Eve (tracing)

Eve is Vercel’s TypeScript framework for durable agents. Respan maps Eve turns and sessions into agent traces, translates the nested AI SDK model and tool spans, and preserves exact delegated-session lineage.

Setup

$
npm install @respan/respan @respan/instrumentation-eve eve@0.26.1 ai@7.0.26

$
export RESPAN_API_KEY="YOUR_RESPAN_API_KEY"

1
import { Respan } from "@respan/respan";
2
import {
3
  EveInstrumentor,
4
  withEveLineage,
5
} from "@respan/instrumentation-eve";
6
import { defineInstrumentation } from "eve/instrumentation";
7
8
const respan = new Respan({
9
  apiKey: process.env.RESPAN_API_KEY,
10
  instrumentations: [new EveInstrumentor()],
11
});
12
13
await respan.initialize();
14
15
export default withEveLineage(
16
  defineInstrumentation({
17
    recordInputs: false,
18
    recordOutputs: false,
19
  }),
20
);

$
npx eve dev

Capture model content

The safe setup above does not export model inputs or outputs. Enable either Eve option only after confirming that exporting conversation content matches your privacy and retention requirements.

1
export default withEveLineage(
2
  defineInstrumentation({
3
    recordInputs: true,
4
    recordOutputs: true,
5
  }),
6
);

With both options enabled, Respan records serialized chat input and output. Tool spans include their structured arguments and result.

Delegated subagents

Keep withEveLineage(...) enabled when your agent declares subagents. Eve starts a delegated session in a separate OpenTelemetry trace; the helper uses Eve’s exact parent-session and parent-turn lineage so Respan can attach the child agent subtree to its caller.

If lineage is missing or ambiguous, Respan leaves the sessions as separate traces rather than guessing. Without the helper, each Eve session uses its own session ID as its trace group.

If you already define a step.started event hook, put it inside defineInstrumentation. The helper composes the hook and preserves its returned runtime context.

1
export default withEveLineage(
2
  defineInstrumentation({
3
    events: {
4
      "step.started"(input) {
5
        return {
6
          runtimeContext: {
7
            "app.tenant_id":
8
              input.session.auth.current?.attributes.tenantId ?? "",
9
          },
10
        };
11
      },
12
    },
13
  }),
14
);

Do not put secrets in runtime context. User-authored runtime-context values are exported as Eve metadata even when input and output recording are disabled.

What is captured

• Eve turns as agent spans and model steps as task spans
• AI SDK chat spans with model, provider, streaming state, and token usage
• Tool calls with structured input and output
• Eve’s instrumentation functionId as the workflow name on every nested span
• Current and root session identifiers for thread and trace grouping
• Delegated child agents nested under the calling Eve turn when exact lineage is available
• Eve version, environment, turn, step, channel, and authored runtime-context metadata

Configuration