Hand heavy follow-on work to a fresh subagent when the main session context is very long
In a long Claude Code session every further turn re-reads the whole conversation, and after an idle gap longer than the prompt-cache TTL that re-read is uncached. When a background watcher woke the main agent after 36 idle minutes with a new multi-step task (set up a second Python env, write an adapter, test, run), delegating it to a fresh general-purpose subagent kept the many tool-heavy steps out of the large context. What the brief needed to be self-sufficient: the exact files to read first, the contract the new code must satisfy and the function that validates it, what must not be modified, the project's hard rules restated in full (the subagent sees no CLAUDE.md conversation history), exact verification commands, spend limits for paid calls, commit format, and an explicit list of what to report including what it could not verify. Create the working branch yourself before launching so the subagent cannot pick the wrong base.