@khaentertainment/pi-greenlight

The go-ahead from whoever's in charge

When your coding agent tries something risky, it asks its boss first.

Try it | The problem | The solution | Full reference

The problem

You let AI agents work in your repo and step away. One of them is fixing a test, and decides the fastest fix is to delete a folder and force-push to main.

pi-jev already catches that for Pi. Before a risky command runs, it asks a fast decision model (Jev) whether it's a good idea, then warns you or asks you to confirm.

That works when you're at the keyboard. More and more, you aren't. Agents run under other agents: an orchestrator splits the work and hands pieces to helpers. When a helper's command gets flagged, a warning goes unread and a confirm prompt waits forever. The safety check either does nothing or freezes the work.

The solution

Greenlight sends the decision up the chain of command.

When a command looks risky, Greenlight pauses it and asks the agent's orchestrator: approve, deny, or ask a human. The orchestrator knows the plan, so it can tell a planned cleanup from a mistake. If it doesn't answer in time, you're asked instead, the same as in pi-jev.

It works with:

• Traycer, the first supported agentic development environment (ADE). A helper agent's flagged command goes straight to its parent agent in Traycer.
• A parent Pi agent, when one Pi runs another.
• Any orchestrator that speaks A2A, the open agent-to-agent protocol.
• You, if you're working one-on-one with an agent.

Orca and Xirp, two more ADEs, are next.

Greenlight also lets you reach Jev through the provider you already use: TypeSafe, OpenRouter, Cloudflare, Vercel, or OpenCode. Jev is the default, not the only option: Mercury Decide can run the output check, and Clef Flash through OpenRouter is calibrated for the gate with the same thresholds. The calibration data covers all three.

Everything pi-jev does is still here. It checks risky commands in about a third of a second, warns when a command prints a password, and never blocks your agent if the check itself fails.

Try it in two minutes

1. Install it:

   pi install npm:@khaentertainment/pi-greenlight

2. Give it a key. Run /login typesafe in Pi, or set TYPESAFE_API_KEY.

3. Keep working as usual. Greenlight starts by only warning you, so nothing gets blocked. Run /greenlight log to see what it caught.

4. Running agents under Traycer or another orchestrator? Hand the decisions up the chain:

   /greenlight mode escalate

   Working alone? Use /greenlight mode enforce to be asked before risky commands run.

Install Greenlight now, and give your agents a boss to answer to.

Want the details?

Orchestrator setup, providers, settings, and the measurements behind every threshold are in the reference guide.

---

A fork of y0usaf/pi-jev, which built the command check this is based on. MIT licensed.