engine to block dangerous commands and protect paths in all modes.
• Subagent-based auto-approval for uncertain tool calls.
• Pluggable Decision API for Auto mode (subagent, Jev or HTTP providers) with a JSONL audit log and fail-closed behavior.
• Wildcard and argument-based tools allow/deny rules.
• /pi-permission-suite debug on|off to log Decision API calls.

Installation

pi install git:github.com/ansidev/pi-permission-suite

Don't use npm:pi-permission-suite: that package is published from the upstream project and doesn't include this fork's changes.

Usage

https://github.com/ansidev/pi-permission-suite/blob/main/README.md