Lab 10 — Compaction
← Back to labs index

Lab 10 — Compaction

Use this lab’s provider.env.example and the preflight. In a disposable workspace write two service files, with matching publish/consume queue orders-demo. Start interactive Copilot, summarize both files and use /context to inspect context usage. Ask several follow-ups, then enter /compact inside the CLI, not in your shell. Ask again for cited queue values and compare to both files. Record what was retained and whether file reads recurred. Note any usage or caching counters the provider exposes; never infer cache hits from timing alone.

Compaction may discard details. It is not build-your-own Lab 10’s fixed-schema transcript rewrite or an unlimited context window.