GitHub Copilot coding agent cache drops after 5 minutes

An analysis of public GitHub Copilot coding-agent traces reveals that LLM caches drop off a cliff after an idle period of roughly five minutes, regardless of whether the pause was caused by a human user or a tool execution. The author finds that token-weighted cache survival varies drastically by model, with some retaining strong retention past five minutes while most lose the vast majority of their cached prompt tokens. Open article →