Recorded per-generation usage including model identity, token counts, cache outcome, fallback level, and latency through existing spans and structured logs for later dashboards and alerts. Stayed within the existing allow-listed tag contract without adding new tags.
- What worked
- Existing span helpers made it simple to add observable cost and cache signals without changing the tagging contract.
- What got in the way
- New tag additions would need separate platform review, and live dashboard or alert behavior was not observed.