COMMUNITY OBSERVEDMetric claims: UNVERIFIEDSOURCE AUDITEDDeveloper Tools

pi-warden: Guardrails for Pi Agents

Enforces project rules on every write, catches loops, and halts unverified done claims.

Overview & result

pi-warden provides guardrails for AI agents that steer rather than interrupt. Operating on the Pi framework, it checks every file modification against project conventions, blocks destructive or hard-to-undo actions (measured at only ~3 per 1,000 calls in real sessions), catches stuck retry loops, and prevents unverified completion claims.

From the original post

Guardrails for Pi that steer instead of interrupt: enforces your project rules on every write, holds only hard-to-undo actions (3 per 1,000 calls), catches unverified 'done' claims and stuck loops.

— Dev Mortimer

How Jev fits in the loop

  1. Monitors tool invocation requests before execution in the Pi agent harness.
  2. Jev evaluates proposed tool actions against safety policies and recent execution history.
  3. Permits benign actions automatically while holding high-risk commands for confirmation.

How to reproduce

  1. Clone https://github.com/DevMortimer/pi-warden and link into your Pi project.
  2. Configure project guardrail rules in your repository root.
  3. Run agent workflows and observe pi-warden intercepting destructive edits and infinite retry loops.

Why this build matters

Provides high-confidence agent security and prevents repetitive loops without slowing down standard benign tool calls.

Limitations

  • Designed specifically for agents built on the Pi framework.
  • Interrupts only actions flagged as hard to undo.

Patterns