Keep Claude Code’s cache warm. On Fable 5.1, a 1-hour cache write costs 80x a read.

Cache Tax Refusal Receipt GIF Step away from a long Claude Code session for an hour and your next message pays to rewrite the whole context. Same thing when a background agent keeps working while the main conversation sits idle. I built cache-tax for sessions I'm working in over a stretch, with breaks or background work in between. Free and open source. Built with Claude Code, for Claude Code. GitHub + install Keep it warm while you're away. /keepwarm arms six hours; /keepwarm 90m sets a shorter window. After 50 idle minutes, it sends a request over your session without adding heartbeat messages to the conversation. Keep it warm while you wait. Leave warming armed while a separate agent or long workflow runs and the main conversation sits idle. Catch a cold return. With 50k or more tokens of context idle past the hour, it stops your next ordinary message once with the estimated rewrite price. Resend to continue, or /clear and start from a note. Real receipts: the GIF shows a $6.61 estimate and a $6.28 reported cache write after resending. In a separate background-agent run, my returning message read 156,886 cached tokens , 62 minutes after the previous main response, past the one-hour expiry, so the ping did its job. Use it for sessions you'll return to. Keeping abandoned sessions warm wastes tokens. Each ping checks its usage and stops the loop if it reads nothing or writes at least 10% of what it reads. /keepwarm off stops it. Warming needs CLAUDE_CODE_ENABLE_FUNCTION_HOOKS=1 , a one-hour cache and Claude Code left running. Pings cost tokens, including uncapped output. No function-hooks flag? The hook version gives you price warnings, optional cold-send blocking and a status-line countdown. Keep-warm needs the Mod. The 80x figure compares Fable 5.1's one-hour write and read rates. These are API prices, not extra subscription charges. I haven't measured quota savings; prefix changes can still break the cache. Anthropic's explanation of prompt caching