discrepancy between sniffly and cmonitor

I'm using a similar tool to sniffly that calculates token usage & cost:
https://github.com/Maciek-roboblog/Claude-Code-Usage-Monitor

I get quite different results. E.g. for august with cmonitor:
input: 18,819 tokens
output: 247,049 tokens
cache create: 6,514,045 tokens
cache read: 113,056,085 tokens
total tokens: 119,835,998 tokens
cost: $62.11

With sniffly:
22.9K input • 708.4K output (I only used claude in august).
$116.01

Which numbers should I trust?

Start by comparing sniffly's August figures with the linked Claude-Code-Usage-Monitor results, focusing on how input, output, cache creation, and cache read tokens are counted. Trace the usage and cost calculation entry points in sniffly, then verify the result against the reported August data; done means the discrepancy is explained and any needed correction is covered by a reproducible check.

Written by the indexing model from the issue text.