Response.Attempts reports how many attempts produced the response when a retrying decorator counts them. It is deliberately a small integer with explicit zero semantics.

Contract

type Response struct {
	// ...
	Attempts int
}

Value	Meaning
`0`	Serving client does not count attempts
`1`	First attempt succeeded
`>1`	Retries were needed before success

response, err := client.Invoke(ctx, request)
if err != nil {
	return err
}
switch {
case response.Attempts == 0:
	// Do not infer retry behavior.
case response.Attempts == 1:
	// First attempt succeeded.
default:
	metrics.RetryCount("inference", response.Attempts-1)
}

Streaming exposes the same concept as stream.StreamResult.Attempts, where it counts establishment attempts before the stream opened. It is not a count of chunks, tool calls, or provider token generations.

Proof

Source: inference/client.go, inference/stream/result.go
Tests: inference/stream/stream_test.go, inference/retry/invoke_test.go