Attribute each model call to its tutoring run before posting the error; otherwise, the failure record can explain what broke while leaving the cost of the failed lesson invisible. TL;DR: preserve the original exception, attach a bounded run identifier and per-call evidence, and verify delivery separately from the lesson outcome. Keep accounting evidence when error delivery is disabled.
The implementation question often starts with a Node.js catch block and an HTTP client such as axios. The useful contract is an explicit POST with message, stack, environment, and context. This guide demonstrates that capture-and-post boundary in Go, using an edtech agent loop whose operational question is which calls a failed lesson paid for.
What page fired, and which tutoring run paid for it?
Start the postmortem with the page, its triggering condition, and the affected run identifiers. An error event doesn't establish that someone was paged, that a learner was affected, or that an expensive call caused the failure.…