What does the gate tool calls with Jev recipe build?

A per-call gate for a risky tool: instead of pausing every call for human approval, it sends the proposed call, the ticket, and your policy to Jev and routes on the returned probabilities.

Which Jev model does the recipe use?

typesafe/jev-1.13 on OpenRouter, through the Decisions endpoint, so the same OPENROUTER_API_KEY that runs the agent model also runs the check.

What are the three gate outcomes?

Approve, block, or review. The tool runs only on approve, the model receives a refusal with a reason on block, and the loop pauses with status: 'awaiting_hitl' on review.

How does the gate turn probabilities into decisions?

Jev receives the state and three yes-or-no questions; fixed thresholds turn each probability into a decision. The recipe's example asks about the proposed refund against the ticket and policy text.

Why not keep a static human-in-the-loop rule?

The page says static rules check one thing, like the tool name or a condition on arguments, but the same refund tool can be safe in one call and not the next, and the difference lives in the ticket, which Jev can read.

What do I need to run this recipe?

An existing TypeScript agent using @openrouter/agent with a HITL tool, zod 4, @types/node or @types/bun, an OPENROUTER_API_KEY, and a StateAccessor for conversation state.

Does the gate parse generated text?

No. The recipe stresses that Jev returns a probability per question rather than prose, so the gate is a few numeric comparisons your code controls, with reasons kept for the audit log.

Docs

CommunityJev Cookbooks

Gate Agent Tool Calls with Jev

An OpenRouter cookbook recipe that replaces a blanket human-approval rule with a per-call Jev check: the proposed tool call, ticket, and policy go to Jev, whose probabilities route to approve, block, or review.

See highlights

Open openrouter.ai

Format: Docs
Reading time: 2 min
Published by: Community
Last verified: Oct 2026

Tagscommunity catalog-import cobanov openrouter agents hitl

Works forJev for Decision gating Jev for Verification Jev for Tool calling

Used inJev for Software & DevTools Jev for Customer Support

Languages: TypeScript
Platforms: OpenRouter API
Capabilities: Requires API key

Highlights

Jev is on OpenRouter as typesafe/jev-1.13 through the Decisions endpoint, so the agent model and the gate share one API key.
gateRefund runs two deterministic order checks, then asks Jev three yes-or-no questions about the refund, ticket, and policy.
Thresholds route the result: run on approve, refuse on block, and pause with status: 'awaiting_hitl' on review.
The client posts to https://openrouter.ai/api/alpha/decisions and validates the reply with zod, throwing on bad probabilities.
The recipe reports one Decisions request per call at under $0.0001 in its runs, and each decision includes a reason for the audit log.

Watch out

Grounded in the truncated cookbook capture from 2026-10-15; only the setup and the start of the request client are visible, the full gateRefund implementation, fixtures, and test harness were cut off, and the sub-$0.0001 cost is the page's own figure.

Questions about this entry

Short answers drawn from the linked project or page.

From the Jev guides

Original write-ups on what this entry is for.

Retrieval with Decisions: Filter, Rerank, and Verify

Oct 12, 2026

Retrieval is a sequence of bounded judgments. How a decision model slots into passage filtering, semantic find, reranking, and citation checks, and what stays in deterministic code.

Jev Cookbooks#retrieval#reranking#rag

Read article

One Request, Many Questions: Batching, Fan-Out, and Composite Scoring

Sep 12, 2026

Jev bills input tokens and gives output away, and questions in one request run in parallel. The design consequence is simple: ask more per call, then decide in code.

Jev Cookbooks#batching#patterns#search