Lab 7 — Observability and caching
← Back to labs index

Lab 7 — Observability and caching

Use this lab’s provider.env.example and the preflight. In a disposable workspace create service.txt with a synthetic port. Inspect copilot help monitoring. In interactive copilot -C "$workdir" ask for the port twice in fresh sessions; use /usage and /context to inspect reported usage. If you choose JSONL output, inspect copilot --output-format json --help and keep the stream private; do not run -p with blanket approval on a real repository.

Record exposed usage, actual tool invocations, latency and provider details. Do not infer cache savings or provider-reported tokens from model output. Copilot’s monitoring and CLI JSON events are not automatically the build-your-own OTel span tree, per-tool cost allocation or Foundry tracing. Mark unexposed fields unavailable and never publish prompts or tool results from logs.