Harness adapter

Unlike the other three ports, the evaluation adapter is an HTTP contract, not an in-process port. Simple targets can run inside the CloudGrid AI harness; complex agents and workflows can expose a black-box adapter URL.

Core Endpoint

Required behaviors

• W3C trace context (traceparent / tracestate) is preserved on every call.
• Idempotency is required for retries — repeated calls with the same idempotency key return the same result.
• No full traces in response bodies. The target should emit spans through OTLP. Adapter responses carry the final output and bounded summaries.
• Expected output is omitted by default. The evaluated system should not see labels unless the evaluation settings explicitly allow it.
• Provider credentials stay outside CloudGrid. Adapter deployments own their provider keys and private tool configuration.
• No storage shortcuts. Results return through the adapter protocol and CloudGrid message contracts; telemetry still returns through normal OTLP ingest, storage-write persistence, and storage-read GraphQL views.

Example: a minimal harness skeleton

import { serve } from "@hono/node-server";
import { Hono } from "hono";

const app = new Hono();

app.post("/v1/run", async (c) => {
  const body = await c.req.json();
  // 1) replay the agent against the input
  // 2) emit OTLP spans (with traceparent from request)
  // 3) return a run summary
  return c.json({
    runId: body.idempotencyKey,
    status: "ok",
    tokens: { in: 1240, out: 312 },
    durationMs: 412,
  });
});

serve({ fetch: app.fetch, port: 8088 });

Idempotency keys