Claude Prompt Cache Keepalive — Development skill for Claude Code

keeping-prompt-cache-warm

An agent skill for [Claude Code](https://claude.com/claude-code) that keeps a session's prompt cache alive while you are away, so a large conversation does not have to be re-read from scratch when you come back.

The problem

Claude Code caches the conversation prefix for **exactly 60 minutes** after the last model request. If you leave a long session idle for longer than that (overnight, a meeting, a long background job), the next message re-reads the whole context uncached: slow, and it burns usage quota.

What does and does not work

Verified in the Claude Code VS Code extension on Windows, by reading the session transcript usage (`cache_read_input_tokens` vs `cache_creation_input_tokens`):

Mechanism	Wakes an idle session?
Background Bash/PowerShell job (run_in_background: true) when it exits	Yes
CronCreate recurring job	No: fired 0 times in 8.5 hours
ScheduleWakeup	Only inside /loop
Monitor	Expires in 30-60 minutes
Subagents / workflows running in the background	Do not refresh the main thread cache; only their final notification wakes it

A gap of 60 min 20 s was already enough to lose the cache, so there is no slack beyond the hour.

How it works

The skill arms a **ladder of independent background timers** in one shot (not a chain, so one missed wake does not stop the rest). Each timer:

• sleeps until its scheduled minute, polling a stop file every minute;
• holds an OS sleep lock while waiting (Windows SetThreadExecutionState, macOS caffeinate, Linux systemd-inhibit);
• exits printing KEEPALIVE_TICK, whose completion notification wakes the model for a one-line, tool-free reply that refreshes the cache.

Install

git clone https://github.com/santiquiroz/claude-prompt-cache-keepalive ~/.claude/skills/keeping-prompt-cache-warm

Then just tell Claude something like "I'm going to sleep, keep the cache warm for 8 hours".

The wake