Assistant operations

This report turns the production traces in assistant_calls into the operating view for Module 11. The table is intentionally small and append-only: one row per assistant LLM call, with tokens, estimated cost, latency, status, editor context, and request context. That is enough for daily-batch visibility without adding a heavy ops stack to a tutorial that is trying to prove the platform boundary.

The charts below answer four questions an analyst or engineering lead would ask before trusting an editorial assistant in the editor workflow: which model is driving token spend, whether latency is drifting, whether errors are isolated to one assistant, and which recent calls deserve a closer look.[^cost-methodology]

Costliest recent calls

Call-detail view

The drill-down table keeps the request and editor context visible next to the cost and latency fields. It is deliberately row-level rather than aggregated: when a daily chart moves, this is the place to inspect the individual traces that made it move.

[^cost-methodology]: Cost estimates use the checked-in per-model rate table in tutorial/assistants/llm_rates.json and the refresh process documented in tutorial/assistants/README.md.