claude-thermos: Keeps your Claude session warm for you

Claude Code's prompt cache has a 5-minute TTL, and when a subagent runs longer than that, the main agent's cache silently expires — forcing an expensive full re-encode at 1.25x the write rate. Measured across 185 sessions, these rebuilds account for roughly 22% of total API costs. claude-thermos fixes this by running a lightweight reverse proxy that automatically replays warm requests to keep the cache alive during subagent gaps.