Herdr Auto Warm Cache

Cache Keepalive

**Your idle agents lose their prompt cache after an hour. This types one message into them before that happens.**

A [herdr](https://herdr.dev) plugin. Nothing is ever typed into a pane that is waiting on you, and your focus is never taken.

---

The problem, in numbers

An agent session's conversation is cached against the exact token prefix the request carries. Reading that cache bills at **0.1×** base input; writing it at the one-hour TTL bills at **2.0×** — and the TTL refreshes every time the cache is read. So a session stays cheap while you are talking to it, and the first message after an hour of silence pays 2.0× on the whole accumulated conversation.

Nothing surfaces this. The turn simply costs twenty times the one before it.

Measured across 32,037 billed turns in thirty days on one machine: **87 cold returns, 24.2M tokens re-written, average gap 5.0 hours.** Three real examples, where a message arrived *just* inside the hour:

gap	tokens read	paid	cold would have cost
57.5 min	341,693	$0.176	$3.42
57.3 min	510,815	$0.320	$5.17
51.7 min	570,879	$0.340	$5.77

Count your own before you install anything:

for f in ~/.claude/projects/*/*.jsonl; do jq -rc 'select(.type=="assistant" and .message.usage != null) | [(.timestamp|sub("\\.[0-9]+Z$";"Z")|fromdateiso8601), (.message.usage.cache_read_input_tokens // 0), (.message.usage.cache_creation_input_tokens // 0)] | @tsv' "$f" 2>/dev/null | awk 'NR>1 && $1-p>3600 && $2==0 && $3>50000 {n++; t+=$3} {p=$1} END{if(n) printf "%d cold returns, %.1fM re-written\n", n, t/1e6}'; done

What it does

at	what happens
50 min idle	the pane is marked drowsy in a state file, so a bar can badge it
55 min	a notification asks. Nothing is typed yet.
57 min	unanswered, the message is typed and sent in one step

The notification offers whichever outcome will *not* happen on its own, so the click is never