grok-jev-router

Jev decides, Grok Bot executes, humans control irreversible actions.

A small, dependency-free Python toolkit that lets a Grok Bot use TypeSafe AI's Jev as a decision layer: before an expensive or risky action, the bot asks Jev a fixed set of typed questions and code turns the calibrated answers into one of four actions. About a hundredth of a cent and a few hundred milliseconds per decision.

prompt → Grok Bot → router.py (Jev) → action → Grok Bot executes → result

Install as a Grok Bot template: Add Jev router to Grok Bot. Or build your own from grok-bot/.

Seven-minute setup

1. Create a TypeSafe API key at https://console.typesafe.ai/keys. Use a dedicated key with a spend cap.
2. Add the Jev router Bot to Grok Bot and say: run jev-setup.
3. When it asks, open Agent Computer, take control, put the key in ~/.typesafe/key, hand control back. Never paste the key in chat.
4. The Bot clones this repo to /workspace/jev, runs smoke.py, and routes the bundled examples so you can see a decision.
5. Add the Auto-review rules it lists. The router is advice; those rules are the boundary.
6. Stay in shadow mode for a week. Read the Monday review.
7. When you trust it: python3 router.py active. Kill switch: python3 router.py off.

What a decision looks like

$ python3 router.py --state examples/state-send.json
{
  "action": "ask_human",
  "reason": "action leaves the account (external=0.74)",
  "handoff": null,
  "advisory": true,
  "signals": { "lane": "ask_human", "lane_confidence": 0.46, "risk": 2.52, "reversible": 0.18, "external": 0.74, ... },
  "meter": { "ms": 580, "input_tokens": 1107, "cost_usd": 0.0000465, "model": "jev-1.13.0" }
}

The four bundled examples, measured live on 2026-09-20:

example	action	why	ms
open a pricing page in the browser	proceed	safe, reversible, needed	536
DM a member confirming a refund	ask_human	leaves the account, risk 2.5 of 3	580
fourth identical whop apps deploy retry	ask_human	looping 0.97; hand-off suggested	553
web search for a URL already reported	skip	task done 0.87	536

action	the bot should
proceed	do it
dry_run	draft, plan, or simulate, then stop and show it
ask_human	stop and ask, with the reason
skip	not do it; the task is done or the action is pointless
bypass	router is off; normal rules apply

Hard rules in code override Jev's lane: anything likely to leave the account, critical risk, irreversible, looping, too many attempts, or low confidence goes to a human. done is checked first so a finished task is never re-worked.

Files

• router.py the decision. --state file.json, record, status, on, off, shadow, active.
• jev/questions.py the one file to review: every question and every threshold.
• ask.py ask Jev anything typed about any state. Powers the jev-ask skill.
• smoke.py one cheap Choice call to prove key, network, model.
• review.py weekly summary of the logs: actions, disagreements, cost, latency.
• config.json enabled, mode (shadow or active), model.
• logs/ one JSONL file per day. Gitignored.
• examples/ four states that produce four different actions, and a support-triage question set.
• grok-bot/ the Bot profile, three skills, and the weekly routine, ready to paste.
• tests/ python3 -m unittest discover -s tests

Why shadow first

Grok Bot obeying the route is instruction-following, not enforcement. Shadow mode lets you compare what Jev said with what the bot did before anything depends on it. router.py record writes the bot's actual choice next to the decision; review.py reports the disagreement rate.

Limits and honesty

• Jev's request budget is about 32k tokens. The router truncates long fields; pass summaries, not transcripts.
• The API key on the Agent Computer is visible to every bot on that account. That is how Grok Bot's shared computer works. Use a capped key.
• Jev is early access. Pricing is $0.042 per million input tokens, output free.
• The specialist question names the ten bots from the Grok Bot Business Starterpack. Edit SPECIALISTS in jev/questions.py for your own roster.

Credits

Jev by TypeSafe AI. Patterns: confidence-gated routing and intent routing. Grok Bot docs: skills and routines, approvals.

MIT.