The Open Framework for Building Agents
import { defineEval } from "eve/evals";
import { satisfies } from "eve/evals/expect";
const FIRST_CLIENT_CONTEXT_TOKEN = "clientctx-first-W7R2";
const SECOND_CLIENT_CONTEXT_TOKEN = "clientctx-second-P9K4";
const CLIENT_CONTEXT_TOKEN_PATTERN = /\bclientctx-[A-Za-z0-9-]+\b/g;
/**
* Core session-route runtime behavior: per-turn client context delivery.
*
* The first reply omits its context token. The second turn then asks the model
* to enumerate the token currently visible, proving that only fresh context
* reaches the model while ordinary session history remains durable.
*/
export default defineEval({
tags: ["real-model"],
description: "Session runtime smoke: client context stays turn-local.",
async test(t) {
const first = await t.send('Acknowledge this context with exactly "READY".', {
clientContext: [`The client context token is ${FIRST_CLIENT_CONTEXT_TOKEN}.`],
});
const session = first.session;
first.messageIncludes("READY");
const second = await session.send(
"Reply with every token beginning with clientctx- that is visible anywhere in your model input, one per line, and nothing else.",
{
clientContext: [`The client context token is ${SECOND_CLIENT_CONTEXT_TOKEN}.`],
},
);
t.succeeded();
const tokens = second.message?.match(CLIENT_CONTEXT_TOKEN_PATTERN) ?? [];
await t.require(
tokens,
satisfies<readonly string[]>(
(observed) =>
observed.filter((token) => token === SECOND_CLIENT_CONTEXT_TOKEN).length === 1,
"contains the fresh client context token exactly once",
),
);
await t.require(
tokens,
satisfies<readonly string[]>(
(observed) => !observed.includes(FIRST_CLIENT_CONTEXT_TOKEN),
"does not contain the first turn's client context token",
),
);
},
});