Instrumentation
What the harness records for every run, where to read it, how to export.
Every run produces a structured record — intent classifications, confidence per turn, step transitions, escalation triggers, the final outcome. The record exists from the first customer turn to the terminal step, and persists for ninety days on Pro (forty-five with BYO keys).
Reading the log
Account → Workflows → [workflow] → Runs lists every recorded run with its outcome and step path. Clicking a run opens the full structured record alongside the turn-by-turn transcript. The same data is the source of the Live Preview's rail visualization.
Exporting
Account → Usage → Export produces a JSONL file of all runs in the selected window. Every line is one run, with nested fields for transitions, classifications, and escalations. The shape is documented in the Reference section.
What is not recorded
- Raw model responses outside the harness — only the structured fields the workflow extracted are kept.
- Customer PII beyond what the workflow's flow definition explicitly collects.
- Anything from the Live Preview, which runs entirely in-browser and persists nothing.