Post-Call Workflows vs Oversight

Two different things are called "post-call" and conflating them is why deterministic testing looks impossible at first glance. The post-call workflow (common-webhook, end-of-call orchestration) takes actions: it books, it writes to the CRM, it classifies and sends notifications. On a test call that workflow is mocked or skipped, precisely so a test does not mutate real systems, which means the completed outcomes it would have produced simply do not exist to assert against. Oversight (analyze-call) is a different layer: post-call analysis, which reads the call and computes evaluations and takes no actions at all. That distinction is what makes it safe and correct to run on test calls, and it is why Oversight, not the workflow, is the engine for deterministic test verification. Today test traffic deliberately skips it (shouldHideCall suppresses the insights emit for simulation and test-phone calls, to protect the insights UI and LLM-eval cost), so auto-firing on test calls is a targeted relaxation of one condition rather than new infrastructure. The boundary it leaves is honest and worth stating plainly: records written in-call (the transfer destination log, the tool calls and their arguments) are always readable, while workflow-produced outcomes are not, so the deterministic check for booking is the agent's intent (did it call the booking tool with the right arguments) rather than the completed row.