Pi Laya Guardrails

A thin, standalone Pi extension that sends user prompts and LLM-generated tool calls to a local Laya guardrail server. It does not depend on any other Pi extension.

The server owns model inference, policy, deterministic protections, and audit logging. See the Laya server.

Install

Install the companion server first; from the repository root:

cd laya-server
uv venv .venv
uv pip install --python .venv/bin/python "laya==0.3.3"
.venv/bin/python server.py

Then install the published Pi package:

pi install npm:pi-laya-guardrails

For local development instead:

pi install ./laya-guardrails

For a one-off run from this directory:

pi --no-extensions -e ./extensions/laya-guardrails.ts

Modes

• confirm (default): a denied prompt or tool call opens an Allow/Block decision.
• enforce: denied prompts and tool calls are blocked immediately.

Blocking does not terminate the Pi session. Blocked prompts are shown in the UI, truncated to 500 characters. Direct user !/!! commands are intentionally outside this extension's scope; the extension guards LLM calls only.

Set the initial mode with:

PI_LAYA_GUARDRAILS_MODE=confirm pi

Change it during a session:

/laya-guardrails status
/laya-guardrails confirm
/laya-guardrails enforce

Extension configuration

| Environment variable | Default | Purpose | |---|---|---| | PI_LAYA_SOCKET | /tmp/pi-laya-guardrails.sock | Server Unix socket | | PI_LAYA_TIMEOUT_MS | 2000 | Classification request timeout | | PI_LAYA_GUARDRAILS_MODE | confirm | Initial confirm or enforce mode |

The extension fails closed when the server is unavailable, times out, or returns an invalid response.

Tests

npm test