Cache Stats

SillyTavern does not persist Anthropic (or OpenAI-style) prompt-cache usage on chat messages. This extension tees the generate stream, reads cache hit / write from the reply, stores it on that message, and puts a database icon on the message-actions row.

Tap the icon to see hit rate, cache read, cache write, uncached input, and output for that reply.

Older messages stay blank until you generate or swipe again — there is nothing to recover after the fact.

Install

In SillyTavern: Extensions → Install extension, paste:

https://github.com/crossps/st-cache-stats

Enable Cache Stats, then refresh. The next generate or swipe records stats.

Works on PC and phone installs the same way. No settings, no extra server.

What it captures

• Anthropic SSE message_start / message_delta usage (cache_read_input_tokens, cache_creation_input_tokens, 5m / 1h writes)
• OpenAI-style prompt_tokens_details.cached_tokens when that is what the backend returns

Quiet / impersonate generations are skipped. Stats live on message.extra.cache_usage (and the matching swipe extra), so they travel with the chat file.