Back
developer-toolsagentsOpen source

pi-warden

Supervise a Pi coding agent with project-rule and action judgments.

What Jev decides

Whether a proposed action breaks a rule or raises risk, and whether progress or a completion claim is supported.

pi-warden checks proposed writes, project rules, repeated failures, and unverified completion claims, then usually steers the agent to correct itself.

CHOICENOUL
Jev role
Assess context-dependent agent actions and work quality for Warden policy.
Input
The proposed action or result, relevant project rules, and recent agent context.
Output
Typed risk and quality judgments consumed by Warden policy.
Decision frequency
At eligible writes, tool results, and completion checks when judgments are enabled.
Fallback
Local pattern guards continue without a model key; policy can steer or hold actions.
Author reported · Source reviewed

How it works

Input
The proposed action or result, relevant project rules, and recent agent context.
Jev
CHOICE + NOUL
Output
Typed risk and quality judgments consumed by Warden policy.

Overview

pi-warden is a Pi extension that combines local guards with optional model judgments. It checks tool actions before they run, compares edits with project rules, notices repeated failures, and challenges unsupported done claims. Most findings are fed back to the agent; irreversible actions can be held.

Offline guards work without a key. Jev-backed judgments require enabling Warden and a TypeSafe key; the optional capability conscience is beta and off by default. The model gives a judgment, while configured policy determines the intervention.

Language
TypeScript
License
MIT
Author
DevMortimer
Sources checked Sep 23, 2026

Sources

Related projects