Cost metadata, not customer conversations.

Operational fields include project, provider, model, endpoint, token counts, latency, cost, function name, file name, line number, cache and outcome diagnostics, and permitted scalar custom metadata.

The SDK does not send or store prompts, responses, message arrays, tool definitions, provider hashes, fingerprints or diagnostic objects. Do not place personal, confidential or prompt content in endpoint names or custom metadata.

template_only, redacted_sample and off remain accepted so existing deployments do not break. Current SDK releases treat every mode as metadata-only and emit no prompt text.

Classification uses stable workload identifiers and optional task overrides. Recommendations use cost, usage and evaluation metadata; they do not automatically rewrite prompts or change production models.